chore: release v1.0.0

- Clamp KDE candidates to parameter bounds before evaluating l(x)/g(x),
  matching the univariate TPE behavior; without this, candidates scored
  well at out-of-bounds locations but became suboptimal when clamped
- Sort HashMap iterations by ParamId before consuming the seeded RNG to
  eliminate non-deterministic sampling caused by global ParamId counter
- Replace wall-clock timing assertion in async concurrency test with
  atomic max-active counter to avoid CI flakiness
- Gate unused HashSet import behind cfg(feature = "async")

.claude/commands/commit-push-pr.md

@@ -0,0 +1,101 @@
+---
+allowed-tools: Bash(git add:*), Bash(git status:*), Bash(git commit:*), Bash(git diff:*), Bash(git log:*), Bash(git config:*), Bash(git push:*), Bash(git branch:*), Bash(git rev-parse:*), Bash(gh pr:*)
+description: Create a git commit, push, and open a PR
+---
+
+## Context
+
+- Current git status: !`git status`
+- Current git diff (staged and unstaged changes): !`git diff HEAD`
+- Current branch: !`git branch --show-current`
+- Remote tracking: !`git rev-parse --abbrev-ref --symbolic-full-name @{u} 2>/dev/null || echo "no upstream"`
+- Recent commits (for style reference): !`git log --oneline -10`
+- Author commits (for style reference): !`git log --author="$(git config user.email)" --oneline -10`
+- All commits on this branch not on main: !`git log --oneline main..HEAD 2>/dev/null`
+- User input (optional, can specify target branch): $ARGUMENTS
+
+## Your task
+
+Based on the above changes, create a git commit, push, and open a pull request.
+
+1. If there are no staged changes, stage only the relevant changed files by name (never use `git add -A` or `git add .`).
+2. Do NOT commit files that likely contain secrets (.env, credentials.json, etc).
+3. Write a commit message following the guidelines below, then commit.
+4. Push to the remote. Use `git push -u origin <branch>` if there is no upstream set.
+5. Create a pull request using `gh pr create` targeting the base branch:
+   - If the user specified a target branch in their input, use that.
+   - Otherwise, default to `main`.
+6. The PR title should match the commit title (or summarize all branch commits if multiple).
+7. Write the PR body following the PR format below.
+8. Return the PR URL when done.
+
+## Commit message guidelines
+
+- Use imperative mood (e.g., "add feature" not "added feature")
+- First line: brief summary, max 72 characters
+- Focus on the "why" and "what", not the "how"
+- Be specific but concise
+- Many commits do not need a body if the title is self-explanatory
+  - Litmus test: "Would a developer understand this commit from the title + diff?" If yes, skip the body.
+- Do NOT include "Generated with ...", "Co-Authored-By ...", or any AI attribution
+
+### Conventional commit prefixes
+
+Match the prefix to the nature of the change. These must align with `cliff.toml` commit parsers so they appear correctly in the changelog.
+
+**Changelog: "Added"**
+- `feat:` / `feat(scope):` — new feature
+
+**Changelog: "Fixed"**
+- `fix:` / `fix(scope):` — bug fix
+
+**Changelog: "Changed"**
+- `refactor:` / `refactor(scope):` — code restructuring without behavior change
+- `perf:` / `perf(scope):` — performance improvement
+- `docs:` / `docs(scope):` — documentation changes (README, guides, etc.)
+- `style:` / `style(scope):` — formatting only (no logic change)
+- `chore:` / `chore(scope):` — tooling, deps, config, scripts, AI config (.claude/, CLAUDE.md)
+
+**Changelog: "Removed"**
+- `revert:` / `revert(scope):` — revert a previous commit
+
+**Excluded from changelog**
+- `test:` / `test(scope):` — adding or updating tests
+- `ci:` / `ci(scope):` — CI/CD pipeline changes
+
+### Body rules
+
+- When a body is needed (multiple important things in one commit), use bullet points.
+- Body is separated from the title by a blank line.
+- One bullet point per concept/change.
+- Don't explain obvious things like "added unit tests for X".
+
+### Commit format
+
+Always pass the commit message via a HEREDOC:
+
+```
+git commit -m "$(cat <<'EOF'
+<title line>
+
+<optional body>
+EOF
+)"
+```
+
+## PR format
+
+Use a HEREDOC for the body. Look at ALL commits on the branch (not just the latest) to write the summary.
+
+```
+gh pr create --title "<title>" --base <target-branch> --body "$(cat <<'EOF'
+## Summary
+<1-3 bullet points covering all branch commits>
+EOF
+)"
+```
+
+## Before you commit
+
+1. Run `cargo +nightly fmt --all && cargo +nightly clippy --all-features --all-targets --fix --allow-dirty` to auto-format and fix lint issues.
+2. Run all tests and doctests to ensure nothing is broken.

.claude/commands/commit-push.md

@@ -0,0 +1,82 @@
+---
+allowed-tools: Bash(git add:*), Bash(git status:*), Bash(git commit:*), Bash(git diff:*), Bash(git log:*), Bash(git config:*), Bash(git push:*), Bash(git branch:*), Bash(git rev-parse:*)
+description: Create a git commit and push to remote
+---
+
+## Context
+
+- Current git status: !`git status`
+- Current git diff (staged and unstaged changes): !`git diff HEAD`
+- Current branch: !`git branch --show-current`
+- Remote tracking: !`git rev-parse --abbrev-ref --symbolic-full-name @{u} 2>/dev/null || echo "no upstream"`
+- Recent commits (for style reference): !`git log --oneline -10`
+- Author commits (for style reference): !`git log --author="$(git config user.email)" --oneline -10`
+- User input (optional): $ARGUMENTS
+
+## Your task
+
+Based on the above changes, create a single git commit and push it to the remote.
+
+1. If there are no staged changes, stage only the relevant changed files by name (never use `git add -A` or `git add .`).
+2. Do NOT commit files that likely contain secrets (.env, credentials.json, etc).
+3. Write a commit message following the guidelines below, then commit.
+4. Push to the remote. Use `git push -u origin <branch>` if there is no upstream set.
+
+## Commit message guidelines
+
+- Use imperative mood (e.g., "add feature" not "added feature")
+- First line: brief summary, max 72 characters
+- Focus on the "why" and "what", not the "how"
+- Be specific but concise
+- Many commits do not need a body if the title is self-explanatory
+  - Litmus test: "Would a developer understand this commit from the title + diff?" If yes, skip the body.
+- Do NOT include "Generated with ...", "Co-Authored-By ...", or any AI attribution
+
+### Conventional commit prefixes
+
+Match the prefix to the nature of the change. These must align with `cliff.toml` commit parsers so they appear correctly in the changelog.
+
+**Changelog: "Added"**
+- `feat:` / `feat(scope):` — new feature
+
+**Changelog: "Fixed"**
+- `fix:` / `fix(scope):` — bug fix
+
+**Changelog: "Changed"**
+- `refactor:` / `refactor(scope):` — code restructuring without behavior change
+- `perf:` / `perf(scope):` — performance improvement
+- `docs:` / `docs(scope):` — documentation changes (README, guides, etc.)
+- `style:` / `style(scope):` — formatting only (no logic change)
+- `chore:` / `chore(scope):` — tooling, deps, config, scripts, AI config (.claude/, CLAUDE.md)
+
+**Changelog: "Removed"**
+- `revert:` / `revert(scope):` — revert a previous commit
+
+**Excluded from changelog**
+- `test:` / `test(scope):` — adding or updating tests
+- `ci:` / `ci(scope):` — CI/CD pipeline changes
+
+### Body rules
+
+- When a body is needed (multiple important things in one commit), use bullet points.
+- Body is separated from the title by a blank line.
+- One bullet point per concept/change.
+- Don't explain obvious things like "added unit tests for X".
+
+### Commit format
+
+Always pass the commit message via a HEREDOC:
+
+```
+git commit -m "$(cat <<'EOF'
+<title line>
+
+<optional body>
+EOF
+)"
+```
+
+## Before you commit
+
+1. Run `cargo +nightly fmt --all && cargo +nightly clippy --all-features --all-targets --fix --allow-dirty` to auto-format and fix lint issues.
+2. Run all tests and doctests to ensure nothing is broken.

.claude/commands/commit.md

@@ -0,0 +1,81 @@
+---
+allowed-tools: Bash(git add:*), Bash(git status:*), Bash(git commit:*), Bash(git diff:*), Bash(git log:*), Bash(git config:*), Bash(git branch:*)
+description: Create a git commit
+---
+
+## Context
+
+- Current git status: !`git status`
+- Current git diff (staged and unstaged changes): !`git diff HEAD`
+- Current branch: !`git branch --show-current`
+- Recent commits (for style reference): !`git log --oneline -10`
+- Author commits (for style reference): !`git log --author="$(git config user.email)" --oneline -10`
+- User input (optional): $ARGUMENTS
+
+## Your task
+
+Based on the above changes, create a single git commit.
+
+1. If there are no staged changes, stage only the relevant changed files by name (never use `git add -A` or `git add .`).
+2. Do NOT commit files that likely contain secrets (.env, credentials.json, etc).
+3. Write a commit message following the guidelines below, then commit.
+4. Do NOT push to remote.
+
+## Commit message guidelines
+
+- Use imperative mood (e.g., "add feature" not "added feature")
+- First line: brief summary, max 72 characters
+- Focus on the "why" and "what", not the "how"
+- Be specific but concise
+- Many commits do not need a body if the title is self-explanatory
+  - Litmus test: "Would a developer understand this commit from the title + diff?" If yes, skip the body.
+- Do NOT include "Generated with ...", "Co-Authored-By ...", or any AI attribution
+
+### Conventional commit prefixes
+
+Match the prefix to the nature of the change. These must align with `cliff.toml` commit parsers so they appear correctly in the changelog.
+
+**Changelog: "Added"**
+- `feat:` / `feat(scope):` — new feature
+
+**Changelog: "Fixed"**
+- `fix:` / `fix(scope):` — bug fix
+
+**Changelog: "Changed"**
+- `refactor:` / `refactor(scope):` — code restructuring without behavior change
+- `perf:` / `perf(scope):` — performance improvement
+- `docs:` / `docs(scope):` — documentation changes (README, guides, etc.)
+- `style:` / `style(scope):` — formatting only (no logic change)
+- `chore:` / `chore(scope):` — tooling, deps, config, scripts, AI config (.claude/, CLAUDE.md)
+
+**Changelog: "Removed"**
+- `revert:` / `revert(scope):` — revert a previous commit
+
+**Excluded from changelog**
+- `test:` / `test(scope):` — adding or updating tests
+- `ci:` / `ci(scope):` — CI/CD pipeline changes
+
+### Body rules
+
+- When a body is needed (multiple important things in one commit), use bullet points.
+- Body is separated from the title by a blank line.
+- One bullet point per concept/change.
+- Don't explain obvious things like "added unit tests for X".
+
+### Commit format
+
+Always pass the commit message via a HEREDOC:
+
+```
+git commit -m "$(cat <<'EOF'
+<title line>
+
+<optional body>
+EOF
+)"
+```
+
+## Before you commit
+
+1. Run `cargo +nightly fmt --all && cargo +nightly clippy --all-features --all-targets --fix --allow-dirty` to auto-format and fix lint issues.
+2. Run all tests and doctests to ensure nothing is broken.

.github/workflows/ci.yml

@@ -4,10 +4,10 @@ on:
   push:
     branches: [main, master]
   pull_request:
-    branches: [main, master]
 
 permissions:
   contents: read
+  pull-requests: write
 
 env:
   CARGO_TERM_COLOR: always
@@ -41,11 +41,6 @@ jobs:
   test:
     name: Test
     runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        features:
-          - ""
-          - "async"
     steps:
       - uses: actions/checkout@v6
       - name: Install Rust
@@ -54,12 +49,9 @@ jobs:
           rustup update stable
       - uses: Swatinem/rust-cache@v2
       - name: Run tests
-        run: |
-          if [ -z "${{ matrix.features }}" ]; then
-            cargo test --verbose
-          else
-            cargo test --verbose --features "${{ matrix.features }}"
-          fi
+        run: cargo test --verbose --all-features --all-targets
+      - name: Run doc tests
+        run: cargo test --verbose --all-features --doc
 
   examples:
     name: Examples
@@ -71,10 +63,24 @@ jobs:
           rustup override set stable
           rustup update stable
       - uses: Swatinem/rust-cache@v2
-      - name: Run sync example (ml_hyperparameter_tuning)
-        run: cargo run --example ml_hyperparameter_tuning
-      - name: Run async example (async_api_optimization)
-        run: cargo run --example async_api_optimization --features async
+      - name: Run basic_optimization
+        run: cargo run --example basic_optimization
+      - name: Run parameter_types
+        run: cargo run --example parameter_types --features derive
+      - name: Run sampler_comparison
+        run: cargo run --example sampler_comparison
+      - name: Run pruning
+        run: cargo run --example pruning
+      - name: Run early_stopping
+        run: cargo run --example early_stopping
+      - name: Run async_parallel
+        run: cargo run --example async_parallel --features async
+      - name: Run journal_storage
+        run: cargo run --example journal_storage --features journal
+      - name: Run ask_and_tell
+        run: cargo run --example ask_and_tell
+      - name: Run multi_objective
+        run: cargo run --example multi_objective
 
   docs:
     name: Docs
@@ -91,8 +97,12 @@ jobs:
           RUSTDOCFLAGS: -D warnings
 
   feature-check:
-    name: Feature Combinations
+    name: Feature Combinations (${{ matrix.partition }}/4)
     runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        partition: [1, 2, 3, 4]
     steps:
       - uses: actions/checkout@v6
       - name: Install Rust
@@ -103,7 +113,7 @@ jobs:
         uses: taiki-e/install-action@cargo-hack
       - uses: Swatinem/rust-cache@v2
       - name: Check feature powerset
-        run: cargo hack check --feature-powerset --no-dev-deps
+        run: cargo hack check --feature-powerset --no-dev-deps --partition ${{ matrix.partition }}/4
 
   coverage:
     name: Coverage
@@ -128,14 +138,14 @@ jobs:
           fail_ci_if_error: true
 
   msrv:
-    name: MSRV (1.88)
+    name: MSRV (1.89)
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v6
-      - name: Install Rust 1.88
+      - name: Install Rust 1.89
         run: |
-          rustup override set 1.88
-          rustup update 1.88
+          rustup override set 1.89
+          rustup update 1.89
       - uses: Swatinem/rust-cache@v2
       - name: Check compilation
         run: cargo check --all-features
@@ -250,6 +260,99 @@ jobs:
       - name: Typos Check
         run: typos src/
 
+  bench-check:
+    name: Benchmark Smoke Test
+    if: github.event_name == 'push'
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v6
+      - name: Install Rust
+        run: |
+          rustup override set stable
+          rustup update stable
+      - uses: Swatinem/rust-cache@v2
+      - name: Run benchmarks (smoke test)
+        run: cargo bench --all-features --bench samplers --bench optimization
+
+  bench-compare:
+    name: Benchmark Comparison
+    if: github.event_name == 'pull_request'
+    runs-on: ubuntu-latest
+    steps:
+      - name: Install Rust
+        run: |
+          rustup override set stable
+          rustup update stable
+
+      - name: Install critcmp
+        run: cargo install critcmp
+
+      # --- Run benchmarks on the BASE (target) branch ---
+      - uses: actions/checkout@v6
+        with:
+          ref: ${{ github.event.pull_request.base.sha }}
+          clean: false
+
+      - uses: Swatinem/rust-cache@v2
+        with:
+          key: bench-base
+
+      - name: Bench baseline
+        run: cargo bench --all-features --bench samplers --bench optimization -- --save-baseline base
+
+      # --- Run benchmarks on the HEAD (PR) branch ---
+      - uses: actions/checkout@v6
+        with:
+          ref: ${{ github.event.pull_request.head.sha }}
+          clean: false
+
+      - uses: Swatinem/rust-cache@v2
+        with:
+          key: bench-head
+
+      - name: Bench PR head
+        run: cargo bench --all-features --bench samplers --bench optimization -- --save-baseline head
+
+      # --- Compare and post comment ---
+      - name: Compare benchmarks
+        id: compare
+        run: |
+          EOF=$(dd if=/dev/urandom bs=15 count=1 status=none | base64)
+          echo "result<<$EOF" >> "$GITHUB_OUTPUT"
+          critcmp base head --color never >> "$GITHUB_OUTPUT"
+          echo "$EOF" >> "$GITHUB_OUTPUT"
+
+      - name: Find existing comment
+        uses: peter-evans/find-comment@v4
+        id: find-comment
+        with:
+          issue-number: ${{ github.event.pull_request.number }}
+          comment-author: github-actions[bot]
+          body-includes: "## Benchmark Comparison"
+
+      - name: Post or update PR comment
+        uses: peter-evans/create-or-update-comment@v5
+        with:
+          comment-id: ${{ steps.find-comment.outputs.comment-id }}
+          issue-number: ${{ github.event.pull_request.number }}
+          edit-mode: replace
+          body: |
+            ## Benchmark Comparison
+
+            **Base:** `${{ github.event.pull_request.base.sha }}` | **Head:** `${{ github.event.pull_request.head.sha }}`
+
+            <details>
+            <summary>Click to expand full results</summary>
+
+            ```
+            ${{ steps.compare.outputs.result }}
+            ```
+
+            </details>
+
+            > Benchmarks run on `ubuntu-latest` via [Criterion](https://github.com/bheisler/criterion.rs) + [critcmp](https://github.com/BurntSushi/critcmp).
+            > Results may vary due to shared CI runners. Look for consistent >5% changes.
+
   publish:
     name: Publish to crates.io
     runs-on: ubuntu-latest
@@ -270,6 +373,7 @@ jobs:
       - cross-platform
       - cross-targets
       - typos
+      - bench-check
     steps:
       - uses: actions/checkout@v6
       - name: Install Rust
@@ -281,8 +385,10 @@ jobs:
       - name: Check if version already published
         id: check
         run: |
-          CARGO_VERSION=$(cargo metadata --no-deps --format-version 1 | jq -r '.packages[0].version')
+          CARGO_VERSION=$(cargo metadata --no-deps --format-version 1 | jq -r '.packages[] | select(.name == "optimizer") | .version')
+          DERIVE_VERSION=$(cargo metadata --no-deps --format-version 1 | jq -r '.packages[] | select(.name == "optimizer-derive") | .version')
           echo "version=$CARGO_VERSION" >> "$GITHUB_OUTPUT"
+          echo "derive_version=$DERIVE_VERSION" >> "$GITHUB_OUTPUT"
 
           HTTP_STATUS=$(curl -s -o /dev/null -w "%{http_code}" "https://crates.io/api/v1/crates/optimizer/$CARGO_VERSION")
           if [ "$HTTP_STATUS" = "200" ]; then
@@ -293,7 +399,33 @@ jobs:
             echo "skip=false" >> "$GITHUB_OUTPUT"
           fi
 
-      - name: Publish
+          DERIVE_HTTP_STATUS=$(curl -s -o /dev/null -w "%{http_code}" "https://crates.io/api/v1/crates/optimizer-derive/$DERIVE_VERSION")
+          if [ "$DERIVE_HTTP_STATUS" = "200" ]; then
+            echo "optimizer-derive $DERIVE_VERSION already exists on crates.io"
+            echo "skip_derive=true" >> "$GITHUB_OUTPUT"
+          else
+            echo "optimizer-derive $DERIVE_VERSION not found on crates.io, will publish"
+            echo "skip_derive=false" >> "$GITHUB_OUTPUT"
+          fi
+
+      - name: Publish optimizer-derive
+        if: steps.check.outputs.skip == 'false' && steps.check.outputs.skip_derive == 'false'
+        run: |
+          cargo publish -p optimizer-derive --allow-dirty 2>&1 | tee publish_output.txt || {
+            if grep -q "already uploaded" publish_output.txt; then
+              echo "optimizer-derive already published, skipping"
+              exit 0
+            fi
+            exit 1
+          }
+        env:
+          CARGO_REGISTRY_TOKEN: ${{ secrets.CARGO_REGISTRY_TOKEN }}
+
+      - name: Wait for crates.io index update
+        if: steps.check.outputs.skip == 'false' && steps.check.outputs.skip_derive == 'false'
+        run: sleep 30
+
+      - name: Publish optimizer
         if: steps.check.outputs.skip == 'false'
         run: |
           cargo publish --allow-dirty 2>&1 | tee publish_output.txt || {

CHANGELOG.md

@@ -0,0 +1,230 @@
+# Changelog
+
+All notable changes to this project will be documented in this file.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
+and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+## [0.9.1] - 2026-02-12
+
+### Fixed
+
+- Fix journal file handling to support both reading and writing modes
+
+## [0.9.0] - 2026-02-12
+
+### Added
+
+- Storage trait abstraction with pluggable backends
+- JSONL journal storage backend for persistent studies (behind `journal` feature)
+- Differential Evolution sampler
+- Gaussian Process sampler with Expected Improvement acquisition function
+- NSGA-III sampler for many-objective optimization with reference-point decomposition
+- MOEA/D sampler for decomposition-based multi-objective optimization
+- `StudyBuilder` for fluent study construction
+
+### Changed
+
+- Move `next_trial_id` counter from `Study` into `Storage` trait
+- Replace `rand 0.10` with `fastrand 2.3` for simpler, faster random number generation
+- Remove `visualization` and `fanova` feature flags (now always available)
+- Simplify CI test commands by removing feature matrix
+
+### Fixed
+
+- Seed importance tests to eliminate flakiness
+- Use atomic counter for temp paths in journal tests
+
+### Removed
+
+- Reverted SQLite storage backend (may return in a future release)
+
+## [0.8.1] - 2026-02-11
+
+### Fixed
+
+- Resolve broken rustdoc intra-doc links for `Error::NoCompletedTrials`
+
+## [0.8.0] - 2026-02-11
+
+### Added
+
+- Multi-objective optimization with NSGA-II
+- Multi-Objective TPE (MOTPE) sampler
+- Pareto front analysis utilities
+- CSV and JSON data export for visualization
+- HTML visualization reports with Plotly.js charts
+- fANOVA (functional ANOVA) parameter importance via random forest
+
+### Fixed
+
+- Rename variable to pass typos check
+
+## [0.7.2] - 2026-02-11
+
+### Changed
+
+- Update `nalgebra` dependency to version 0.34
+- Add advisory ignore for unmaintained transitive dependency
+
+## [0.7.1] - 2026-02-11
+
+### Changed
+
+- Update `tracing` dependency to version 0.1.29
+
+## [0.7.0] - 2026-02-11
+
+### Added
+
+- CMA-ES (Covariance Matrix Adaptation Evolution Strategy) sampler behind `cma-es` feature flag
+- Sobol quasi-random sampler behind `sobol` feature flag
+- BOHB (Bayesian Optimization with HyperBand) sampler for budget-aware optimization
+- Parameter importance analysis via Spearman rank correlation
+- Constraint handling with feasibility-aware trial ranking
+- `optimize_with_retries()` for automatic retry of failed trials
+- `optimize_with_checkpoint()` with atomic save writes for crash recovery
+- `Study::summary()` and `Display` impl for study overview
+- `IntoIterator` for `&Study` and `iter()` method
+- Tracing integration behind `tracing` feature flag
+- Serde serialization support behind `serde` feature flag
+- Benchmark suite with criterion and standard test functions
+
+## [0.6.0] - 2026-02-11
+
+### Added
+
+- Pruning system with `Pruner` trait and `NopPruner` default implementation
+- Intermediate value reporting on trials for pruner integration
+- `TrialPruned` error variant and `Pruned` trial state
+- `ThresholdPruner` for fixed-bound trial pruning
+- `MedianPruner` for statistics-based trial pruning
+- `PercentilePruner` for configurable percentile-based trial pruning
+- `PatientPruner` for patience-based trial pruning
+- `SuccessiveHalvingPruner` for SHA-based trial pruning
+- `HyperbandPruner` for multi-bracket trial pruning
+- `WilcoxonPruner` for statistics-based trial pruning
+- `Study::minimize()` and `Study::maximize()` constructor shortcuts
+- `From<RangeInclusive>` for `FloatParam` and `IntParam`
+- Timeout-based optimization with `optimize_until`
+- `Study::top_trials(n)` for retrieving the best N trials
+- Ask-and-tell interface with `ask()` and `tell()` methods
+- Trial user attributes for logging and analysis
+- `enqueue_trial()` for pre-specified parameter evaluation
+
+## [0.5.1] - 2026-02-10
+
+### Changed
+
+- Update random number generation to use `rand::make_rng()` and upgrade `rand` to version 0.10
+
+## [0.5.0] - 2026-02-06
+
+### Added
+
+- Typed parameter API with `FloatParam`, `IntParam`, `CategoricalParam`, `BoolParam`, and `EnumParam`
+- `#[derive(Categorical)]` proc macro for deriving categorical parameters from enums (behind `derive` feature)
+- `.name()` builder method on all parameter types for custom labels
+- `CompletedTrial::get()` for typed parameter access
+- `Display` impl on `ParamValue`
+- Prelude module at `optimizer::prelude::*`
+
+### Changed
+
+- Reorganize sampler module structure and update imports
+- Remove `log` dependency
+
+## [0.4.0] - 2026-02-02
+
+### Added
+
+- Multivariate TPE sampler for correlated parameter search
+- Gamma strategies for TPE sampler (linear, sqrt, fixed) with examples
+- Example: async API parameter optimization
+- Example: ML hyperparameter tuning
+
+### Fixed
+
+- Handle end value in `suggest` method to avoid panic on underflow
+
+## [0.3.1] - 2026-01-31
+
+### Added
+
+- Grid search sampler for exhaustive parameter exploration
+- `suggest_bool()` method for boolean parameter suggestion
+- `SuggestableRange` trait and `suggest_range()` method for parameter suggestion from ranges
+- Documentation, keywords, categories, and readme fields in `Cargo.toml`
+
+### Changed
+
+- Replace `TpeError` with unified `Error` type across the library
+- Add permissions for read access to contents in CI and scheduled workflows
+
+## [0.3.0] - 2026-01-30
+
+### Added
+
+- Cross-target compilation checks in CI
+
+### Changed
+
+- Internal code refactoring and cleanup
+
+## [0.2.0] - 2026-01-30
+
+### Added
+
+- CI step for unused dependencies check with `cargo-machete`
+- CI step to publish to crates.io
+- Default typo extension for 'Tpe'
+
+### Changed
+
+- Remove `serde` dependency (was unused)
+- Remove unused `ordered-float` dependency
+
+## [0.1.1] - 2026-01-30
+
+### Fixed
+
+- Minor release fixes
+
+## [0.1.0] - 2026-01-30
+
+### Added
+
+- Initial implementation of the optimization library
+- TPE (Tree-structured Parzen Estimator) sampler with optional fixed bandwidth KDE
+- Random sampler
+- Async optimization support
+- CI workflows for testing, coverage (Codecov), auditing, and publishing
+- README with project overview, features, and quick start guide
+
+---
+
+## Release Workflow
+
+This changelog is maintained automatically with [git-cliff](https://git-cliff.org/).
+
+1. Write commits using [Conventional Commits](https://www.conventionalcommits.org/) (`feat:`, `fix:`, `refactor:`, etc.).
+2. Tag the release: `git tag v0.X.0`
+3. Regenerate the changelog: `make changelog`
+4. Commit the updated `CHANGELOG.md` and push.
+
+[0.9.1]: https://github.com/raimannma/rust-optimizer/compare/v0.9.0...v0.9.1
+[0.9.0]: https://github.com/raimannma/rust-optimizer/compare/v0.8.1...v0.9.0
+[0.8.1]: https://github.com/raimannma/rust-optimizer/compare/v0.8.0...v0.8.1
+[0.8.0]: https://github.com/raimannma/rust-optimizer/compare/v0.7.2...v0.8.0
+[0.7.2]: https://github.com/raimannma/rust-optimizer/compare/v0.7.1...v0.7.2
+[0.7.1]: https://github.com/raimannma/rust-optimizer/compare/v0.7.0...v0.7.1
+[0.7.0]: https://github.com/raimannma/rust-optimizer/compare/v0.6.0...v0.7.0
+[0.6.0]: https://github.com/raimannma/rust-optimizer/compare/v0.5.1...v0.6.0
+[0.5.1]: https://github.com/raimannma/rust-optimizer/compare/v0.5.0...v0.5.1
+[0.5.0]: https://github.com/raimannma/rust-optimizer/compare/v0.4.0...v0.5.0
+[0.4.0]: https://github.com/raimannma/rust-optimizer/compare/v0.3.1...v0.4.0
+[0.3.1]: https://github.com/raimannma/rust-optimizer/compare/v0.3.0...v0.3.1
+[0.3.0]: https://github.com/raimannma/rust-optimizer/compare/v0.2.0...v0.3.0
+[0.2.0]: https://github.com/raimannma/rust-optimizer/compare/v0.1.1...v0.2.0
+[0.1.1]: https://github.com/raimannma/rust-optimizer/compare/v0.1.0...v0.1.1
+[0.1.0]: https://github.com/raimannma/rust-optimizer/releases/tag/v0.1.0

CLAUDE.md

@@ -0,0 +1,3 @@
+## Git Guidelines
+
+- If the user asks to make a commit, do not add any co-author by claude or any other co-author note.

Cargo.toml

@@ -3,41 +3,44 @@ members = ["optimizer-derive"]
 
 [package]
 name = "optimizer"
-version = "0.7.0"
+version = "1.0.0"
 edition = "2024"
-rust-version = "1.88"
+rust-version = "1.89"
 license = "MIT"
-authors = ["Manuel Raimann <raimannma@outlook.de"]
-description = "A Rust library for optimization algorithms."
+authors = ["Manuel Raimann <raimannma@outlook.de>"]
+description = "Bayesian and population-based optimization library with an Optuna-like API for hyperparameter tuning and black-box optimization"
 repository = "https://github.com/raimannma/rust-optimizer"
 documentation = "https://docs.rs/optimizer"
 keywords = ["optimization", "hyperparameter", "tpe", "grid-search", "bayesian"]
-categories = ["algorithm", "science", "data-structures"]
+categories = ["algorithms", "science", "mathematics"]
 readme = "README.md"
 
 [dependencies]
-rand = "0.10"
+fastrand = "2.3"
 thiserror = "2"
 parking_lot = "0.12"
-tokio = { version = "1", features = ["sync", "rt-multi-thread"], optional = true }
+tokio = { version = "1.30", features = ["sync", "rt-multi-thread"], optional = true }
 optimizer-derive = { version = "0.1.0", path = "optimizer-derive", optional = true }
 serde = { version = "1", features = ["derive"], optional = true }
 serde_json = { version = "1", optional = true }
-tracing = { version = "0.1", optional = true }
+tracing = { version = "0.1.29", optional = true }
 sobol_burley = { version = "0.5", optional = true }
-nalgebra = { version = "0.33", optional = true }
+nalgebra = { version = "0.34", optional = true }
+fs2 = { version = "0.4", optional = true }
 
 [features]
 default = []
 async = ["dep:tokio"]
 derive = ["dep:optimizer-derive"]
 serde = ["dep:serde", "dep:serde_json"]
+journal = ["dep:fs2", "serde"]
 tracing = ["dep:tracing"]
 sobol = ["dep:sobol_burley"]
 cma-es = ["dep:nalgebra"]
+gp = ["dep:nalgebra"]
 
 [dev-dependencies]
-tokio = { version = "1", features = ["rt-multi-thread", "macros", "time"] }
+tokio = { version = "1.30", features = ["rt-multi-thread", "macros", "time"] }
 optimizer-derive = { version = "0.1.0", path = "optimizer-derive" }
 serde_json = "1"
 criterion = { version = "0.8", features = ["html_reports"] }
@@ -51,15 +54,44 @@ name = "optimization"
 harness = false
 
 [[example]]
-name = "async_api_optimization"
-path = "examples/async_api_optimization.rs"
+name = "basic_optimization"
+path = "examples/basic_optimization.rs"
+
+[[example]]
+name = "parameter_types"
+path = "examples/parameter_types.rs"
+required-features = ["derive"]
+
+[[example]]
+name = "sampler_comparison"
+path = "examples/sampler_comparison.rs"
+
+[[example]]
+name = "pruning"
+path = "examples/pruning.rs"
+
+[[example]]
+name = "early_stopping"
+path = "examples/early_stopping.rs"
+
+[[example]]
+name = "async_parallel"
+path = "examples/async_parallel.rs"
 required-features = ["async"]
 
 [[example]]
-name = "ml_hyperparameter_tuning"
-path = "examples/ml_hyperparameter_tuning.rs"
+name = "journal_storage"
+path = "examples/journal_storage.rs"
+required-features = ["journal"]
 
 [[example]]
-name = "parameter_api"
-path = "examples/parameter_api.rs"
-required-features = ["derive"]
+name = "ask_and_tell"
+path = "examples/ask_and_tell.rs"
+
+[[example]]
+name = "multi_objective"
+path = "examples/multi_objective.rs"
+
+[[test]]
+name = "journal_tests"
+required-features = ["journal"]

LICENSE.md

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 optimizer
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

Makefile

@@ -0,0 +1,6 @@
+.PHONY: changelog
+
+# Regenerate CHANGELOG.md from git tags using git-cliff.
+# Install: cargo install git-cliff
+changelog:
+	git-cliff --output CHANGELOG.md

README.md

@@ -1,155 +1,73 @@
 # optimizer
 
-A Rust library for black-box optimization with multiple sampling strategies.
+Bayesian and population-based optimization library with an Optuna-like API
+for hyperparameter tuning and black-box optimization. Supports 12 samplers,
+8 pruners, multi-objective optimization, async parallelism, and persistent storage.
 
 [![Docs](https://docs.rs/optimizer/badge.svg)](https://docs.rs/optimizer)
 [![Crates.io](https://img.shields.io/crates/v/optimizer.svg)](https://crates.io/crates/optimizer)
 [![codecov](https://codecov.io/gh/raimannma/rust-optimizer/graph/badge.svg?token=WOE77XJ4M6)](https://codecov.io/gh/raimannma/rust-optimizer)
 
-## Features
-
-- Optuna-like API for hyperparameter optimization
-- Multiple sampling strategies:
-  - **Random Search** - Simple random sampling for baseline comparisons
-  - **TPE (Tree-Parzen Estimator)** - Bayesian optimization for efficient search
-  - **Grid Search** - Exhaustive search over a specified parameter grid
-- Float, integer, categorical, boolean, and enum parameter types
-- Log-scale and stepped parameter sampling
-- Sync and async optimization with parallel trial evaluation
-- `#[derive(Categorical)]` for enum parameters
-
 ## Quick Start
 
 ```rust
-use optimizer::parameter::{FloatParam, Parameter};
-use optimizer::sampler::tpe::TpeSampler;
-use optimizer::{Direction, Study};
+use optimizer::prelude::*;
 
-// Create a study with TPE sampler
-let sampler = TpeSampler::builder().seed(42).build().unwrap();
-let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+let study: Study<f64> = Study::new(Direction::Minimize);
+let x = FloatParam::new(-10.0, 10.0).name("x");
 
-// Define parameter search space
-let x_param = FloatParam::new(-10.0, 10.0);
+study.optimize(50, |trial| {
+    let val = x.suggest(trial)?;
+    Ok::<_, Error>((val - 3.0).powi(2))
+}).unwrap();
 
-// Optimize x^2 for 20 trials
-study
-    .optimize_with_sampler(20, |trial| {
-        let x = x_param.suggest(trial)?;
-        Ok::<_, optimizer::Error>(x * x)
-    })
-    .unwrap();
-
-// Get the best result
 let best = study.best_trial().unwrap();
-println!("Best value: {}", best.value);
-for (id, label) in &best.param_labels {
-    println!("  {}: {:?}", label, best.params[id]);
-}
+println!("Best x = {:.4}, f(x) = {:.4}", best.get(&x).unwrap(), best.value);
 ```
 
-## Samplers
+## Features at a Glance
 
-### Random Search
-
-```rust
-use optimizer::{Direction, Study};
-use optimizer::sampler::random::RandomSampler;
-
-let study: Study<f64> = Study::with_sampler(
-    Direction::Minimize,
-    RandomSampler::with_seed(42),
-);
-```
-
-### TPE (Tree-Parzen Estimator)
-
-```rust
-use optimizer::{Direction, Study};
-use optimizer::sampler::tpe::TpeSampler;
-
-let sampler = TpeSampler::builder()
-    .gamma(0.15)           // Quantile for good/bad split
-    .n_startup_trials(20)  // Random trials before TPE kicks in
-    .n_ei_candidates(32)   // Candidates to evaluate
-    .seed(42)
-    .build()
-    .unwrap();
-
-let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
-```
-
-#### Gamma Strategies
-
-The gamma parameter controls what fraction of trials are considered "good" when building the TPE model. Instead of a fixed value, you can use adaptive strategies:
-
-| Strategy | Description | Formula |
-|----------|-------------|---------|
-| `FixedGamma` | Constant value (default: 0.25) | `γ = constant` |
-| `LinearGamma` | Linear interpolation over trials | `γ = γ_min + (γ_max - γ_min) * min(n/n_max, 1)` |
-| `SqrtGamma` | Optuna-style inverse sqrt scaling | `γ = min(γ_max, factor/√n / n)` |
-| `HyperoptGamma` | Hyperopt-style adaptive | `γ = min(γ_max, (base + 1) / n)` |
-
-```rust
-use optimizer::sampler::tpe::{TpeSampler, SqrtGamma, LinearGamma};
-
-// Optuna-style gamma that decreases with more trials
-let sampler = TpeSampler::builder()
-    .gamma_strategy(SqrtGamma::default())
-    .build()
-    .unwrap();
-
-// Linear interpolation from 0.1 to 0.3 over 100 trials
-let sampler = TpeSampler::builder()
-    .gamma_strategy(LinearGamma::new(0.1, 0.3, 100).unwrap())
-    .build()
-    .unwrap();
-```
-
-You can also implement custom strategies:
-
-```rust
-use optimizer::sampler::tpe::{TpeSampler, GammaStrategy};
-
-#[derive(Debug, Clone)]
-struct MyGamma { base: f64 }
-
-impl GammaStrategy for MyGamma {
-    fn gamma(&self, n_trials: usize) -> f64 {
-        (self.base + 0.01 * n_trials as f64).min(0.5)
-    }
-    fn clone_box(&self) -> Box<dyn GammaStrategy> {
-        Box::new(self.clone())
-    }
-}
-
-let sampler = TpeSampler::builder()
-    .gamma_strategy(MyGamma { base: 0.1 })
-    .build()
-    .unwrap();
-```
-
-### Grid Search
-
-```rust
-use optimizer::{Direction, Study};
-use optimizer::sampler::grid::GridSearchSampler;
-
-let sampler = GridSearchSampler::builder()
-    .n_points_per_param(10)  // Number of points per parameter dimension
-    .build();
-
-let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
-```
+- **[Samplers](https://docs.rs/optimizer/latest/optimizer/sampler/)** — Random, TPE, Multivariate TPE, Grid, Sobol, CMA-ES, Gaussian Process, Differential Evolution, BOHB, NSGA-II, NSGA-III, MOEA/D
+- **[Pruners](https://docs.rs/optimizer/latest/optimizer/pruner/)** — Median, Percentile, Threshold, Patient, Hyperband, Successive Halving, Wilcoxon, Nop
+- **[Parameters](https://docs.rs/optimizer/latest/optimizer/parameter/)** — Float, Int, Categorical, Bool, and Enum types with `.name()` labels and typed access
+- **[Multi-objective](https://docs.rs/optimizer/latest/optimizer/multi_objective/)** — Pareto front extraction with NSGA-II/III and MOEA/D
+- **[Async & parallel](https://docs.rs/optimizer/latest/optimizer/struct.Study.html#method.optimize_parallel)** — Concurrent trial evaluation with Tokio
+- **[Storage backends](https://docs.rs/optimizer/latest/optimizer/storage/)** — In-memory (default) or JSONL journal for persistence and resumption
+- **[Visualization](https://docs.rs/optimizer/latest/optimizer/fn.generate_html_report.html)** — HTML reports with optimization history and parameter importance
+- **[Analysis](https://docs.rs/optimizer/latest/optimizer/struct.Study.html#method.fanova)** — fANOVA and Spearman correlation for parameter importance
 
 ## Feature Flags
 
-- `async` - Enable async optimization methods (requires tokio)
-- `derive` - Enable `#[derive(Categorical)]` for enum parameters
+| Flag | Enables | Default |
+|------|---------|---------|
+| `async` | Async/parallel optimization (Tokio) | No |
+| `derive` | `#[derive(Categorical)]` for enum parameters | No |
+| `serde` | Serialization of trials and parameters | No |
+| `journal` | JSONL storage backend (implies `serde`) | No |
+| `sobol` | Sobol quasi-random sampler | No |
+| `cma-es` | CMA-ES sampler (requires `nalgebra`) | No |
+| `gp` | Gaussian Process sampler (requires `nalgebra`) | No |
+| `tracing` | Structured logging with `tracing` | No |
 
-## Documentation
+## Examples
 
-Full API documentation is available at [docs.rs/optimizer](https://docs.rs/optimizer).
+```sh
+cargo run --example basic_optimization                       # Minimize a quadratic — simplest possible usage
+cargo run --example parameter_types --features derive        # All 5 param types + #[derive(Categorical)]
+cargo run --example sampler_comparison                       # Compare Random, TPE, and Grid on the same problem
+cargo run --example pruning                                  # Trial pruning with MedianPruner
+cargo run --example early_stopping                           # Halt a study when a target is reached
+cargo run --example async_parallel --features async          # Evaluate trials concurrently with tokio
+cargo run --example journal_storage --features journal       # Persist trials to disk and resume later
+cargo run --example ask_and_tell                             # Decouple sampling from evaluation
+cargo run --example multi_objective                          # Optimize competing objectives + Pareto front
+```
+
+## Learn More
+
+- [API documentation](https://docs.rs/optimizer)
+- [Changelog](CHANGELOG.md)
+- [GitHub Issues](https://github.com/raimannma/rust-optimizer/issues)
 
 ## License
 

benches/optimization.rs

@@ -2,10 +2,10 @@
 mod test_functions;
 
 use criterion::{BenchmarkId, Criterion, criterion_group, criterion_main};
-use optimizer::parameter::Parameter;
+use optimizer::Study;
+use optimizer::parameter::{FloatParam, Parameter};
 use optimizer::sampler::random::RandomSampler;
 use optimizer::sampler::tpe::TpeSampler;
-use optimizer::{FloatParam, Study};
 
 fn make_params(dims: usize) -> Vec<FloatParam> {
     (0..dims)
@@ -23,7 +23,7 @@ fn bench_tpe_sphere(c: &mut Criterion) {
             b.iter(|| {
                 let study = Study::minimize(TpeSampler::builder().seed(42).build().unwrap());
                 study
-                    .optimize(100, |trial| {
+                    .optimize(100, |trial: &mut optimizer::Trial| {
                         let x: Vec<f64> = params
                             .iter()
                             .map(|p| p.suggest(trial))
@@ -48,7 +48,7 @@ fn bench_tpe_rosenbrock(c: &mut Criterion) {
             b.iter(|| {
                 let study = Study::minimize(TpeSampler::builder().seed(42).build().unwrap());
                 study
-                    .optimize(100, |trial| {
+                    .optimize(100, |trial: &mut optimizer::Trial| {
                         let x: Vec<f64> = params
                             .iter()
                             .map(|p| p.suggest(trial))
@@ -72,7 +72,7 @@ fn bench_random_vs_tpe(c: &mut Criterion) {
         b.iter(|| {
             let study = Study::minimize(RandomSampler::with_seed(42));
             study
-                .optimize(100, |trial| {
+                .optimize(100, |trial: &mut optimizer::Trial| {
                     let x: Vec<f64> = params
                         .iter()
                         .map(|p| p.suggest(trial))
@@ -88,7 +88,7 @@ fn bench_random_vs_tpe(c: &mut Criterion) {
         b.iter(|| {
             let study = Study::minimize(TpeSampler::builder().seed(42).build().unwrap());
             study
-                .optimize(100, |trial| {
+                .optimize(100, |trial: &mut optimizer::Trial| {
                     let x: Vec<f64> = params
                         .iter()
                         .map(|p| p.suggest(trial))

benches/samplers.rs

@@ -1,12 +1,11 @@
 use std::collections::HashMap;
 
 use criterion::{BenchmarkId, Criterion, criterion_group, criterion_main};
-use optimizer::parameter::Parameter;
-use optimizer::sampler::Sampler;
-use optimizer::sampler::grid::GridSearchSampler;
+use optimizer::parameter::{FloatParam, Parameter};
+use optimizer::sampler::grid::GridSampler;
 use optimizer::sampler::random::RandomSampler;
 use optimizer::sampler::tpe::TpeSampler;
-use optimizer::{CompletedTrial, FloatParam};
+use optimizer::sampler::{CompletedTrial, Sampler};
 
 /// Build a synthetic history of `n` completed trials over `dims` float parameters.
 fn build_history(n: usize, dims: usize) -> Vec<CompletedTrial<f64>> {
@@ -33,7 +32,7 @@ fn build_history(n: usize, dims: usize) -> Vec<CompletedTrial<f64>> {
         let value: f64 = param_values
             .values()
             .map(|v| {
-                let optimizer::ParamValue::Float(f) = v else {
+                let optimizer::parameter::ParamValue::Float(f) = v else {
                     unreachable!()
                 };
                 f * f
@@ -98,7 +97,7 @@ fn bench_grid_sample(c: &mut Criterion) {
             |b, _| {
                 b.iter(|| {
                     // Fresh sampler each iteration since grid tracks used points
-                    let sampler = GridSearchSampler::builder()
+                    let sampler = GridSampler::builder()
                         .n_points_per_param(grid_points)
                         .build();
                     sampler.sample(&dist, 0, &history)

cliff.toml

@@ -0,0 +1,99 @@
+# git-cliff configuration
+# https://git-cliff.org/docs/configuration
+
+[changelog]
+# changelog header
+header = """
+# Changelog
+
+All notable changes to this project will be documented in this file.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
+and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).\n
+"""
+# template for the changelog body
+# https://keats.github.io/tera/docs/#introduction
+body = """
+{%- macro remote_url() -%}
+  https://github.com/raimannma/rust-optimizer
+{%- endmacro -%}
+
+{% if version -%}
+    ## [{{ version | trim_start_matches(pat="v") }}] - {{ timestamp | date(format="%Y-%m-%d") }}
+{% else -%}
+    ## [Unreleased]
+{% endif -%}
+
+{% for group, commits in commits | group_by(attribute="group") %}
+    ### {{ group | striptags | trim }}
+    {% for commit in commits %}
+        - {{ commit.message | split(pat="\n") | first | trim }}\
+          {%- if commit.breaking %} (**BREAKING**){% endif -%}
+    {% endfor %}
+{% endfor %}
+"""
+# template for the changelog footer
+footer = """
+{%- macro remote_url() -%}
+  https://github.com/raimannma/rust-optimizer
+{%- endmacro -%}
+
+{% for release in releases -%}
+    {% if release.version -%}
+        {% if release.previous.version -%}
+            [{{ release.version | trim_start_matches(pat="v") }}]: \
+                {{ self::remote_url() }}/compare/{{ release.previous.version }}...{{ release.version }}
+        {% else -%}
+            [{{ release.version | trim_start_matches(pat="v") }}]: \
+                {{ self::remote_url() }}/releases/tag/{{ release.version }}
+        {% endif -%}
+    {% else -%}
+        {% if release.previous.version -%}
+            [Unreleased]: {{ self::remote_url() }}/compare/{{ release.previous.version }}...HEAD
+        {% endif -%}
+    {% endif -%}
+{% endfor %}
+"""
+# remove the leading and trailing whitespace from the templates
+trim = true
+
+[git]
+# parse the commits based on https://www.conventionalcommits.org
+conventional_commits = true
+# filter out the commits that are not conventional
+filter_unconventional = false
+# process each line of a commit as an individual commit
+split_commits = false
+# regex for preprocessing the commit messages
+commit_preprocessors = [
+    { pattern = '\((\w+\s)?#([0-9]+)\)', replace = "([#${2}](https://github.com/raimannma/rust-optimizer/issues/${2}))" },
+]
+# regex for parsing and grouping commits
+commit_parsers = [
+    { message = "^feat", group = "Added" },
+    { message = "^fix", group = "Fixed" },
+    { message = "^refactor", group = "Changed" },
+    { message = "^perf", group = "Changed" },
+    { message = "^doc", group = "Changed" },
+    { message = "^style", group = "Changed" },
+    { message = "^ci", skip = true },
+    { message = "^chore\\(release\\)", skip = true },
+    { message = "^chore: release", skip = true },
+    { message = "^chore", group = "Changed" },
+    { message = "^revert", group = "Removed" },
+    { body = ".*security", group = "Security" },
+]
+# protect breaking changes from being skipped due to matching a skipping commit_parser
+protect_breaking_commits = true
+# filter out the commits that are not matched by commit parsers
+filter_commits = true
+# regex for matching git tags
+tag_pattern = "v[0-9].*"
+# regex for skipping tags
+skip_tags = ""
+# regex for ignoring tags
+ignore_tags = ""
+# sort the tags topologically
+topo_order = false
+# sort the commits inside sections by oldest first
+sort_commits = "oldest"

deny.toml

@@ -1,7 +1,11 @@
 [advisories]
 version = 2
 db-path = "~/.cargo/advisory-db"
-ignore = []
+ignore = [
+    # paste is unmaintained but it's a transitive dep via simba -> nalgebra
+    # with no fix available upstream yet
+    "RUSTSEC-2024-0436",
+]
 
 [licenses]
 version = 2
@@ -9,8 +13,8 @@ allow = [
     "MIT",
     "Apache-2.0",
     "Apache-2.0 WITH LLVM-exception",
-    "BSD-2-Clause",
     "Unicode-3.0",
+    "Zlib",
 ]
 confidence-threshold = 0.8
 

examples/ask_and_tell.rs

@@ -0,0 +1,51 @@
+//! Ask-and-tell interface — decouple sampling from evaluation.
+//!
+//! Use `ask()` to get a trial with sampled parameters, evaluate it however
+//! you like (workers, GPUs, external processes), then `tell()` the result.
+//! This is useful for batch evaluation or custom scheduling.
+//!
+//! Run with: `cargo run --example ask_and_tell`
+
+use optimizer::prelude::*;
+
+fn main() -> optimizer::Result<()> {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+
+    let x = FloatParam::new(-5.0, 5.0).name("x");
+    let y = FloatParam::new(-5.0, 5.0).name("y");
+
+    for batch in 0..3 {
+        let batch_size = 5;
+        let mut trials = Vec::with_capacity(batch_size);
+
+        // ask() creates trials with sampled parameters
+        for _ in 0..batch_size {
+            let mut trial = study.ask();
+            let xv = x.suggest(&mut trial)?;
+            let yv = y.suggest(&mut trial)?;
+            trials.push((trial, xv, yv));
+        }
+
+        // Evaluate the batch (could be sent to workers, GPUs, etc.)
+        for (trial, xv, yv) in trials {
+            let value = xv * xv + yv * yv;
+            study.tell(trial, Ok::<_, &str>(value));
+        }
+
+        println!(
+            "Batch {}: evaluated {batch_size} trials (total: {})",
+            batch + 1,
+            study.n_trials(),
+        );
+    }
+
+    let best = study.best_trial()?;
+    println!(
+        "Best: f({:.3}, {:.3}) = {:.6}",
+        best.get(&x).unwrap(),
+        best.get(&y).unwrap(),
+        best.value,
+    );
+
+    Ok(())
+}

examples/async_api_optimization.rs

@@ -1,368 +0,0 @@
-//! Async API Parameter Optimization Example
-//!
-//! This example shows how to use async/parallel optimization to tune
-//! configuration parameters for a web service. Each evaluation simulates
-//! an async operation (like deploying and load-testing a service).
-//!
-//! # Key Concepts Demonstrated
-//!
-//! - Async optimization with `optimize_parallel`
-//! - Running multiple trials concurrently for faster optimization
-//! - Boolean and categorical parameter types
-//! - Measuring speedup from parallelism
-//!
-//! # When to Use Async Optimization
-//!
-//! Use async/parallel optimization when your objective function involves:
-//! - Network requests (API calls, database queries)
-//! - File I/O operations
-//! - External service calls
-//! - Any operation where you're waiting for I/O rather than computing
-//!
-//! With parallelism, you can evaluate multiple configurations simultaneously,
-//! significantly reducing total optimization time.
-//!
-//! Run with: `cargo run --example async_api_optimization --features async`
-
-use std::time::{Duration, Instant};
-
-use optimizer::prelude::*;
-
-// ============================================================================
-// Configuration: Service parameters we want to tune
-// ============================================================================
-
-/// Configuration for a web service.
-///
-/// In a real application, these parameters would control:
-/// - Memory allocation (cache sizes)
-/// - Connection management (pool sizes, timeouts)
-/// - Request handling (batching, compression)
-/// - Protocol options (HTTP version, load balancing)
-struct ServiceConfig {
-    cache_size_mb: i64,
-    connection_pool_size: i64,
-    request_timeout_ms: i64,
-    retry_count: i64,
-    batch_size: i64,
-    compression_level: i64,
-    use_http2: bool,
-    load_balancing: String,
-}
-
-// ============================================================================
-// Objective Function: Evaluate a service configuration
-// ============================================================================
-
-/// Simulates deploying and load-testing a service configuration.
-///
-/// In a real scenario, this function might:
-/// 1. Deploy the configuration to a staging environment
-/// 2. Run load tests against the service
-/// 3. Collect metrics (latency, throughput, error rate)
-/// 4. Return a composite score
-///
-/// The async sleep simulates the I/O time of these operations.
-/// This is where parallel execution helps - while one trial is waiting
-/// for I/O, other trials can run.
-#[allow(clippy::too_many_arguments)]
-async fn evaluate_service(config: &ServiceConfig) -> f64 {
-    // Simulate async I/O (deployment, load testing, metric collection)
-    tokio::time::sleep(Duration::from_millis(50)).await;
-
-    // Calculate a score based on how close we are to optimal values
-    // Lower score = better configuration
-    let mut score = 0.0;
-
-    // Cache size: too small = cache misses, too large = wasted memory
-    // Optimal around 512MB
-    let cache_optimal = 512.0;
-    score += ((config.cache_size_mb as f64 - cache_optimal) / 256.0).powi(2);
-
-    // Connection pool: too small = contention, too large = resource waste
-    // Optimal around 100
-    let pool_optimal = 100.0;
-    score += ((config.connection_pool_size as f64 - pool_optimal) / 50.0).powi(2);
-
-    // Timeout: too short = false failures, too long = slow recovery
-    // Optimal around 5000ms
-    let timeout_optimal = 5000.0;
-    score += ((config.request_timeout_ms as f64 - timeout_optimal) / 2000.0).powi(2);
-
-    // Retries: too few = fragile, too many = amplifies failures
-    // Optimal around 3
-    let retry_optimal = 3.0;
-    score += ((config.retry_count as f64 - retry_optimal) / 2.0).powi(2);
-
-    // Batch size: trade-off between latency and throughput
-    // Optimal around 64
-    let batch_optimal = 64.0;
-    score += ((config.batch_size as f64 - batch_optimal) / 32.0).powi(2);
-
-    // Compression level: trade-off between CPU and bandwidth
-    // Optimal around 6
-    let compression_optimal = 6.0;
-    score += ((config.compression_level as f64 - compression_optimal) / 3.0).powi(2);
-
-    // HTTP/2 is generally better for our use case
-    if !config.use_http2 {
-        score += 0.5;
-    }
-
-    // Load balancing strategy affects performance
-    score += match config.load_balancing.as_str() {
-        "round_robin" => 0.0,       // Best for our use case
-        "least_connections" => 0.1, // Good alternative
-        "ip_hash" => 0.2,           // OK for session affinity
-        "random" => 0.3,            // Not ideal
-        _ => 1.0,
-    };
-
-    // Add noise to simulate real-world variability
-    let noise = (config.cache_size_mb as f64 * 0.1).sin() * 0.05;
-
-    score + noise
-}
-
-/// The async objective function for each trial.
-///
-/// For async optimization, the objective function must:
-/// 1. Take ownership of the Trial (not a mutable reference)
-/// 2. Return a Future
-/// 3. Return both the Trial and the result value as a tuple
-///
-/// This ownership pattern allows the trial to be used across await points.
-#[allow(clippy::too_many_arguments)]
-async fn objective(
-    mut trial: Trial,
-    cache_size_mb_param: &IntParam,
-    connection_pool_size_param: &IntParam,
-    request_timeout_ms_param: &IntParam,
-    retry_count_param: &IntParam,
-    batch_size_param: &IntParam,
-    compression_level_param: &IntParam,
-    use_http2_param: &BoolParam,
-    load_balancing_param: &CategoricalParam<&str>,
-) -> optimizer::Result<(Trial, f64)> {
-    // Sample configuration parameters using parameter definitions
-    let cache_size_mb = cache_size_mb_param.suggest(&mut trial)?;
-    let connection_pool_size = connection_pool_size_param.suggest(&mut trial)?;
-    let request_timeout_ms = request_timeout_ms_param.suggest(&mut trial)?;
-    let retry_count = retry_count_param.suggest(&mut trial)?;
-    let batch_size = batch_size_param.suggest(&mut trial)?;
-    let compression_level = compression_level_param.suggest(&mut trial)?;
-    let use_http2 = use_http2_param.suggest(&mut trial)?;
-    let load_balancing = load_balancing_param.suggest(&mut trial)?;
-
-    // Build configuration
-    let config = ServiceConfig {
-        cache_size_mb,
-        connection_pool_size,
-        request_timeout_ms,
-        retry_count,
-        batch_size,
-        compression_level,
-        use_http2,
-        load_balancing: load_balancing.to_string(),
-    };
-
-    // Evaluate (this is the async part)
-    let score = evaluate_service(&config).await;
-
-    // Return both the trial and the score
-    Ok((trial, score))
-}
-
-// ============================================================================
-// Helper Functions
-// ============================================================================
-
-/// Prints the results of the optimization.
-fn print_results(study: &Study<f64>, elapsed: Duration, n_trials: usize) {
-    println!("\n{}", "=".repeat(60));
-    println!("\nOptimization completed!");
-    println!("Total trials: {}", study.n_trials());
-    println!("Time elapsed: {elapsed:.2?}");
-
-    // Calculate speedup from parallelism
-    // Each trial takes ~50ms, so sequential would take n_trials * 50ms
-    let sequential_time = n_trials as f64 * 0.050;
-    let actual_time = elapsed.as_secs_f64();
-    println!(
-        "Effective parallelism: {:.1}x speedup",
-        sequential_time / actual_time
-    );
-}
-
-/// Prints the best configuration found.
-#[allow(clippy::too_many_arguments)]
-fn print_best_config(
-    study: &Study<f64>,
-    cache_size_mb_param: &IntParam,
-    connection_pool_size_param: &IntParam,
-    request_timeout_ms_param: &IntParam,
-    retry_count_param: &IntParam,
-    batch_size_param: &IntParam,
-    compression_level_param: &IntParam,
-    use_http2_param: &BoolParam,
-    load_balancing_param: &CategoricalParam<&str>,
-) -> optimizer::Result<()> {
-    let best = study.best_trial()?;
-
-    println!("\nBest configuration found:");
-    println!("  Score: {:.6}", best.value);
-    println!("\n  Parameters:");
-    println!(
-        "    cache_size_mb: {}",
-        best.get(cache_size_mb_param).unwrap()
-    );
-    println!(
-        "    connection_pool_size: {}",
-        best.get(connection_pool_size_param).unwrap()
-    );
-    println!(
-        "    request_timeout_ms: {}",
-        best.get(request_timeout_ms_param).unwrap()
-    );
-    println!("    retry_count: {}", best.get(retry_count_param).unwrap());
-    println!("    batch_size: {}", best.get(batch_size_param).unwrap());
-    println!(
-        "    compression_level: {}",
-        best.get(compression_level_param).unwrap()
-    );
-    println!("    use_http2: {}", best.get(use_http2_param).unwrap());
-    println!(
-        "    load_balancing: {}",
-        best.get(load_balancing_param).unwrap()
-    );
-
-    Ok(())
-}
-
-/// Prints the top N trials.
-fn print_top_trials(study: &Study<f64>, n: usize) {
-    println!("\nTop {n} trials:");
-
-    let mut trials = study.trials();
-    trials.sort_by(|a, b| a.value.partial_cmp(&b.value).unwrap());
-
-    for (i, trial) in trials.iter().take(n).enumerate() {
-        println!(
-            "  {}. Trial #{}: score = {:.6}",
-            i + 1,
-            trial.id,
-            trial.value
-        );
-    }
-}
-
-// ============================================================================
-// Main: Set up and run the async optimization
-// ============================================================================
-
-#[tokio::main]
-async fn main() -> optimizer::Result<()> {
-    println!("=== Async API Parameter Optimization Example ===\n");
-
-    // Step 1: Create a TPE sampler
-    let sampler = TpeSampler::builder()
-        .n_startup_trials(8)
-        .gamma(0.2)
-        .seed(123)
-        .build()
-        .expect("Failed to build TPE sampler");
-
-    // Step 2: Create a study to minimize the score
-    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
-
-    // Step 3: Define parameter search spaces
-    let cache_size_mb_param = IntParam::new(64, 1024).name("cache_size_mb").step(64);
-    let connection_pool_size_param = IntParam::new(10, 200).name("connection_pool_size").step(10);
-    let request_timeout_ms_param = IntParam::new(1000, 10000)
-        .name("request_timeout_ms")
-        .step(500);
-    let retry_count_param = IntParam::new(0, 5).name("retry_count");
-    let batch_size_param = IntParam::new(1, 256).name("batch_size").log_scale();
-    let compression_level_param = IntParam::new(0, 9).name("compression_level");
-    let use_http2_param = BoolParam::new().name("use_http2");
-    let load_balancing_param = CategoricalParam::new(vec![
-        "round_robin",
-        "least_connections",
-        "random",
-        "ip_hash",
-    ])
-    .name("load_balancing");
-
-    // Clone params for use after the closure moves them
-    let cache_size_mb_p = cache_size_mb_param.clone();
-    let connection_pool_size_p = connection_pool_size_param.clone();
-    let request_timeout_ms_p = request_timeout_ms_param.clone();
-    let retry_count_p = retry_count_param.clone();
-    let batch_size_p = batch_size_param.clone();
-    let compression_level_p = compression_level_param.clone();
-    let use_http2_p = use_http2_param.clone();
-    let load_balancing_p = load_balancing_param.clone();
-
-    // Step 4: Configure optimization
-    let n_trials = 40;
-    let concurrency = 4; // Run 4 trials in parallel
-
-    println!("Starting parallel optimization with {concurrency} concurrent evaluations...\n");
-
-    let start = Instant::now();
-
-    // Step 5: Run parallel async optimization
-    //
-    // optimize_parallel:
-    // - Runs up to `concurrency` trials simultaneously
-    // - Each trial calls the objective function
-    // - Uses a semaphore to limit concurrent evaluations
-    // - Collects results as trials complete
-    //
-    // The sampler gets access to trial history for informed sampling.
-    study
-        .optimize_parallel(n_trials, concurrency, move |trial| {
-            let cache_size_mb_param = cache_size_mb_param.clone();
-            let connection_pool_size_param = connection_pool_size_param.clone();
-            let request_timeout_ms_param = request_timeout_ms_param.clone();
-            let retry_count_param = retry_count_param.clone();
-            let batch_size_param = batch_size_param.clone();
-            let compression_level_param = compression_level_param.clone();
-            let use_http2_param = use_http2_param.clone();
-            let load_balancing_param = load_balancing_param.clone();
-            async move {
-                objective(
-                    trial,
-                    &cache_size_mb_param,
-                    &connection_pool_size_param,
-                    &request_timeout_ms_param,
-                    &retry_count_param,
-                    &batch_size_param,
-                    &compression_level_param,
-                    &use_http2_param,
-                    &load_balancing_param,
-                )
-                .await
-            }
-        })
-        .await?;
-
-    let elapsed = start.elapsed();
-
-    // Step 5: Print results
-    print_results(&study, elapsed, n_trials);
-    print_best_config(
-        &study,
-        &cache_size_mb_p,
-        &connection_pool_size_p,
-        &request_timeout_ms_p,
-        &retry_count_p,
-        &batch_size_p,
-        &compression_level_p,
-        &use_http2_p,
-        &load_balancing_p,
-    )?;
-    print_top_trials(&study, 5);
-
-    Ok(())
-}

examples/async_parallel.rs

@@ -0,0 +1,46 @@
+//! Async parallel optimization — evaluate multiple trials concurrently.
+//!
+//! Uses `optimize_parallel` with tokio to run several trials at once,
+//! reducing wall-clock time when the objective involves blocking work.
+//! Each sync closure is internally wrapped in `spawn_blocking`.
+//!
+//! Run with: `cargo run --example async_parallel --features async`
+
+use optimizer::prelude::*;
+
+#[tokio::main]
+async fn main() -> optimizer::Result<()> {
+    let study: Study<f64> = Study::minimize(TpeSampler::new());
+
+    let x = FloatParam::new(-5.0, 5.0).name("x");
+    let y = FloatParam::new(-5.0, 5.0).name("y");
+
+    let n_trials = 30;
+    let concurrency = 4;
+
+    println!("Running {n_trials} trials with {concurrency} concurrent workers...");
+
+    let xc = x.clone();
+    let yc = y.clone();
+    study
+        .optimize_parallel(
+            n_trials,
+            concurrency,
+            move |trial: &mut optimizer::Trial| {
+                let xv = xc.suggest(trial)?;
+                let yv = yc.suggest(trial)?;
+                Ok::<_, optimizer::Error>(xv * xv + yv * yv)
+            },
+        )
+        .await?;
+
+    let best = study.best_trial()?;
+    println!(
+        "Best: f({:.3}, {:.3}) = {:.6}",
+        best.get(&x).unwrap(),
+        best.get(&y).unwrap(),
+        best.value,
+    );
+
+    Ok(())
+}

examples/basic_optimization.rs

@@ -0,0 +1,32 @@
+//! Basic optimization example — the "hello world" of the optimizer crate.
+//!
+//! Minimizes a simple quadratic function f(x) = (x - 3)² using the default
+//! random sampler. No feature flags are required.
+//!
+//! Run with: `cargo run --example basic_optimization`
+
+use optimizer::prelude::*;
+
+fn main() {
+    // Create a study that minimizes the objective function.
+    // The default sampler is random; for smarter sampling, pass a TpeSampler.
+    let study: Study<f64> = Study::new(Direction::Minimize);
+
+    // Search for x in [-10, 10]. The optimizer will suggest values from this range.
+    let x = FloatParam::new(-10.0, 10.0).name("x");
+
+    // Run 50 trials, each evaluating f(x) = (x - 3)²
+    study
+        .optimize(50, |trial: &mut optimizer::Trial| {
+            let x_val = x.suggest(trial)?;
+            let value = (x_val - 3.0).powi(2);
+            Ok::<_, Error>(value)
+        })
+        .unwrap();
+
+    // Retrieve and display the best result
+    let best = study.best_trial().unwrap();
+    println!("Best trial #{}", best.id);
+    println!("  x     = {:.4}", best.get(&x).unwrap());
+    println!("  f(x)  = {:.4}", best.value);
+}

examples/benchmark_convergence.rs

@@ -1,124 +0,0 @@
-use std::ops::ControlFlow;
-use std::time::Instant;
-
-use optimizer::parameter::Parameter;
-use optimizer::sampler::random::RandomSampler;
-use optimizer::sampler::tpe::TpeSampler;
-use optimizer::{FloatParam, Study};
-
-/// Standard optimization test functions.
-mod functions {
-    pub fn sphere(x: &[f64]) -> f64 {
-        x.iter().map(|xi| xi * xi).sum()
-    }
-
-    pub fn rosenbrock(x: &[f64]) -> f64 {
-        x.windows(2)
-            .map(|w| 100.0 * (w[1] - w[0] * w[0]).powi(2) + (1.0 - w[0]).powi(2))
-            .sum()
-    }
-
-    pub fn rastrigin(x: &[f64]) -> f64 {
-        let n = x.len() as f64;
-        10.0 * n
-            + x.iter()
-                .map(|xi| xi * xi - 10.0 * (2.0 * std::f64::consts::PI * xi).cos())
-                .sum::<f64>()
-    }
-}
-
-fn run_convergence(
-    name: &str,
-    sampler_name: &str,
-    study: Study<f64>,
-    params: &[FloatParam],
-    objective: fn(&[f64]) -> f64,
-    n_trials: usize,
-) {
-    let start = Instant::now();
-
-    study
-        .optimize_with_callback(
-            n_trials,
-            |trial| {
-                let x: Vec<f64> = params
-                    .iter()
-                    .map(|p| p.suggest(trial))
-                    .collect::<Result<_, _>>()
-                    .unwrap();
-                Ok::<_, optimizer::Error>(objective(&x))
-            },
-            |study, _trial| {
-                let elapsed = start.elapsed().as_millis();
-                let best = study.best_value().unwrap();
-                let n = study.n_trials();
-                println!("{n},{best},{elapsed},{sampler_name},{name}");
-                ControlFlow::Continue(())
-            },
-        )
-        .unwrap();
-}
-
-fn main() {
-    println!("trial,best_value,elapsed_ms,sampler,function");
-
-    let dims = 5;
-    let params: Vec<FloatParam> = (0..dims)
-        .map(|i| FloatParam::new(-5.0, 5.0).name(format!("x{i}")))
-        .collect();
-    let n_trials = 200;
-
-    // Sphere: Random vs TPE
-    run_convergence(
-        "sphere_5d",
-        "random",
-        Study::minimize(RandomSampler::with_seed(1)),
-        &params,
-        functions::sphere,
-        n_trials,
-    );
-    run_convergence(
-        "sphere_5d",
-        "tpe",
-        Study::minimize(TpeSampler::builder().seed(1).build().unwrap()),
-        &params,
-        functions::sphere,
-        n_trials,
-    );
-
-    // Rosenbrock: Random vs TPE
-    run_convergence(
-        "rosenbrock_5d",
-        "random",
-        Study::minimize(RandomSampler::with_seed(2)),
-        &params,
-        functions::rosenbrock,
-        n_trials,
-    );
-    run_convergence(
-        "rosenbrock_5d",
-        "tpe",
-        Study::minimize(TpeSampler::builder().seed(2).build().unwrap()),
-        &params,
-        functions::rosenbrock,
-        n_trials,
-    );
-
-    // Rastrigin: Random vs TPE
-    run_convergence(
-        "rastrigin_5d",
-        "random",
-        Study::minimize(RandomSampler::with_seed(3)),
-        &params,
-        functions::rastrigin,
-        n_trials,
-    );
-    run_convergence(
-        "rastrigin_5d",
-        "tpe",
-        Study::minimize(TpeSampler::builder().seed(3).build().unwrap()),
-        &params,
-        functions::rastrigin,
-        n_trials,
-    );
-}

examples/early_stopping.rs

@@ -0,0 +1,58 @@
+//! Early stopping — halt an entire study once a target is reached.
+//!
+//! Implements the [`Objective`] trait on a custom struct and uses the
+//! [`after_trial`](Objective::after_trial) hook to return
+//! `ControlFlow::Break(())` when the best value drops below a threshold.
+//!
+//! Run with: `cargo run --example early_stopping`
+
+use std::ops::ControlFlow;
+
+use optimizer::prelude::*;
+
+/// An objective that minimises `(x - 3)^2` and stops early once the
+/// value drops below `target`.
+struct EarlyStopObjective {
+    x: FloatParam,
+    target: f64,
+}
+
+impl Objective<f64> for EarlyStopObjective {
+    type Error = Error;
+
+    fn evaluate(&self, trial: &mut Trial) -> Result<f64> {
+        let v = self.x.suggest(trial)?;
+        Ok((v - 3.0).powi(2))
+    }
+
+    fn after_trial(&self, _study: &Study<f64>, trial: &CompletedTrial<f64>) -> ControlFlow<()> {
+        if trial.value < self.target {
+            println!("Target {} reached at trial #{}", self.target, trial.id);
+            ControlFlow::Break(())
+        } else {
+            ControlFlow::Continue(())
+        }
+    }
+}
+
+fn main() -> optimizer::Result<()> {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+    let x = FloatParam::new(-10.0, 10.0).name("x");
+
+    let objective = EarlyStopObjective {
+        x: x.clone(),
+        target: 0.01,
+    };
+
+    study.optimize(100, objective)?;
+
+    let best = study.best_trial()?;
+    println!(
+        "Stopped after {} trials — best f({:.4}) = {:.6}",
+        study.n_trials(),
+        best.get(&x).unwrap(),
+        best.value,
+    );
+
+    Ok(())
+}

examples/journal_storage.rs

@@ -0,0 +1,68 @@
+//! Journal storage — persist trials to disk and resume later.
+//!
+//! `JournalStorage` writes every trial to a JSONL file so that a study can
+//! be resumed after a crash or across separate runs.
+//!
+//! Run with: `cargo run --example journal_storage --features journal`
+
+use optimizer::prelude::*;
+
+fn main() -> optimizer::Result<()> {
+    let path = std::env::temp_dir().join("optimizer_journal_example.jsonl");
+
+    // Clean up from any previous run
+    let _ = std::fs::remove_file(&path);
+
+    let x = FloatParam::new(-5.0, 5.0).name("x");
+
+    // --- First run: optimize 20 trials and persist to disk ---
+    {
+        let storage = JournalStorage::<f64>::new(&path);
+        let study: Study<f64> = Study::builder()
+            .minimize()
+            .sampler(TpeSampler::new())
+            .storage(storage)
+            .build();
+
+        study.optimize(20, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            Ok::<_, optimizer::Error>(xv * xv)
+        })?;
+
+        println!(
+            "First run: {} trials saved to {}",
+            study.n_trials(),
+            path.display(),
+        );
+    }
+
+    // --- Second run: resume from the journal file ---
+    {
+        let storage = JournalStorage::<f64>::open(&path)?;
+        let study: Study<f64> = Study::builder()
+            .minimize()
+            .sampler(TpeSampler::new())
+            .storage(storage)
+            .build();
+
+        let before = study.n_trials();
+        study.optimize(10, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            Ok::<_, optimizer::Error>(xv * xv)
+        })?;
+
+        let best = study.best_trial()?;
+        println!(
+            "Resumed: {} → {} trials, best f({:.4}) = {:.6}",
+            before,
+            study.n_trials(),
+            best.get(&x).unwrap(),
+            best.value,
+        );
+    }
+
+    // Clean up
+    let _ = std::fs::remove_file(&path);
+
+    Ok(())
+}

examples/ml_hyperparameter_tuning.rs

@@ -1,275 +0,0 @@
-//! Machine Learning Hyperparameter Tuning Example
-//!
-//! This example shows how to use the optimizer library to find the best
-//! hyperparameters for a machine learning model. We simulate a gradient
-//! boosting model (like XGBoost or LightGBM) and search for optimal settings.
-//!
-//! # Key Concepts Demonstrated
-//!
-//! - Creating a Study with a TPE (Tree-Parzen Estimator) sampler
-//! - Defining an objective function that the optimizer will minimize
-//! - Using different parameter types: floats, integers, log-scale, stepped
-//! - Using callbacks to monitor progress and implement early stopping
-//!
-//! # How It Works
-//!
-//! 1. Create a `Study` - this manages the optimization process
-//! 2. Define an objective function that takes a `Trial` and returns a score
-//! 3. Inside the objective, use `trial.suggest_*()` to sample parameters
-//! 4. The optimizer runs many trials, learning which parameter regions work best
-//! 5. After optimization, retrieve the best parameters found
-//!
-//! Run with: `cargo run --example ml_hyperparameter_tuning`
-
-use std::ops::ControlFlow;
-
-use optimizer::prelude::*;
-
-// ============================================================================
-// Configuration: Hyperparameters we want to tune
-// ============================================================================
-
-/// Holds all the hyperparameters for our model.
-///
-/// In a real application, you would pass these to your ML framework
-/// (e.g., XGBoost, LightGBM, scikit-learn).
-struct ModelConfig {
-    learning_rate: f64,
-    max_depth: i64,
-    n_estimators: i64,
-    subsample: f64,
-    colsample_bytree: f64,
-    min_child_weight: i64,
-    reg_alpha: f64,
-    reg_lambda: f64,
-}
-
-// ============================================================================
-// Objective Function: What we want to optimize
-// ============================================================================
-
-/// Simulates training a model and returns the validation loss.
-///
-/// In a real scenario, this function would:
-/// 1. Create a model with the given hyperparameters
-/// 2. Train it on your training data
-/// 3. Evaluate it on validation data
-/// 4. Return the validation metric (e.g., RMSE, log loss, accuracy)
-///
-/// The optimizer will try to MINIMIZE this value (we set Direction::Minimize).
-#[allow(clippy::too_many_arguments)]
-fn evaluate_model(config: &ModelConfig) -> f64 {
-    // Simulated optimal hyperparameters:
-    // learning_rate ~ 0.05, max_depth ~ 6, n_estimators ~ 200
-    // subsample ~ 0.8, colsample_bytree ~ 0.8, min_child_weight ~ 3
-    // reg_alpha ~ 0.1, reg_lambda ~ 1.0
-
-    let mut loss = 0.15; // Base loss
-
-    // Each term penalizes deviation from the optimal value
-    loss += (config.learning_rate - 0.05).powi(2) * 100.0;
-    loss += ((config.max_depth - 6) as f64).powi(2) * 0.01;
-    loss += ((config.n_estimators - 200) as f64).powi(2) * 0.00001;
-    loss += (config.subsample - 0.8).powi(2) * 10.0;
-    loss += (config.colsample_bytree - 0.8).powi(2) * 10.0;
-    loss += ((config.min_child_weight - 3) as f64).powi(2) * 0.05;
-    loss += (config.reg_alpha - 0.1).powi(2) * 5.0;
-    loss += (config.reg_lambda - 1.0).powi(2) * 2.0;
-
-    // Add some noise to simulate real-world variability
-    let noise = (config.learning_rate * 1000.0).sin() * 0.01;
-
-    loss + noise
-}
-
-/// The objective function that the optimizer calls for each trial.
-///
-/// This function:
-/// 1. Uses parameter definitions passed as arguments
-/// 2. Builds a model configuration from the suggested values
-/// 3. Evaluates the model and returns the loss
-///
-/// The optimizer learns from the results to suggest better parameters
-/// in future trials.
-#[allow(clippy::too_many_arguments)]
-fn objective(
-    trial: &mut Trial,
-    learning_rate_param: &FloatParam,
-    max_depth_param: &IntParam,
-    n_estimators_param: &IntParam,
-    subsample_param: &FloatParam,
-    colsample_bytree_param: &FloatParam,
-    min_child_weight_param: &IntParam,
-    reg_alpha_param: &FloatParam,
-    reg_lambda_param: &FloatParam,
-) -> optimizer::Result<f64> {
-    let learning_rate = learning_rate_param.suggest(trial)?;
-    let max_depth = max_depth_param.suggest(trial)?;
-    let n_estimators = n_estimators_param.suggest(trial)?;
-    let subsample = subsample_param.suggest(trial)?;
-    let colsample_bytree = colsample_bytree_param.suggest(trial)?;
-    let min_child_weight = min_child_weight_param.suggest(trial)?;
-    let reg_alpha = reg_alpha_param.suggest(trial)?;
-    let reg_lambda = reg_lambda_param.suggest(trial)?;
-
-    // Build configuration and evaluate
-    let config = ModelConfig {
-        learning_rate,
-        max_depth,
-        n_estimators,
-        subsample,
-        colsample_bytree,
-        min_child_weight,
-        reg_alpha,
-        reg_lambda,
-    };
-
-    let loss = evaluate_model(&config);
-
-    Ok(loss)
-}
-
-// ============================================================================
-// Callback Function: Monitor progress and implement early stopping
-// ============================================================================
-
-/// Called after each successful trial completes.
-///
-/// Use callbacks to:
-/// - Log progress to console or file
-/// - Save checkpoints
-/// - Implement early stopping when a good solution is found
-/// - Track metrics over time
-///
-/// Return `ControlFlow::Continue(())` to keep optimizing.
-/// Return `ControlFlow::Break(())` to stop early.
-fn on_trial_complete(study: &Study<f64>, trial: &CompletedTrial<f64>) -> ControlFlow<()> {
-    // Print trial number and objective value
-    print!("{:>5} ", study.n_trials());
-    for value in trial.params.values() {
-        print!("{value:>12} ");
-    }
-    println!("{:>12.6}", trial.value);
-
-    // Early stopping: if we find an excellent solution, stop early
-    if trial.value < 0.16 {
-        println!("\nEarly stopping: found excellent solution!");
-        return ControlFlow::Break(());
-    }
-
-    ControlFlow::Continue(())
-}
-
-// ============================================================================
-// Main: Set up and run the optimization
-// ============================================================================
-
-fn main() -> optimizer::Result<()> {
-    println!("=== ML Hyperparameter Tuning Example ===\n");
-
-    // Step 1: Create a sampler
-    //
-    // TPE (Tree-Parzen Estimator) is a Bayesian optimization algorithm.
-    // It learns from previous trials to suggest better parameters.
-    // - n_startup_trials: Number of random trials before TPE kicks in
-    // - gamma: What fraction of trials are considered "good" (lower = more selective)
-    // - seed: For reproducibility
-    let sampler = TpeSampler::builder()
-        .n_startup_trials(10)
-        .gamma(0.25)
-        .seed(42)
-        .build()
-        .expect("Failed to build TPE sampler");
-
-    // Step 2: Create a study
-    //
-    // The study manages the optimization process. We want to MINIMIZE
-    // the loss (lower is better). Use Direction::Maximize for metrics
-    // where higher is better (like accuracy).
-    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
-
-    // Print header
-    println!("Starting hyperparameter optimization...\n");
-    println!(
-        "{:>5} {:>12} (parameters...)    {:>12}",
-        "Trial", "Params", "Loss"
-    );
-    println!("{}", "-".repeat(60));
-
-    // Step 3: Define parameter search spaces
-    let learning_rate_param = FloatParam::new(0.001, 0.3)
-        .name("learning_rate")
-        .log_scale();
-    let max_depth_param = IntParam::new(3, 12).name("max_depth");
-    let n_estimators_param = IntParam::new(50, 500).name("n_estimators").step(50);
-    let subsample_param = FloatParam::new(0.5, 1.0).name("subsample");
-    let colsample_bytree_param = FloatParam::new(0.5, 1.0).name("colsample_bytree");
-    let min_child_weight_param = IntParam::new(1, 10).name("min_child_weight");
-    let reg_alpha_param = FloatParam::new(1e-3, 10.0).name("reg_alpha").log_scale();
-    let reg_lambda_param = FloatParam::new(1e-3, 10.0).name("reg_lambda").log_scale();
-
-    // Step 4: Run optimization
-    //
-    // optimize_with_callback runs the objective function for up to
-    // n_trials iterations. After each trial, it calls the callback.
-    // The sampler gets access to trial history for informed sampling.
-    let n_trials = 50;
-
-    study.optimize_with_callback(
-        n_trials,
-        |trial| {
-            objective(
-                trial,
-                &learning_rate_param,
-                &max_depth_param,
-                &n_estimators_param,
-                &subsample_param,
-                &colsample_bytree_param,
-                &min_child_weight_param,
-                &reg_alpha_param,
-                &reg_lambda_param,
-            )
-        },
-        on_trial_complete,
-    )?;
-
-    // Step 4: Get the best result
-    println!("\n{}", "=".repeat(110));
-    println!("\nOptimization completed!");
-    println!("Total trials: {}", study.n_trials());
-
-    let best = study.best_trial()?;
-    println!("\nBest trial:");
-    println!("  Loss: {:.6}", best.value);
-    println!("  Parameters:");
-    println!(
-        "    learning_rate: {:.6}",
-        best.get(&learning_rate_param).unwrap()
-    );
-    println!("    max_depth: {}", best.get(&max_depth_param).unwrap());
-    println!(
-        "    n_estimators: {}",
-        best.get(&n_estimators_param).unwrap()
-    );
-    println!("    subsample: {:.6}", best.get(&subsample_param).unwrap());
-    println!(
-        "    colsample_bytree: {:.6}",
-        best.get(&colsample_bytree_param).unwrap()
-    );
-    println!(
-        "    min_child_weight: {}",
-        best.get(&min_child_weight_param).unwrap()
-    );
-    println!("    reg_alpha: {:.6}", best.get(&reg_alpha_param).unwrap());
-    println!(
-        "    reg_lambda: {:.6}",
-        best.get(&reg_lambda_param).unwrap()
-    );
-
-    // Step 5: Use the best parameters (in a real app)
-    //
-    // Now you would take best.params and use them to train your final model
-    // on the full dataset.
-
-    Ok(())
-}

examples/multi_objective.rs

@@ -0,0 +1,49 @@
+//! Multi-objective optimization — optimize competing objectives simultaneously.
+//!
+//! `MultiObjectiveStudy` returns the Pareto front: the set of solutions where
+//! no objective can be improved without worsening another.
+//!
+//! Run with: `cargo run --example multi_objective`
+
+use optimizer::multi_objective::MultiObjectiveStudy;
+use optimizer::prelude::*;
+
+fn main() -> optimizer::Result<()> {
+    let study = MultiObjectiveStudy::new(vec![Direction::Minimize, Direction::Minimize]);
+
+    let x = FloatParam::new(0.0, 1.0).name("x");
+
+    // Classic bi-objective: f1(x) = x², f2(x) = (x-1)²
+    // The Pareto front is the curve where improving f1 worsens f2.
+    study.optimize(50, |trial: &mut optimizer::Trial| {
+        let xv = x.suggest(trial)?;
+        let f1 = xv * xv;
+        let f2 = (xv - 1.0) * (xv - 1.0);
+        Ok::<_, optimizer::Error>(vec![f1, f2])
+    })?;
+
+    let front = study.pareto_front();
+    println!(
+        "Ran {} trials, Pareto front has {} solutions:",
+        study.n_trials(),
+        front.len(),
+    );
+
+    let mut sorted = front.clone();
+    sorted.sort_by(|a, b| a.values[0].partial_cmp(&b.values[0]).unwrap());
+
+    for (i, trial) in sorted.iter().take(5).enumerate() {
+        println!(
+            "  {}: x={:.3}, f1={:.4}, f2={:.4}",
+            i + 1,
+            trial.get(&x).unwrap(),
+            trial.values[0],
+            trial.values[1],
+        );
+    }
+    if sorted.len() > 5 {
+        println!("  ... and {} more", sorted.len() - 5);
+    }
+
+    Ok(())
+}

examples/parameter_api.rs

@@ -1,56 +0,0 @@
-use optimizer::prelude::*;
-use optimizer_derive::Categorical;
-
-#[derive(Clone, Debug, Categorical)]
-enum Activation {
-    Relu,
-    Sigmoid,
-    Tanh,
-}
-
-fn main() {
-    let study: Study<f64> = Study::new(Direction::Minimize);
-
-    // Define parameters outside the objective function
-    let lr_param = FloatParam::new(1e-5, 1e-1).name("lr").log_scale();
-    let n_layers_param = IntParam::new(1, 5).name("n_layers");
-    let units_param = IntParam::new(32, 512).name("units").step(32);
-    let optimizer_param = CategoricalParam::new(vec!["sgd", "adam", "rmsprop"]).name("optimizer");
-    let activation_param = EnumParam::<Activation>::new().name("activation");
-    let batch_size_param = IntParam::new(16, 256).name("batch_size").log_scale();
-    let use_dropout_param = BoolParam::new().name("use_dropout");
-
-    study
-        .optimize(20, |trial| {
-            let lr = lr_param.suggest(trial)?;
-            let n_layers = n_layers_param.suggest(trial)?;
-            let units = units_param.suggest(trial)?;
-            let optimizer = optimizer_param.suggest(trial)?;
-            let use_dropout = use_dropout_param.suggest(trial)?;
-            let activation = activation_param.suggest(trial)?;
-            let batch_size = batch_size_param.suggest(trial)?;
-
-            // Simulate a loss function
-            let loss = lr * (n_layers as f64) + (units as f64) * 0.001
-                - if use_dropout { 0.1 } else { 0.0 };
-
-            println!(
-                "Trial {}: lr={lr:.6}, layers={n_layers}, units={units}, opt={optimizer}, \
-                 dropout={use_dropout}, activation={activation:?}, batch={batch_size} -> loss={loss:.4}",
-                trial.id()
-            );
-
-            Ok::<_, Error>(loss)
-        })
-        .unwrap();
-
-    let best = study.best_trial().unwrap();
-    println!("\nBest trial: value={:.4}", best.value);
-    println!("  lr: {:.6}", best.get(&lr_param).unwrap());
-    println!("  n_layers: {}", best.get(&n_layers_param).unwrap());
-    println!("  units: {}", best.get(&units_param).unwrap());
-    println!("  optimizer: {}", best.get(&optimizer_param).unwrap());
-    println!("  activation: {:?}", best.get(&activation_param).unwrap());
-    println!("  batch_size: {}", best.get(&batch_size_param).unwrap());
-    println!("  use_dropout: {}", best.get(&use_dropout_param).unwrap());
-}

examples/parameter_types.rs

@@ -0,0 +1,73 @@
+//! Parameter types example — demonstrates all five parameter types and the derive feature.
+//!
+//! Shows `FloatParam`, `IntParam`, `CategoricalParam`, `BoolParam`, and `EnumParam`
+//! with `.name()` labels, `#[derive(Categorical)]` for enums, and typed access
+//! to results via `CompletedTrial::get()`.
+//!
+//! Run with: `cargo run --example parameter_types --features derive`
+
+use optimizer::prelude::*;
+use optimizer_derive::Categorical;
+
+/// Activation functions — `#[derive(Categorical)]` auto-generates the
+/// `Categorical` trait, mapping each variant to a sequential index.
+#[derive(Clone, Debug, Categorical)]
+enum Activation {
+    Relu,
+    Sigmoid,
+    Tanh,
+    Gelu,
+}
+
+fn main() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+
+    // --- Define one of each parameter type, each with a human-readable .name() ---
+
+    // Float: learning rate on a log scale (common for ML hyperparameters)
+    let lr = FloatParam::new(1e-5, 1e-1).log_scale().name("lr");
+
+    // Int: number of hidden layers (stepped by 1, the default)
+    let n_layers = IntParam::new(1, 5).name("n_layers");
+
+    // Categorical: optimizer algorithm chosen from a list of strings
+    let optimizer = CategoricalParam::new(vec!["sgd", "adam", "rmsprop"]).name("optimizer");
+
+    // Bool: whether to apply dropout
+    let use_dropout = BoolParam::new().name("use_dropout");
+
+    // Enum: activation function — uses #[derive(Categorical)] above
+    let activation = EnumParam::<Activation>::new().name("activation");
+
+    // --- Run the optimization ---
+    study
+        .optimize(30, |trial: &mut optimizer::Trial| {
+            let lr_val = lr.suggest(trial)?;
+            let layers = n_layers.suggest(trial)?;
+            let opt = optimizer.suggest(trial)?;
+            let dropout = use_dropout.suggest(trial)?;
+            let act = activation.suggest(trial)?;
+
+            // Simulated loss that depends on all parameters
+            let loss = lr_val * f64::from(layers as i32)
+                + if opt == "adam" { -0.05 } else { 0.0 }
+                + if dropout { -0.02 } else { 0.0 }
+                + match act {
+                    Activation::Gelu => -0.03,
+                    Activation::Relu => -0.01,
+                    _ => 0.0,
+                };
+
+            Ok::<_, Error>(loss)
+        })
+        .unwrap();
+
+    // --- Retrieve best trial and read back each parameter with typed .get() ---
+    let best = study.best_trial().unwrap();
+    println!("Best trial #{} — loss = {:.6}", best.id, best.value);
+    println!("  lr         = {:.6}", best.get(&lr).unwrap());
+    println!("  n_layers   = {}", best.get(&n_layers).unwrap());
+    println!("  optimizer  = {}", best.get(&optimizer).unwrap());
+    println!("  use_dropout = {}", best.get(&use_dropout).unwrap());
+    println!("  activation = {:?}", best.get(&activation).unwrap());
+}

examples/pruning.rs

@@ -0,0 +1,67 @@
+//! Trial pruning — stop unpromising trials early with `MedianPruner`.
+//!
+//! When your objective involves an iterative loop (e.g. training epochs),
+//! the pruner compares intermediate values across trials and kills the
+//! ones that fall below the median — saving compute on bad configurations.
+//!
+//! Run with: `cargo run --example pruning`
+
+use optimizer::prelude::*;
+
+fn main() -> optimizer::Result<()> {
+    // MedianPruner prunes trials whose intermediate value falls below the
+    // median of previously completed trials at the same step.
+    let study: Study<f64> = Study::builder()
+        .minimize()
+        .sampler(RandomSampler::with_seed(42))
+        .pruner(
+            MedianPruner::new(Direction::Minimize)
+                .n_warmup_steps(3) // run at least 3 epochs before pruning
+                .n_min_trials(3), // need 3 completed trials before pruning kicks in
+        )
+        .build();
+
+    let lr = FloatParam::new(1e-4, 1.0).name("learning_rate");
+    let momentum = FloatParam::new(0.0, 0.99).name("momentum");
+
+    let n_epochs: u64 = 20;
+
+    study.optimize(30, |trial: &mut optimizer::Trial| {
+        let lr_val = lr.suggest(trial)?;
+        let mom = momentum.suggest(trial)?;
+
+        // Simulated training loop — good hyperparameters converge to low loss,
+        // bad ones plateau high, giving the pruner something to cut.
+        let mut loss = 1.0;
+        for epoch in 0..n_epochs {
+            let lr_penalty = (lr_val.log10() - 0.01_f64.log10()).powi(2);
+            let mom_penalty = (mom - 0.8).powi(2);
+            let base_loss = 0.02 + 0.05 * lr_penalty + 1.5 * mom_penalty;
+            let progress = (epoch as f64 + 1.0) / n_epochs as f64;
+            loss = base_loss + (1.0 - base_loss) * (-3.5 * progress).exp();
+
+            // Report intermediate value so the pruner can evaluate this trial.
+            trial.report(epoch, loss);
+
+            // Check whether the pruner recommends stopping early.
+            if trial.should_prune() {
+                Err(TrialPruned)?;
+            }
+        }
+
+        Ok::<_, Error>(loss)
+    })?;
+
+    // --- Results ---
+    let best = study.best_trial()?;
+    println!(
+        "Recorded {} trials ({} pruned)",
+        study.n_trials(),
+        study.n_pruned_trials()
+    );
+    println!("Best trial #{}: loss = {:.6}", best.id, best.value);
+    println!("  learning_rate = {:.6}", best.get(&lr).unwrap());
+    println!("  momentum      = {:.4}", best.get(&momentum).unwrap());
+
+    Ok(())
+}

examples/sampler_comparison.rs

@@ -0,0 +1,78 @@
+//! Sampler comparison example — benchmarks Random, TPE, and Grid samplers on the same problem.
+//!
+//! Runs the Sphere function f(x, y) = x² + y² with each sampler and compares the best
+//! value found. This shows how sampler choice affects optimization quality.
+//!
+//! Run with: `cargo run --example sampler_comparison`
+
+use optimizer::prelude::*;
+
+/// Shared objective function: Sphere function with global minimum at (0, 0).
+/// Simple enough to solve well, but 2-D so samplers have room to differ.
+fn sphere(x: f64, y: f64) -> f64 {
+    x.powi(2) + y.powi(2)
+}
+
+/// Run an optimization study and return the best value found.
+fn run_study(study: Study<f64>, n_trials: usize) -> f64 {
+    // Use asymmetric ranges so the Grid sampler tracks each parameter independently.
+    let x = FloatParam::new(-5.0, 5.0).name("x");
+    let y = FloatParam::new(-3.0, 3.0).name("y");
+
+    study
+        .optimize(n_trials, |trial: &mut optimizer::Trial| {
+            let x_val = x.suggest(trial)?;
+            let y_val = y.suggest(trial)?;
+            Ok::<_, Error>(sphere(x_val, y_val))
+        })
+        .unwrap();
+
+    let best = study.best_trial().unwrap();
+    println!(
+        "  Best trial #{:>3}: x = {:>7.4}, y = {:>7.4}, f(x,y) = {:.6}",
+        best.id,
+        best.get(&x).unwrap(),
+        best.get(&y).unwrap(),
+        best.value,
+    );
+    best.value
+}
+
+fn main() {
+    let n_trials: usize = 100;
+    println!("Comparing samplers on Sphere(x, y) = x² + y²  ({n_trials} trials each)");
+    println!();
+
+    // --- Random sampler (baseline) ---
+    // Pure random search: samples uniformly at random. Fast but not guided.
+    println!("1. Random sampler:");
+    let random_best = run_study(Study::minimize(RandomSampler::with_seed(42)), n_trials);
+
+    // --- TPE sampler (Bayesian) ---
+    // Tree-structured Parzen Estimator: builds a probabilistic model of good vs bad
+    // regions and focuses sampling where improvements are likely.
+    println!("\n2. TPE sampler (Bayesian):");
+    let tpe = TpeSampler::builder()
+        .n_startup_trials(10) // random exploration for the first 10 trials
+        .n_ei_candidates(24) // candidates evaluated per Expected Improvement step
+        .gamma(0.25) // top 25% of trials define the "good" distribution
+        .seed(42)
+        .build()
+        .unwrap();
+    let tpe_best = run_study(Study::minimize(tpe), n_trials);
+
+    // --- Grid sampler (exhaustive) ---
+    // Evaluates evenly spaced grid points. Each parameter gets its own grid that
+    // is sampled in order, so n_points_per_param must be >= n_trials.
+    println!("\n3. Grid sampler (exhaustive):");
+    let grid = GridSampler::builder()
+        .n_points_per_param(n_trials) // one grid point per trial per parameter
+        .build();
+    let grid_best = run_study(Study::minimize(grid), n_trials);
+
+    // --- Summary ---
+    println!("\n--- Summary ---");
+    println!("  Random : {random_best:.6}");
+    println!("  TPE    : {tpe_best:.6}");
+    println!("  Grid   : {grid_best:.6}");
+}

optimizer-derive/Cargo.toml

@@ -2,7 +2,7 @@
 name = "optimizer-derive"
 version = "0.1.0"
 edition = "2024"
-rust-version = "1.88"
+rust-version = "1.89"
 license = "MIT"
 description = "Derive macros for the optimizer crate"
 repository = "https://github.com/raimannma/rust-optimizer"

optimizer-derive/src/lib.rs

@@ -4,13 +4,13 @@ use syn::{Data, DeriveInput, Fields, parse_macro_input};
 
 /// Derive macro for the `Categorical` trait on fieldless enums.
 ///
-/// Generates an implementation of `optimizer::Categorical` that maps
+/// Generates an implementation of `optimizer::parameter::Categorical` that maps
 /// enum variants to/from sequential indices.
 ///
 /// # Example
 ///
 /// ```ignore
-/// use optimizer::Categorical;
+/// use optimizer::parameter::Categorical;
 ///
 /// #[derive(Clone, Categorical)]
 /// enum Color {
@@ -49,7 +49,7 @@ pub fn derive_categorical(input: TokenStream) -> TokenStream {
     let (impl_generics, ty_generics, where_clause) = input.generics.split_for_impl();
 
     let expanded = quote! {
-        impl #impl_generics optimizer::Categorical for #name #ty_generics #where_clause {
+        impl #impl_generics optimizer::parameter::Categorical for #name #ty_generics #where_clause {
             const N_CHOICES: usize = #n_choices;
 
             fn from_index(index: usize) -> Self {

rustfmt.toml

@@ -1,8 +1,8 @@
 # Nightly Features
-imports_granularity = "Module"
-group_imports = "StdExternalCrate"
+#imports_granularity = "Module"
+#group_imports = "StdExternalCrate"
 #format_strings = true
-format_code_in_doc_comments = true
+#format_code_in_doc_comments = true
 
 # Stable Features
 merge_derives = true

src/error.rs

@@ -1,6 +1,22 @@
+//! Error types for the optimizer crate.
+//!
+//! All fallible operations in the crate return [`Result<T>`], which is an
+//! alias for `core::result::Result<T, Error>`. The [`Error`] enum covers
+//! parameter validation, sampling conflicts, pruning signals, and
+//! feature-gated I/O errors.
+
+/// Errors returned by optimizer operations.
+///
+/// Most variants are returned during parameter validation or trial
+/// management. The [`TrialPruned`](Error::TrialPruned) variant has special
+/// significance — it signals early stopping and is typically raised via
+/// the [`TrialPruned`](super::TrialPruned) convenience type.
+#[non_exhaustive]
 #[derive(Debug, thiserror::Error)]
 pub enum Error {
-    /// Returned when the lower bound is greater than the upper bound.
+    /// The lower bound exceeds the upper bound in a
+    /// [`FloatParam`](crate::parameter::FloatParam) or
+    /// [`IntParam`](crate::parameter::IntParam).
     #[error("invalid bounds: low ({low}) must be less than or equal to high ({high})")]
     InvalidBounds {
         /// The lower bound value.
@@ -9,19 +25,22 @@ pub enum Error {
         high: f64,
     },
 
-    /// Returned when log scale is used with non-positive bounds.
+    /// Log-scale is enabled but the lower bound is not positive (float) or
+    /// is less than 1 (integer).
     #[error("invalid log bounds: low must be positive for log scale")]
     InvalidLogBounds,
 
-    /// Returned when step size is not positive.
+    /// The step size provided to a parameter is not positive.
     #[error("invalid step: step must be positive")]
     InvalidStep,
 
-    /// Returned when categorical choices are empty.
+    /// A [`CategoricalParam`](crate::parameter::CategoricalParam) was created
+    /// with an empty choices vector.
     #[error("categorical choices cannot be empty")]
     EmptyChoices,
 
-    /// Returned when a parameter is suggested with a different configuration.
+    /// The same [`ParamId`](crate::parameter::ParamId) was suggested twice
+    /// with a different distribution configuration.
     #[error("parameter conflict for '{name}': {reason}")]
     ParameterConflict {
         /// The name of the conflicting parameter.
@@ -30,27 +49,29 @@ pub enum Error {
         reason: String,
     },
 
-    /// Returned when requesting the best trial but no trials have completed.
+    /// [`Study::best_trial`](crate::Study::best_trial) or similar was called
+    /// before any trial completed successfully.
     #[error("no completed trials available")]
     NoCompletedTrials,
 
-    /// Returned when gamma is not in the valid range (0.0, 1.0).
+    /// The gamma value for TPE sampling is outside the open interval (0, 1).
     #[error("invalid gamma: {0} must be in (0.0, 1.0)")]
     InvalidGamma(f64),
 
-    /// Returned when bandwidth is not positive.
+    /// A KDE bandwidth value is not positive.
     #[error("invalid bandwidth: {0} must be positive")]
     InvalidBandwidth(f64),
 
-    /// Returned when KDE is created with empty samples.
+    /// A kernel density estimator was constructed with no samples.
     #[error("KDE requires at least one sample")]
     EmptySamples,
 
-    /// Returned when multivariate KDE samples have zero dimensions.
+    /// Multivariate KDE samples have zero dimensions.
     #[error("multivariate KDE samples must have at least one dimension")]
     ZeroDimensions,
 
-    /// Returned when multivariate KDE samples have inconsistent dimensions.
+    /// A sample in the multivariate KDE has a different number of dimensions
+    /// than the first sample.
     #[error(
         "dimension mismatch: expected {expected} dimensions but sample {sample_index} has {got}"
     )]
@@ -63,7 +84,7 @@ pub enum Error {
         sample_index: usize,
     },
 
-    /// Returned when bandwidth vector length doesn't match the number of dimensions.
+    /// The bandwidth vector length does not match the number of KDE dimensions.
     #[error("bandwidth dimension mismatch: expected {expected} bandwidths but got {got}")]
     BandwidthDimensionMismatch {
         /// The expected number of bandwidths.
@@ -72,20 +93,40 @@ pub enum Error {
         got: usize,
     },
 
-    /// Returned when a trial is pruned (stopped early by the objective function).
+    /// The objective signalled that this trial should be pruned (stopped
+    /// early). Typically raised via `Err(TrialPruned)?` inside the
+    /// objective closure.
     #[error("trial was pruned")]
     TrialPruned,
 
-    /// Returned when an internal invariant is violated.
+    /// The multi-objective closure returned a different number of values
+    /// than the number of directions configured on the study.
+    #[error("objective dimension mismatch: expected {expected} values, got {got}")]
+    ObjectiveDimensionMismatch {
+        /// The expected number of objective values.
+        expected: usize,
+        /// The actual number of objective values returned.
+        got: usize,
+    },
+
+    /// An internal invariant was violated. This indicates a bug in the
+    /// library rather than a user error.
     #[error("internal error: {0}")]
     Internal(&'static str),
 
-    /// Returned when an async task fails.
+    /// An async worker task failed. Only available with the `async` feature.
     #[cfg(feature = "async")]
     #[error("async task error: {0}")]
     TaskError(String),
+
+    /// A storage I/O operation failed. Only available with the `journal`
+    /// feature.
+    #[cfg(feature = "journal")]
+    #[error("storage error: {0}")]
+    Storage(String),
 }
 
+/// A convenience alias for `core::result::Result<T, Error>`.
 pub type Result<T> = core::result::Result<T, Error>;
 
 /// Convenience type for signalling a pruned trial from an objective function.

src/fanova.rs

@@ -0,0 +1,604 @@
+//! fANOVA (functional ANOVA) parameter importance via random forest.
+//!
+//! fANOVA decomposes the variance of the objective function into
+//! contributions from individual parameters (**main effects**) and
+//! parameter pairs (**interaction effects**). This helps answer the
+//! question: *"Which parameters matter most, and do any parameters
+//! interact?"*
+//!
+//! # Algorithm
+//!
+//! 1. Fit a random forest to the mapping `(parameters) → objective`
+//! 2. Apply functional ANOVA decomposition to the trained forest
+//! 3. Compute main effects: the variance explained by each parameter alone
+//! 4. Compute interaction effects: the additional variance explained by
+//!    pairs of parameters beyond their individual contributions
+//! 5. Normalize so all importances sum to 1.0
+//!
+//! # When to use
+//!
+//! - **After optimization**: call [`Study::fanova()`](crate::Study::fanova)
+//!   or [`Study::fanova_with_config()`](crate::Study::fanova_with_config)
+//!   to identify which parameters had the most impact
+//! - **Interaction detection**: unlike Spearman correlation
+//!   ([`Study::param_importance()`](crate::Study::param_importance)),
+//!   fANOVA can detect non-linear relationships and parameter interactions
+//! - **Hyperparameter tuning**: focus tuning effort on high-importance
+//!   parameters and fix low-importance ones to reasonable defaults
+//!
+//! # Reference
+//!
+//! Hutter, F., Hoos, H. & Leyton-Brown, K. (2014). "An Efficient
+//! Approach for Assessing Hyperparameter Importance." ICML 2014.
+//!
+//! # Example
+//!
+//! ```
+//! use optimizer::prelude::*;
+//!
+//! let study: Study<f64> = Study::new(Direction::Minimize);
+//! let x = FloatParam::new(0.0, 10.0).name("x");
+//! let y = FloatParam::new(0.0, 10.0).name("y");
+//!
+//! study
+//!     .optimize(50, |trial: &mut optimizer::Trial| {
+//!         let xv = x.suggest(trial)?;
+//!         let yv = y.suggest(trial)?;
+//!         // x matters much more than y
+//!         Ok::<_, optimizer::Error>(3.0 * xv + 0.1 * yv)
+//!     })
+//!     .unwrap();
+//!
+//! let result = study.fanova().unwrap();
+//! // Main effects sorted by descending importance
+//! assert_eq!(result.main_effects[0].0, "x");
+//! ```
+
+/// Result of fANOVA analysis.
+///
+/// All importance values are fractions of total variance and sum to 1.0
+/// across main effects and interactions combined.
+#[derive(Debug, Clone)]
+pub struct FanovaResult {
+    /// Per-parameter importance (fraction of total variance explained).
+    ///
+    /// Sorted by descending importance. Each entry is
+    /// `(parameter_name, importance)` where importance is in `[0.0, 1.0]`.
+    pub main_effects: Vec<(String, f64)>,
+    /// Pairwise interaction importance (fraction of total variance explained).
+    ///
+    /// Sorted by descending importance. Each entry is
+    /// `((param_a, param_b), importance)`. Only pairs with non-negligible
+    /// interaction (> 1e-10) are included.
+    pub interactions: Vec<((String, String), f64)>,
+}
+
+/// Configuration for fANOVA analysis.
+///
+/// Use [`Default::default()`] for reasonable settings, or customize
+/// the random forest parameters for specific needs. Pass to
+/// [`Study::fanova_with_config()`](crate::Study::fanova_with_config).
+#[derive(Debug, Clone)]
+pub struct FanovaConfig {
+    /// Number of trees in the random forest (default: 64).
+    pub n_trees: usize,
+    /// Maximum depth of each tree. `None` for unlimited (default: `None`).
+    pub max_depth: Option<usize>,
+    /// Minimum samples required to split a node (default: 2).
+    pub min_samples_split: usize,
+    /// Minimum samples required in a leaf node (default: 1).
+    pub min_samples_leaf: usize,
+    /// Random seed for reproducibility (default: `Some(42)`).
+    pub seed: Option<u64>,
+}
+
+impl Default for FanovaConfig {
+    fn default() -> Self {
+        Self {
+            n_trees: 64,
+            max_depth: None,
+            min_samples_split: 2,
+            min_samples_leaf: 1,
+            seed: Some(42),
+        }
+    }
+}
+
+// --- Decision Tree ---
+
+/// A node in the regression tree (arena-allocated).
+#[derive(Debug, Clone)]
+enum TreeNode {
+    Leaf {
+        value: f64,
+        n_samples: usize,
+    },
+    Split {
+        feature: usize,
+        threshold: f64,
+        left: usize,
+        right: usize,
+        n_samples: usize,
+    },
+}
+
+/// A regression decision tree for fANOVA.
+#[derive(Debug, Clone)]
+struct DecisionTree {
+    nodes: Vec<TreeNode>,
+}
+
+impl DecisionTree {
+    /// Build a tree from the given data using the specified bootstrap indices.
+    fn build(
+        data: &[Vec<f64>],
+        targets: &[f64],
+        indices: &[usize],
+        config: &FanovaConfig,
+        rng: &mut fastrand::Rng,
+    ) -> Self {
+        let mut tree = Self { nodes: Vec::new() };
+        tree.build_node(data, targets, indices, 0, config, rng);
+        tree
+    }
+
+    #[allow(clippy::cast_precision_loss)]
+    fn build_node(
+        &mut self,
+        data: &[Vec<f64>],
+        targets: &[f64],
+        indices: &[usize],
+        depth: usize,
+        config: &FanovaConfig,
+        rng: &mut fastrand::Rng,
+    ) -> usize {
+        let n = indices.len();
+        let mean = indices.iter().map(|&i| targets[i]).sum::<f64>() / n as f64;
+
+        // Stopping conditions
+        if n < config.min_samples_split || config.max_depth.is_some_and(|d| depth >= d) {
+            let idx = self.nodes.len();
+            self.nodes.push(TreeNode::Leaf {
+                value: mean,
+                n_samples: n,
+            });
+            return idx;
+        }
+
+        // Pure node check (all targets identical)
+        #[allow(clippy::float_cmp)]
+        if indices.iter().all(|&i| targets[i] == targets[indices[0]]) {
+            let idx = self.nodes.len();
+            self.nodes.push(TreeNode::Leaf {
+                value: mean,
+                n_samples: n,
+            });
+            return idx;
+        }
+
+        let n_features = data[0].len();
+        #[allow(clippy::cast_possible_truncation, clippy::cast_sign_loss)]
+        let max_features = ((n_features as f64).sqrt().ceil() as usize)
+            .max(1)
+            .min(n_features);
+        let candidates = partial_shuffle(n_features, max_features, rng);
+
+        // Total variance at this node
+        let total_var: f64 = indices.iter().map(|&i| (targets[i] - mean).powi(2)).sum();
+        if total_var == 0.0 {
+            let idx = self.nodes.len();
+            self.nodes.push(TreeNode::Leaf {
+                value: mean,
+                n_samples: n,
+            });
+            return idx;
+        }
+
+        let mut best_score = f64::NEG_INFINITY;
+        let mut best_feature = 0;
+        let mut best_threshold = 0.0;
+
+        for &feat in &candidates {
+            let mut values: Vec<f64> = indices.iter().map(|&i| data[i][feat]).collect();
+            values.sort_by(|a, b| a.partial_cmp(b).unwrap_or(core::cmp::Ordering::Equal));
+            values.dedup();
+
+            if values.len() < 2 {
+                continue;
+            }
+
+            for w in values.windows(2) {
+                let threshold = f64::midpoint(w[0], w[1]);
+                let (l_sum, l_sq, l_n, r_sum, r_sq, r_n) =
+                    split_stats(data, targets, indices, feat, threshold);
+
+                if l_n < config.min_samples_leaf || r_n < config.min_samples_leaf {
+                    continue;
+                }
+
+                let l_var = l_sq - l_sum * l_sum / l_n as f64;
+                let r_var = r_sq - r_sum * r_sum / r_n as f64;
+                let score = total_var - l_var - r_var;
+
+                if score > best_score {
+                    best_score = score;
+                    best_feature = feat;
+                    best_threshold = threshold;
+                }
+            }
+        }
+
+        if best_score <= 0.0 {
+            let idx = self.nodes.len();
+            self.nodes.push(TreeNode::Leaf {
+                value: mean,
+                n_samples: n,
+            });
+            return idx;
+        }
+
+        let (left_indices, right_indices): (Vec<usize>, Vec<usize>) = indices
+            .iter()
+            .partition(|&&i| data[i][best_feature] <= best_threshold);
+
+        if left_indices.is_empty() || right_indices.is_empty() {
+            let idx = self.nodes.len();
+            self.nodes.push(TreeNode::Leaf {
+                value: mean,
+                n_samples: n,
+            });
+            return idx;
+        }
+
+        // Reserve slot for this split node (placeholder replaced below)
+        let node_idx = self.nodes.len();
+        self.nodes.push(TreeNode::Leaf {
+            value: 0.0,
+            n_samples: 0,
+        });
+
+        let left = self.build_node(data, targets, &left_indices, depth + 1, config, rng);
+        let right = self.build_node(data, targets, &right_indices, depth + 1, config, rng);
+
+        self.nodes[node_idx] = TreeNode::Split {
+            feature: best_feature,
+            threshold: best_threshold,
+            left,
+            right,
+            n_samples: n,
+        };
+
+        node_idx
+    }
+
+    /// Compute marginal prediction for a given feature subset.
+    ///
+    /// Features in `subset` use values from `feature_values`.
+    /// Features not in `subset` are marginalized by weighting branches
+    /// proportionally to their training-data fractions.
+    fn marginal_predict(&self, subset: &[usize], feature_values: &[f64]) -> f64 {
+        self.marginal_predict_at(0, subset, feature_values)
+    }
+
+    #[allow(clippy::cast_precision_loss)]
+    fn marginal_predict_at(&self, idx: usize, subset: &[usize], vals: &[f64]) -> f64 {
+        match self.nodes[idx] {
+            TreeNode::Leaf { value, .. } => value,
+            TreeNode::Split {
+                feature,
+                threshold,
+                left,
+                right,
+                n_samples,
+            } => {
+                if subset.contains(&feature) {
+                    if vals[feature] <= threshold {
+                        self.marginal_predict_at(left, subset, vals)
+                    } else {
+                        self.marginal_predict_at(right, subset, vals)
+                    }
+                } else {
+                    let l_n = self.n_samples(left) as f64;
+                    let r_n = self.n_samples(right) as f64;
+                    let total = n_samples as f64;
+                    (l_n / total) * self.marginal_predict_at(left, subset, vals)
+                        + (r_n / total) * self.marginal_predict_at(right, subset, vals)
+                }
+            }
+        }
+    }
+
+    fn n_samples(&self, idx: usize) -> usize {
+        match self.nodes[idx] {
+            TreeNode::Leaf { n_samples, .. } | TreeNode::Split { n_samples, .. } => n_samples,
+        }
+    }
+}
+
+// --- Helper Functions ---
+
+/// Select `k` random indices from `0..n` using partial Fisher-Yates shuffle.
+fn partial_shuffle(n: usize, k: usize, rng: &mut fastrand::Rng) -> Vec<usize> {
+    let mut indices: Vec<usize> = (0..n).collect();
+    let k = k.min(n);
+    for i in 0..k {
+        let j = rng.usize(i..n);
+        indices.swap(i, j);
+    }
+    indices.truncate(k);
+    indices
+}
+
+/// Compute left/right split statistics for variance reduction.
+#[allow(clippy::cast_precision_loss)]
+fn split_stats(
+    data: &[Vec<f64>],
+    targets: &[f64],
+    indices: &[usize],
+    feature: usize,
+    threshold: f64,
+) -> (f64, f64, usize, f64, f64, usize) {
+    let (mut l_sum, mut l_sq, mut l_n) = (0.0, 0.0, 0usize);
+    let (mut r_sum, mut r_sq, mut r_n) = (0.0, 0.0, 0usize);
+
+    for &i in indices {
+        let y = targets[i];
+        if data[i][feature] <= threshold {
+            l_sum += y;
+            l_sq += y * y;
+            l_n += 1;
+        } else {
+            r_sum += y;
+            r_sq += y * y;
+            r_n += 1;
+        }
+    }
+
+    (l_sum, l_sq, l_n, r_sum, r_sq, r_n)
+}
+
+/// Population variance of a slice.
+#[allow(clippy::cast_precision_loss)]
+fn variance(values: &[f64]) -> f64 {
+    if values.is_empty() {
+        return 0.0;
+    }
+    let n = values.len() as f64;
+    let mean = values.iter().sum::<f64>() / n;
+    values.iter().map(|v| (v - mean).powi(2)).sum::<f64>() / n
+}
+
+// --- Public API ---
+
+/// Run fANOVA analysis on pre-processed numerical data.
+///
+/// `data` is `n_samples` rows, each with `n_features` columns.
+/// `targets` has one entry per sample.
+/// `feature_names` maps feature index to human-readable name.
+#[allow(clippy::cast_precision_loss)]
+pub(crate) fn compute_fanova(
+    data: &[Vec<f64>],
+    targets: &[f64],
+    feature_names: &[String],
+    config: &FanovaConfig,
+) -> FanovaResult {
+    let n_samples = data.len();
+    let n_features = data[0].len();
+
+    let mut rng: fastrand::Rng = config
+        .seed
+        .map_or_else(fastrand::Rng::new, fastrand::Rng::with_seed);
+
+    // Build random forest with bootstrap sampling
+    let trees: Vec<DecisionTree> = (0..config.n_trees)
+        .map(|_| {
+            let bootstrap: Vec<usize> = (0..n_samples).map(|_| rng.usize(0..n_samples)).collect();
+            DecisionTree::build(data, targets, &bootstrap, config, &mut rng)
+        })
+        .collect();
+
+    // Compute main effects: V_j = Var[E[f | x_j]]
+    let main_var: Vec<f64> = (0..n_features)
+        .map(|j| {
+            let subset = [j];
+            let preds: Vec<f64> = (0..n_samples)
+                .map(|i| {
+                    trees
+                        .iter()
+                        .map(|t| t.marginal_predict(&subset, &data[i]))
+                        .sum::<f64>()
+                        / trees.len() as f64
+                })
+                .collect();
+            variance(&preds)
+        })
+        .collect();
+
+    // Compute pairwise interaction effects: V_{j,k} - V_j - V_k
+    let mut interactions: Vec<((String, String), f64)> = Vec::new();
+    for j in 0..n_features {
+        for k in (j + 1)..n_features {
+            let subset = [j, k];
+            let preds: Vec<f64> = (0..n_samples)
+                .map(|i| {
+                    trees
+                        .iter()
+                        .map(|t| t.marginal_predict(&subset, &data[i]))
+                        .sum::<f64>()
+                        / trees.len() as f64
+                })
+                .collect();
+            let joint = variance(&preds);
+            let interaction = (joint - main_var[j] - main_var[k]).max(0.0);
+            if interaction > 1e-10 {
+                interactions.push((
+                    (feature_names[j].clone(), feature_names[k].clone()),
+                    interaction,
+                ));
+            }
+        }
+    }
+
+    // Normalize so all importances sum to 1.0
+    let total: f64 =
+        main_var.iter().sum::<f64>() + interactions.iter().map(|(_, v)| *v).sum::<f64>();
+
+    let mut main_effects: Vec<(String, f64)> = feature_names
+        .iter()
+        .zip(&main_var)
+        .map(|(name, &v)| (name.clone(), if total > 0.0 { v / total } else { 0.0 }))
+        .collect();
+    main_effects.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap_or(core::cmp::Ordering::Equal));
+
+    if total > 0.0 {
+        for entry in &mut interactions {
+            entry.1 /= total;
+        }
+    }
+    interactions.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap_or(core::cmp::Ordering::Equal));
+
+    FanovaResult {
+        main_effects,
+        interactions,
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::rng_util;
+
+    #[test]
+    fn single_dominant_parameter() {
+        // f(x, y) = x — only x matters
+        let mut rng = fastrand::Rng::with_seed(0);
+        let n = 100;
+        let data: Vec<Vec<f64>> = (0..n)
+            .map(|_| {
+                vec![
+                    rng_util::f64_range(&mut rng, 0.0, 10.0),
+                    rng_util::f64_range(&mut rng, 0.0, 10.0),
+                ]
+            })
+            .collect();
+        let targets: Vec<f64> = data.iter().map(|row| row[0]).collect();
+
+        let result = compute_fanova(
+            &data,
+            &targets,
+            &["x".into(), "y".into()],
+            &FanovaConfig::default(),
+        );
+
+        assert_eq!(result.main_effects[0].0, "x");
+        assert!(
+            result.main_effects[0].1 > 0.8,
+            "x importance = {}",
+            result.main_effects[0].1
+        );
+    }
+
+    #[test]
+    fn interaction_detection() {
+        // f(x, y) = x * y — both matter and interact
+        let mut rng = fastrand::Rng::with_seed(42);
+        let n = 200;
+        let data: Vec<Vec<f64>> = (0..n)
+            .map(|_| {
+                vec![
+                    rng_util::f64_range(&mut rng, 0.0, 10.0),
+                    rng_util::f64_range(&mut rng, 0.0, 10.0),
+                ]
+            })
+            .collect();
+        let targets: Vec<f64> = data.iter().map(|row| row[0] * row[1]).collect();
+
+        let config = FanovaConfig {
+            n_trees: 128,
+            ..FanovaConfig::default()
+        };
+        let result = compute_fanova(&data, &targets, &["x".into(), "y".into()], &config);
+
+        assert!(
+            !result.interactions.is_empty(),
+            "should detect x*y interaction"
+        );
+        assert!(
+            result.interactions[0].1 > 0.05,
+            "interaction importance = {}",
+            result.interactions[0].1
+        );
+    }
+
+    #[test]
+    fn variance_computation() {
+        assert!((variance(&[1.0, 2.0, 3.0, 4.0, 5.0]) - 2.0).abs() < 1e-10);
+        assert!(variance(&[5.0, 5.0, 5.0]).abs() < 1e-10);
+        assert!(variance(&[]).abs() < 1e-10);
+    }
+
+    #[test]
+    fn three_params_one_dominant() {
+        // f(x, y, z) = 3*x + 0.1*y + 0*z
+        let mut rng = fastrand::Rng::with_seed(7);
+        let n = 150;
+        let data: Vec<Vec<f64>> = (0..n)
+            .map(|_| {
+                vec![
+                    rng_util::f64_range(&mut rng, 0.0, 10.0),
+                    rng_util::f64_range(&mut rng, 0.0, 10.0),
+                    rng_util::f64_range(&mut rng, 0.0, 10.0),
+                ]
+            })
+            .collect();
+        let targets: Vec<f64> = data.iter().map(|r| 3.0 * r[0] + 0.1 * r[1]).collect();
+
+        let result = compute_fanova(
+            &data,
+            &targets,
+            &["x".into(), "y".into(), "z".into()],
+            &FanovaConfig::default(),
+        );
+
+        // x should be the most important
+        assert_eq!(result.main_effects[0].0, "x");
+        assert!(result.main_effects[0].1 > 0.5);
+
+        // z should have near-zero importance
+        let z_imp = result
+            .main_effects
+            .iter()
+            .find(|(name, _)| name == "z")
+            .map_or(0.0, |(_, v)| *v);
+        assert!(z_imp < 0.1, "z importance = {z_imp}");
+    }
+
+    #[test]
+    fn importances_sum_to_one() {
+        let mut rng = fastrand::Rng::with_seed(3);
+        let n = 100;
+        let data: Vec<Vec<f64>> = (0..n)
+            .map(|_| {
+                vec![
+                    rng_util::f64_range(&mut rng, 0.0, 10.0),
+                    rng_util::f64_range(&mut rng, 0.0, 10.0),
+                ]
+            })
+            .collect();
+        let targets: Vec<f64> = data.iter().map(|r| r[0] + r[1]).collect();
+
+        let result = compute_fanova(
+            &data,
+            &targets,
+            &["x".into(), "y".into()],
+            &FanovaConfig::default(),
+        );
+
+        let total: f64 = result.main_effects.iter().map(|(_, v)| *v).sum::<f64>()
+            + result.interactions.iter().map(|(_, v)| *v).sum::<f64>();
+        assert!(
+            (total - 1.0).abs() < 1e-10,
+            "importances should sum to 1.0, got {total}"
+        );
+    }
+}

src/importance.rs

@@ -1,4 +1,26 @@
 //! Parameter importance via Spearman rank correlation.
+//!
+//! Compute the absolute Spearman rank correlation between each parameter
+//! and the objective value to estimate which parameters most influence
+//! the outcome. This is a lightweight, non-parametric alternative to
+//! [`fANOVA`](crate::fanova) that works well for monotonic relationships.
+//!
+//! # How it works
+//!
+//! 1. Rank parameter values and objective values independently
+//! 2. Compute the Pearson correlation on the ranks (= Spearman ρ)
+//! 3. Take the absolute value (direction of correlation is not relevant
+//!    for importance)
+//!
+//! # When to use
+//!
+//! - **Quick importance check**: call
+//!   [`Study::param_importance()`](crate::Study::param_importance) after
+//!   optimization for a fast, interpretable ranking
+//! - **Monotonic relationships**: Spearman captures monotonic (not just
+//!   linear) correlations but may miss non-monotonic effects or interactions
+//! - For interaction detection or non-linear importance, use
+//!   [`fANOVA`](crate::fanova) instead
 
 /// Assign average ranks to a slice of `f64` values (handles ties).
 #[allow(clippy::cast_precision_loss, clippy::float_cmp)]

src/kde/multivariate.rs

@@ -5,8 +5,6 @@
 //! parameter independently, the multivariate KDE models the joint distribution
 //! to better capture correlations between parameters.
 
-use rand::{Rng, RngExt};
-
 use crate::error::{Error, Result};
 
 /// A multivariate Gaussian kernel density estimator for joint distributions.
@@ -32,7 +30,6 @@ use crate::error::{Error, Result};
 /// // Get dimensionality
 /// assert_eq!(kde.n_dims(), 2);
 /// ```
-#[allow(dead_code)] // Fields and methods will be used in subsequent stories (US-003, US-004)
 #[derive(Clone, Debug)]
 pub(crate) struct MultivariateKDE {
     /// The sample points used to construct the KDE.
@@ -45,7 +42,6 @@ pub(crate) struct MultivariateKDE {
     n_dims: usize,
 }
 
-#[allow(dead_code)] // Methods will be used in subsequent stories (US-003, US-004)
 impl MultivariateKDE {
     /// Creates a new multivariate KDE with automatic bandwidth selection using Scott's rule.
     ///
@@ -102,6 +98,7 @@ impl MultivariateKDE {
     /// Returns `Error::ZeroDimensions` if samples have zero dimensions.
     /// Returns `Error::BandwidthDimensionMismatch` if bandwidths length doesn't match dimensions.
     /// Returns `Error::InvalidBandwidth` if any bandwidth is not positive.
+    #[cfg(test)]
     pub(crate) fn with_bandwidths(samples: Vec<Vec<f64>>, bandwidths: Vec<f64>) -> Result<Self> {
         if samples.is_empty() {
             return Err(Error::EmptySamples);
@@ -182,12 +179,13 @@ impl MultivariateKDE {
     }
 
     /// Returns the number of dimensions.
+    #[cfg(test)]
     pub(crate) fn n_dims(&self) -> usize {
         self.n_dims
     }
 
     /// Returns the number of samples.
-    #[allow(dead_code)] // Will be used in subsequent stories
+    #[cfg(test)]
     pub(crate) fn n_samples(&self) -> usize {
         self.samples.len()
     }
@@ -199,7 +197,7 @@ impl MultivariateKDE {
     }
 
     /// Returns a reference to the samples.
-    #[allow(dead_code)] // Will be used in subsequent stories
+    #[cfg(test)]
     pub(crate) fn samples(&self) -> &[Vec<f64>] {
         &self.samples
     }
@@ -290,6 +288,7 @@ impl MultivariateKDE {
     /// # Panics
     ///
     /// Panics if `x.len() != self.n_dims`.
+    #[cfg(test)]
     pub(crate) fn pdf(&self, x: &[f64]) -> f64 {
         self.log_pdf(x).exp()
     }
@@ -308,9 +307,9 @@ impl MultivariateKDE {
     /// # Returns
     ///
     /// A `Vec<f64>` of length `n_dims` representing a sample from the KDE.
-    pub(crate) fn sample<R: Rng>(&self, rng: &mut R) -> Vec<f64> {
+    pub(crate) fn sample(&self, rng: &mut fastrand::Rng) -> Vec<f64> {
         // Select a random sample to center the kernel on
-        let idx = rng.random_range(0..self.samples.len());
+        let idx = rng.usize(0..self.samples.len());
         let center = &self.samples[idx];
 
         // Add independent Gaussian noise to each dimension
@@ -319,8 +318,8 @@ impl MultivariateKDE {
             .iter()
             .zip(self.bandwidths.iter())
             .map(|(&center_j, &bandwidth_j)| {
-                let u1: f64 = rng.random();
-                let u2: f64 = rng.random();
+                let u1: f64 = rng.f64();
+                let u2: f64 = rng.f64();
 
                 // Box-Muller transform: generates standard normal variate
                 let z = (-2.0 * u1.ln()).sqrt() * (2.0 * core::f64::consts::PI * u2).cos();
@@ -724,7 +723,7 @@ mod tests {
     fn test_multivariate_kde_sample_basic() {
         let samples = vec![vec![0.0, 0.0], vec![1.0, 1.0], vec![2.0, 2.0]];
         let kde = MultivariateKDE::new(samples).unwrap();
-        let mut rng = rand::rng();
+        let mut rng = fastrand::Rng::new();
 
         // Sample should have correct dimensionality
         let sample = kde.sample(&mut rng);
@@ -741,7 +740,7 @@ mod tests {
             vec![4.0, 4.0],
         ];
         let kde = MultivariateKDE::new(samples).unwrap();
-        let mut rng = rand::rng();
+        let mut rng = fastrand::Rng::new();
 
         // Samples should generally be in a reasonable range around the data
         for _ in 0..100 {
@@ -766,7 +765,7 @@ mod tests {
         // When KDE has only one sample, all samples should be centered around it
         let samples = vec![vec![5.0, 10.0]];
         let kde = MultivariateKDE::new(samples).unwrap();
-        let mut rng = rand::rng();
+        let mut rng = fastrand::Rng::new();
 
         // Generate many samples and check they cluster around (5.0, 10.0)
         let n_samples = 100;
@@ -803,7 +802,7 @@ mod tests {
             })
             .collect();
         let kde = MultivariateKDE::new(samples).unwrap();
-        let mut rng = rand::rng();
+        let mut rng = fastrand::Rng::new();
 
         // Sample should have correct dimensionality
         for _ in 0..50 {
@@ -823,7 +822,7 @@ mod tests {
         let data = vec![vec![0.0, 0.0], vec![0.0, 0.0], vec![0.0, 0.0]];
         let bandwidths = vec![0.1, 10.0]; // Small bandwidth in x, large in y
         let kde = MultivariateKDE::with_bandwidths(data, bandwidths).unwrap();
-        let mut rng = rand::rng();
+        let mut rng = fastrand::Rng::new();
 
         // Generate samples and check variance in each dimension
         let n_samples = 1000;
@@ -868,7 +867,7 @@ mod tests {
             vec![4.0, 4.0],
         ];
         let kde = MultivariateKDE::new(data).unwrap();
-        let mut rng = rand::rng();
+        let mut rng = fastrand::Rng::new();
 
         // Sample many points and verify the mean is near the center
         let n_samples = 500;
@@ -896,14 +895,12 @@ mod tests {
 
     #[test]
     fn test_multivariate_kde_sample_deterministic_with_seeded_rng() {
-        use rand::SeedableRng;
-
         let data = vec![vec![1.0, 2.0], vec![3.0, 4.0]];
         let kde = MultivariateKDE::new(data).unwrap();
 
         // Use a seeded RNG for reproducibility
-        let mut rng1 = rand::rngs::StdRng::seed_from_u64(42);
-        let mut rng2 = rand::rngs::StdRng::seed_from_u64(42);
+        let mut rng1 = fastrand::Rng::with_seed(42);
+        let mut rng2 = fastrand::Rng::with_seed(42);
 
         // Same seed should produce same samples
         let result1 = kde.sample(&mut rng1);

src/kde/univariate.rs

@@ -3,8 +3,6 @@
 //! This module provides a Gaussian kernel density estimator used by the TPE
 //! sampler to model probability distributions over good and bad trial regions.
 
-use rand::{Rng, RngExt};
-
 use crate::error::{Error, Result};
 
 /// A Gaussian kernel density estimator for continuous distributions.
@@ -26,7 +24,7 @@ use crate::error::{Error, Result};
 /// assert!(density > 0.0);
 ///
 /// // Sample from the estimated distribution
-/// let mut rng = rand::rng();
+/// let mut rng = fastrand::Rng::new();
 /// let sample = kde.sample(&mut rng);
 /// ```
 #[derive(Clone, Debug)]
@@ -132,15 +130,15 @@ impl KernelDensityEstimator {
     /// Sampling works by:
     /// 1. Uniformly selecting one of the kernel centers (samples)
     /// 2. Adding Gaussian noise with the bandwidth as standard deviation
-    pub(crate) fn sample<R: Rng>(&self, rng: &mut R) -> f64 {
+    pub(crate) fn sample(&self, rng: &mut fastrand::Rng) -> f64 {
         // Select a random sample to center the kernel on
-        let idx = rng.random_range(0..self.samples.len());
+        let idx = rng.usize(0..self.samples.len());
         let center = self.samples[idx];
 
         // Add Gaussian noise with bandwidth as standard deviation
         // Using Box-Muller transform for Gaussian sampling
-        let u1: f64 = rng.random();
-        let u2: f64 = rng.random();
+        let u1: f64 = rng.f64();
+        let u2: f64 = rng.f64();
 
         let z = (-2.0 * u1.ln()).sqrt() * (2.0 * core::f64::consts::PI * u2).cos();
         center + z * self.bandwidth
@@ -211,7 +209,7 @@ mod tests {
     fn test_kde_sample_in_reasonable_range() {
         let samples = vec![0.0, 1.0, 2.0, 3.0, 4.0];
         let kde = KernelDensityEstimator::new(samples).unwrap();
-        let mut rng = rand::rng();
+        let mut rng = fastrand::Rng::new();
 
         // Samples should generally be in a reasonable range around the data
         for _ in 0..100 {

src/lib.rs

@@ -9,186 +9,79 @@
 #![deny(clippy::pedantic)]
 #![deny(clippy::std_instead_of_core)]
 
-//! A black-box optimization library with multiple sampling strategies.
+//! Bayesian and population-based optimization library with an Optuna-like API
+//! for hyperparameter tuning and black-box optimization. It ships 12 samplers
+//! (from random search to CMA-ES and NSGA-III), 8 pruners, async/parallel
+//! evaluation, and optional journal-based persistence — all with zero required
+//! feature flags for the common case.
 //!
-//! This library provides an Optuna-like API for hyperparameter optimization
-//! with support for multiple sampling algorithms:
+//! # Getting Started
 //!
-//! - **Random Search** - Simple random sampling for baseline comparisons
-//! - **TPE (Tree-Parzen Estimator)** - Bayesian optimization for efficient search
-//! - **Grid Search** - Exhaustive search over a specified parameter grid
-//! - **Sobol (QMC)** - Quasi-random sampling for better space coverage (requires `sobol` feature)
-//! - **CMA-ES** - Covariance Matrix Adaptation Evolution Strategy for continuous optimization (requires `cma-es` feature)
-//! - **BOHB** - Bayesian Optimization + `HyperBand` for budget-aware TPE sampling
-//!
-//! Additional features include:
-//!
-//! - Float, integer, and categorical parameter types
-//! - Log-scale and stepped parameter sampling
-//! - Synchronous and async optimization
-//! - Parallel trial evaluation with bounded concurrency
-//!
-//! # Quick Start
+//! Minimize a function in five lines — no feature flags needed:
 //!
 //! ```
 //! use optimizer::prelude::*;
 //!
-//! // Create a study with TPE sampler
-//! let sampler = TpeSampler::builder().seed(42).build().unwrap();
-//! let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
-//!
-//! // Define parameter search space
+//! let study: Study<f64> = Study::new(Direction::Minimize);
 //! let x = FloatParam::new(-10.0, 10.0).name("x");
 //!
-//! // Optimize x^2 for 20 trials
 //! study
-//!     .optimize(20, |trial| {
-//!         let x_val = x.suggest(trial)?;
-//!         Ok::<_, Error>(x_val * x_val)
+//!     .optimize(50, |trial: &mut optimizer::Trial| {
+//!         let v = x.suggest(trial)?;
+//!         Ok::<_, Error>((v - 3.0).powi(2))
 //!     })
 //!     .unwrap();
 //!
-//! // Get the best result
 //! let best = study.best_trial().unwrap();
-//! println!("x = {}", best.get(&x).unwrap());
+//! println!("x = {:.4}, f(x) = {:.4}", best.get(&x).unwrap(), best.value);
 //! ```
 //!
-//! # Creating a Study
+//! # Core Concepts
 //!
-//! A [`Study`] manages optimization trials. Create one with an optimization direction:
+//! | Type | Role |
+//! |------|------|
+//! | [`Study`] | Drive an optimization loop: create trials, record results, track the best. |
+//! | [`Trial`] | A single evaluation of the objective function, carrying suggested parameter values. |
+//! | [`Parameter`](parameter::Parameter) | Define the search space — [`FloatParam`](parameter::FloatParam), [`IntParam`](parameter::IntParam), [`CategoricalParam`](parameter::CategoricalParam), [`BoolParam`](parameter::BoolParam), [`EnumParam`](parameter::EnumParam). |
+//! | [`Sampler`](sampler::Sampler) | Strategy for choosing the next point to evaluate (TPE, CMA-ES, random, etc.). |
+//! | [`Direction`] | Whether the study minimizes or maximizes the objective value. |
 //!
-//! ```
-//! use optimizer::sampler::random::RandomSampler;
-//! use optimizer::sampler::tpe::TpeSampler;
-//! use optimizer::{Direction, Study};
+//! # Sampler Guide
 //!
-//! // Minimize with default random sampler
-//! let study: Study<f64> = Study::new(Direction::Minimize);
+//! ## Single-objective samplers
 //!
-//! // Maximize with TPE sampler
-//! let study: Study<f64> = Study::with_sampler(Direction::Maximize, TpeSampler::new());
+//! | Sampler | Algorithm | Best for | Feature flag |
+//! |---------|-----------|----------|--------------|
+//! | [`RandomSampler`](sampler::RandomSampler) | Uniform random | Baselines, high-dimensional | — |
+//! | [`TpeSampler`](sampler::TpeSampler) | Tree-Parzen Estimator | General-purpose Bayesian | — |
+//! | [`GridSearchSampler`](sampler::GridSampler) | Exhaustive grid | Small, discrete spaces | — |
+//! | `SobolSampler` | Sobol quasi-random sequence | Space-filling, low dimensions | `sobol` |
+//! | `CmaEsSampler` | CMA-ES | Continuous, moderate dimensions | `cma-es` |
+//! | `GpSampler` | Gaussian Process + EI | Expensive objectives, few trials | `gp` |
+//! | [`DESampler`](sampler::DESampler) | Differential Evolution | Non-convex, population-based | — |
+//! | [`BohbSampler`](sampler::BohbSampler) | BOHB (TPE + `HyperBand`) | Budget-aware early stopping | — |
 //!
-//! // With seeded sampler for reproducibility
-//! let study: Study<f64> = Study::with_sampler(Direction::Minimize, RandomSampler::with_seed(42));
-//! ```
+//! ## Multi-objective samplers
 //!
-//! # Suggesting Parameters
-//!
-//! Within the objective function, use parameter types to suggest values:
-//!
-//! ```
-//! use optimizer::parameter::{BoolParam, CategoricalParam, FloatParam, IntParam, Parameter};
-//! use optimizer::{Direction, Study};
-//!
-//! let study: Study<f64> = Study::new(Direction::Minimize);
-//!
-//! // Define parameter search spaces
-//! let x_param = FloatParam::new(0.0, 1.0);
-//! let lr_param = FloatParam::new(1e-5, 1e-1).log_scale();
-//! let step_param = FloatParam::new(0.0, 1.0).step(0.1);
-//! let n_param = IntParam::new(1, 10);
-//! let batch_param = IntParam::new(16, 256).log_scale();
-//! let units_param = IntParam::new(32, 512).step(32);
-//! let flag_param = BoolParam::new();
-//! let optimizer_param = CategoricalParam::new(vec!["sgd", "adam", "rmsprop"]);
-//!
-//! study
-//!     .optimize(10, |trial| {
-//!         let x = x_param.suggest(trial)?;
-//!         let lr = lr_param.suggest(trial)?;
-//!         let step = step_param.suggest(trial)?;
-//!         let n = n_param.suggest(trial)?;
-//!         let batch = batch_param.suggest(trial)?;
-//!         let units = units_param.suggest(trial)?;
-//!         let flag = flag_param.suggest(trial)?;
-//!         let optimizer = optimizer_param.suggest(trial)?;
-//!
-//!         Ok::<_, optimizer::Error>(x * n as f64)
-//!     })
-//!     .unwrap();
-//! ```
-//!
-//! # Available Samplers
-//!
-//! ## Random Search
-//!
-//! The simplest sampling strategy, useful for baselines:
-//!
-//! ```
-//! use optimizer::sampler::random::RandomSampler;
-//! use optimizer::{Direction, Study};
-//!
-//! let study: Study<f64> = Study::with_sampler(Direction::Minimize, RandomSampler::with_seed(42));
-//! ```
-//!
-//! ## TPE (Tree-Parzen Estimator)
-//!
-//! Bayesian optimization that learns from previous trials:
-//!
-//! ```
-//! use optimizer::sampler::tpe::TpeSampler;
-//!
-//! let sampler = TpeSampler::builder()
-//!     .gamma(0.15)           // Quantile for good/bad split
-//!     .n_startup_trials(20)  // Random trials before TPE
-//!     .n_ei_candidates(32)   // Candidates to evaluate
-//!     .seed(42)              // Reproducibility
-//!     .build()
-//!     .unwrap();
-//! ```
-//!
-//! ## Grid Search
-//!
-//! Exhaustive search over a discretized parameter space:
-//!
-//! ```
-//! use optimizer::sampler::grid::GridSearchSampler;
-//! use optimizer::{Direction, Study};
-//!
-//! let sampler = GridSearchSampler::builder()
-//!     .n_points_per_param(10)  // Points per parameter dimension
-//!     .build();
-//!
-//! let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
-//! ```
-//!
-//! # Async and Parallel Optimization
-//!
-//! With the `async` feature enabled, you can run trials asynchronously:
-//!
-//! ```ignore
-//! use optimizer::{Study, Direction};
-//! use optimizer::parameter::{FloatParam, Parameter};
-//!
-//! let x_param = FloatParam::new(0.0, 1.0);
-//!
-//! // Sequential async
-//! study.optimize_async(10, |mut trial| {
-//!     let x_param = x_param.clone();
-//!     async move {
-//!         let x = x_param.suggest(&mut trial)?;
-//!         Ok((trial, x * x))
-//!     }
-//! }).await?;
-//!
-//! // Parallel with bounded concurrency
-//! study.optimize_parallel(10, 4, |mut trial| {
-//!     let x_param = x_param.clone();
-//!     async move {
-//!         let x = x_param.suggest(&mut trial)?;
-//!         Ok((trial, x * x))
-//!     }
-//! }).await?;
-//! ```
+//! | Sampler | Algorithm | Best for | Feature flag |
+//! |---------|-----------|----------|--------------|
+//! | [`Nsga2Sampler`](sampler::Nsga2Sampler) | NSGA-II | 2-3 objectives | — |
+//! | [`Nsga3Sampler`](sampler::Nsga3Sampler) | NSGA-III (reference-point) | 3+ objectives | — |
+//! | [`MoeadSampler`](sampler::MoeadSampler) | MOEA/D (decomposition) | Many objectives, structured fronts | — |
+//! | [`MotpeSampler`](sampler::MotpeSampler) | Multi-Objective TPE | Bayesian multi-objective | — |
 //!
 //! # Feature Flags
 //!
-//! - `async`: Enable async optimization methods (requires tokio)
-//! - `derive`: Enable `#[derive(Categorical)]` for enum parameters
-//! - `serde`: Enable `Serialize`/`Deserialize` on public types and `Study::save()`/`Study::load()`
-//! - `sobol`: Enable the Sobol quasi-random sampler for better space coverage
-//! - `cma-es`: Enable the CMA-ES sampler for continuous optimization
-//! - `tracing`: Emit structured log events via the [`tracing`](https://docs.rs/tracing) crate at key optimization points
+//! | Flag | What it enables | Default |
+//! |------|----------------|---------|
+//! | `async` | Async/parallel optimization via tokio (`Study::optimize_async`, `Study::optimize_parallel`) | off |
+//! | `derive` | `#[derive(Categorical)]` for enum parameters | off |
+//! | `serde` | `Serialize`/`Deserialize` on public types, `Study::save`/`Study::load` | off |
+//! | `journal` | `JournalStorage` — JSONL persistence with file locking (enables `serde`) | off |
+//! | `sobol` | `SobolSampler` — quasi-random low-discrepancy sequences | off |
+//! | `cma-es` | `CmaEsSampler` — Covariance Matrix Adaptation Evolution Strategy | off |
+//! | `gp` | `GpSampler` — Gaussian Process surrogate with Expected Improvement | off |
+//! | `tracing` | Structured log events via [`tracing`](https://docs.rs/tracing) at key optimization points | off |
 
 /// Emit a `tracing::info!` event when the `tracing` feature is enabled.
 /// No-op otherwise.
@@ -214,43 +107,36 @@ macro_rules! trace_debug {
     ($($arg:tt)*) => {};
 }
 
-mod distribution;
+pub mod distribution;
 mod error;
+mod fanova;
 mod importance;
 mod kde;
-mod param;
+pub mod multi_objective;
+pub mod objective;
+pub mod param;
 pub mod parameter;
+pub mod pareto;
 pub mod pruner;
+mod rng_util;
 pub mod sampler;
+pub mod storage;
 mod study;
 mod trial;
 mod types;
+mod visualization;
 
 pub use error::{Error, Result, TrialPruned};
+pub use fanova::{FanovaConfig, FanovaResult};
+pub use objective::Objective;
 #[cfg(feature = "derive")]
 pub use optimizer_derive::Categorical;
-pub use param::ParamValue;
-pub use parameter::{
-    BoolParam, Categorical, CategoricalParam, EnumParam, FloatParam, IntParam, ParamId, Parameter,
-};
-pub use pruner::{
-    HyperbandPruner, MedianPruner, NopPruner, PatientPruner, PercentilePruner, Pruner,
-    SuccessiveHalvingPruner, ThresholdPruner, WilcoxonPruner,
-};
-pub use sampler::CompletedTrial;
-pub use sampler::bohb::BohbSampler;
-#[cfg(feature = "cma-es")]
-pub use sampler::cma_es::CmaEsSampler;
-pub use sampler::grid::GridSearchSampler;
-pub use sampler::random::RandomSampler;
-#[cfg(feature = "sobol")]
-pub use sampler::sobol::SobolSampler;
-pub use sampler::tpe::TpeSampler;
-pub use study::Study;
 #[cfg(feature = "serde")]
 pub use study::StudySnapshot;
+pub use study::{Study, StudyBuilder};
 pub use trial::{AttrValue, Trial};
 pub use types::{Direction, TrialState};
+pub use visualization::generate_html_report;
 
 /// Convenient wildcard import for the most common types.
 ///
@@ -262,26 +148,36 @@ pub mod prelude {
     pub use optimizer_derive::Categorical as DeriveCategory;
 
     pub use crate::error::{Error, Result, TrialPruned};
-    pub use crate::param::ParamValue;
+    pub use crate::fanova::{FanovaConfig, FanovaResult};
+    pub use crate::multi_objective::{
+        MultiObjectiveSampler, MultiObjectiveStudy, MultiObjectiveTrial,
+    };
+    pub use crate::objective::Objective;
     pub use crate::parameter::{
-        BoolParam, Categorical, CategoricalParam, EnumParam, FloatParam, IntParam, Parameter,
+        BoolParam, Categorical, CategoricalParam, EnumParam, FloatParam, IntParam, ParamValue,
+        Parameter,
     };
     pub use crate::pruner::{
         HyperbandPruner, MedianPruner, NopPruner, PatientPruner, PercentilePruner, Pruner,
-        SuccessiveHalvingPruner, ThresholdPruner,
+        SuccessiveHalvingPruner, ThresholdPruner, WilcoxonPruner,
     };
-    pub use crate::sampler::CompletedTrial;
-    pub use crate::sampler::bohb::BohbSampler;
     #[cfg(feature = "cma-es")]
-    pub use crate::sampler::cma_es::CmaEsSampler;
-    pub use crate::sampler::grid::GridSearchSampler;
-    pub use crate::sampler::random::RandomSampler;
+    pub use crate::sampler::CmaEsSampler;
+    #[cfg(feature = "gp")]
+    pub use crate::sampler::GpSampler;
     #[cfg(feature = "sobol")]
-    pub use crate::sampler::sobol::SobolSampler;
-    pub use crate::sampler::tpe::TpeSampler;
-    pub use crate::study::Study;
+    pub use crate::sampler::SobolSampler;
+    pub use crate::sampler::{
+        BohbSampler, CompletedTrial, DESampler, DEStrategy, Decomposition, GridSampler,
+        MoeadSampler, MotpeSampler, Nsga2Sampler, Nsga3Sampler, RandomSampler, TpeSampler,
+    };
+    #[cfg(feature = "journal")]
+    pub use crate::storage::JournalStorage;
+    pub use crate::storage::{MemoryStorage, Storage};
     #[cfg(feature = "serde")]
     pub use crate::study::StudySnapshot;
+    pub use crate::study::{Study, StudyBuilder};
     pub use crate::trial::{AttrValue, Trial};
     pub use crate::types::Direction;
+    pub use crate::visualization::generate_html_report;
 }

src/multi_objective.rs

@@ -0,0 +1,413 @@
+//! Multi-objective optimization via a dedicated study type.
+//!
+//! [`MultiObjectiveStudy`] manages trials that return **multiple** objective
+//! values simultaneously. It supports arbitrary numbers of objectives with
+//! per-objective directions (minimize or maximize).
+//!
+//! # Key concepts
+//!
+//! In multi-objective optimization there is usually no single best solution.
+//! Instead, there is a **Pareto front** — the set of solutions where no
+//! objective can be improved without worsening another. Use
+//! [`pareto_front()`](MultiObjectiveStudy::pareto_front) to retrieve these
+//! non-dominated solutions after optimization.
+//!
+//! A solution **dominates** another if it is at least as good in all
+//! objectives and strictly better in at least one. Solutions that are not
+//! dominated by any other are called **Pareto-optimal**.
+//!
+//! # Samplers
+//!
+//! By default a random sampler is used. For smarter search, pass a
+//! [`MultiObjectiveSampler`] such as [`Nsga2Sampler`](crate::sampler::Nsga2Sampler),
+//! [`Nsga3Sampler`](crate::sampler::Nsga3Sampler), or
+//! [`MoeadSampler`](crate::sampler::MoeadSampler) via
+//! [`MultiObjectiveStudy::with_sampler`].
+//!
+//! # Examples
+//!
+//! ```
+//! use optimizer::Direction;
+//! use optimizer::multi_objective::MultiObjectiveStudy;
+//! use optimizer::parameter::{FloatParam, Parameter};
+//!
+//! let study = MultiObjectiveStudy::new(vec![Direction::Minimize, Direction::Minimize]);
+//! let x = FloatParam::new(0.0, 1.0);
+//!
+//! study
+//!     .optimize(20, |trial: &mut optimizer::Trial| {
+//!         let xv = x.suggest(trial)?;
+//!         Ok::<_, optimizer::Error>(vec![xv, 1.0 - xv])
+//!     })
+//!     .unwrap();
+//!
+//! let front = study.pareto_front();
+//! assert!(!front.is_empty());
+//! ```
+
+use core::sync::atomic::{AtomicU64, Ordering};
+use std::collections::HashMap;
+use std::sync::Arc;
+
+use parking_lot::RwLock;
+
+use crate::distribution::Distribution;
+use crate::param::ParamValue;
+use crate::parameter::{ParamId, Parameter};
+use crate::pruner::NopPruner;
+use crate::sampler::random::RandomMultiObjectiveSampler;
+use crate::sampler::{CompletedTrial, Sampler};
+use crate::trial::{AttrValue, Trial};
+use crate::types::{Direction, TrialState};
+
+// ---------------------------------------------------------------------------
+// MultiObjectiveTrial
+// ---------------------------------------------------------------------------
+
+/// A completed trial with multiple objective values.
+///
+/// Each trial stores its sampled parameter values, the vector of
+/// objective values (one per objective), and optional constraint values.
+/// Retrieve typed parameter values with [`get()`](Self::get) and check
+/// constraint feasibility with [`is_feasible()`](Self::is_feasible).
+#[derive(Clone, Debug)]
+#[cfg_attr(feature = "serde", derive(serde::Serialize, serde::Deserialize))]
+pub struct MultiObjectiveTrial {
+    /// The unique identifier for this trial.
+    pub id: u64,
+    /// The sampled parameter values, keyed by parameter id.
+    pub params: HashMap<ParamId, ParamValue>,
+    /// The parameter distributions used, keyed by parameter id.
+    pub distributions: HashMap<ParamId, Distribution>,
+    /// Human-readable labels for parameters, keyed by parameter id.
+    pub param_labels: HashMap<ParamId, String>,
+    /// The objective values (one per objective).
+    pub values: Vec<f64>,
+    /// The state of the trial.
+    pub state: TrialState,
+    /// User-defined attributes stored during the trial.
+    pub user_attrs: HashMap<String, AttrValue>,
+    /// Constraint values for this trial (<=0.0 means feasible).
+    #[cfg_attr(feature = "serde", serde(default))]
+    pub constraints: Vec<f64>,
+}
+
+impl MultiObjectiveTrial {
+    /// Returns the typed value for the given parameter.
+    ///
+    /// Returns `None` if the parameter was not used in this trial.
+    ///
+    /// # Panics
+    ///
+    /// Panics if the stored value is incompatible with the parameter type.
+    pub fn get<P: Parameter>(&self, param: &P) -> Option<P::Value> {
+        self.params.get(&param.id()).map(|v| {
+            param
+                .cast_param_value(v)
+                .expect("parameter type mismatch: stored value incompatible with parameter")
+        })
+    }
+
+    /// Returns `true` if all constraints are satisfied (values <= 0.0).
+    ///
+    /// A trial with no constraints is considered feasible.
+    #[must_use]
+    pub fn is_feasible(&self) -> bool {
+        self.constraints.iter().all(|&c| c <= 0.0)
+    }
+
+    /// Gets a user attribute by key.
+    #[must_use]
+    pub fn user_attr(&self, key: &str) -> Option<&AttrValue> {
+        self.user_attrs.get(key)
+    }
+
+    /// Returns all user attributes.
+    #[must_use]
+    pub fn user_attrs(&self) -> &HashMap<String, AttrValue> {
+        &self.user_attrs
+    }
+}
+
+// ---------------------------------------------------------------------------
+// MultiObjectiveSampler trait
+// ---------------------------------------------------------------------------
+
+/// Trait for samplers aware of multi-objective history.
+///
+/// Separate from [`Sampler`] because multi-objective algorithms (e.g.,
+/// NSGA-II) need access to the full vector of objective values per trial
+/// (`&[MultiObjectiveTrial]`) and the per-objective directions
+/// (`&[Direction]`).
+///
+/// Implementations include [`Nsga2Sampler`](crate::sampler::Nsga2Sampler),
+/// [`Nsga3Sampler`](crate::sampler::Nsga3Sampler), and
+/// [`MoeadSampler`](crate::sampler::MoeadSampler).
+pub trait MultiObjectiveSampler: Send + Sync {
+    /// Samples a parameter value from the given distribution.
+    fn sample(
+        &self,
+        distribution: &Distribution,
+        trial_id: u64,
+        history: &[MultiObjectiveTrial],
+        directions: &[Direction],
+    ) -> ParamValue;
+}
+
+// ---------------------------------------------------------------------------
+// MoSamplerBridge — bridges MultiObjectiveSampler to Sampler trait
+// ---------------------------------------------------------------------------
+
+/// Bridges a [`MultiObjectiveSampler`] to the [`Sampler`] trait so that
+/// `Trial::with_sampler()` can use it.
+struct MoSamplerBridge {
+    inner: Arc<dyn MultiObjectiveSampler>,
+    history: Arc<RwLock<Vec<MultiObjectiveTrial>>>,
+    directions: Vec<Direction>,
+}
+
+impl Sampler for MoSamplerBridge {
+    fn sample(
+        &self,
+        distribution: &Distribution,
+        trial_id: u64,
+        _history: &[CompletedTrial],
+    ) -> ParamValue {
+        let mo_history = self.history.read();
+        self.inner
+            .sample(distribution, trial_id, &mo_history, &self.directions)
+    }
+}
+
+// ---------------------------------------------------------------------------
+// MultiObjectiveStudy
+// ---------------------------------------------------------------------------
+
+/// A study for multi-objective optimization.
+///
+/// Manage trials that return multiple objective values. Supports
+/// arbitrary numbers of objectives with independent minimize/maximize
+/// directions. After optimization, call [`pareto_front()`](Self::pareto_front)
+/// to retrieve the non-dominated solutions.
+///
+/// For single-objective optimization, use [`Study`](crate::Study) instead.
+///
+/// # Examples
+///
+/// ```
+/// use optimizer::Direction;
+/// use optimizer::multi_objective::MultiObjectiveStudy;
+/// use optimizer::parameter::{FloatParam, Parameter};
+///
+/// // Bi-objective: minimize both
+/// let study = MultiObjectiveStudy::new(vec![Direction::Minimize, Direction::Minimize]);
+/// let x = FloatParam::new(0.0, 1.0);
+///
+/// study
+///     .optimize(30, |trial: &mut optimizer::Trial| {
+///         let xv = x.suggest(trial)?;
+///         Ok::<_, optimizer::Error>(vec![xv, 1.0 - xv])
+///     })
+///     .unwrap();
+///
+/// let front = study.pareto_front();
+/// assert!(!front.is_empty());
+/// ```
+pub struct MultiObjectiveStudy {
+    directions: Vec<Direction>,
+    sampler: Arc<dyn MultiObjectiveSampler>,
+    completed_trials: Arc<RwLock<Vec<MultiObjectiveTrial>>>,
+    next_trial_id: AtomicU64,
+}
+
+impl MultiObjectiveStudy {
+    /// Creates a new multi-objective study with the given directions.
+    ///
+    /// Uses a random sampler by default.
+    ///
+    /// # Arguments
+    ///
+    /// * `directions` - One direction per objective (minimize or maximize).
+    #[must_use]
+    pub fn new(directions: Vec<Direction>) -> Self {
+        Self {
+            directions,
+            sampler: Arc::new(RandomMultiObjectiveSampler::new()),
+            completed_trials: Arc::new(RwLock::new(Vec::new())),
+            next_trial_id: AtomicU64::new(0),
+        }
+    }
+
+    /// Creates a new study with a custom multi-objective sampler.
+    #[must_use]
+    pub fn with_sampler(
+        directions: Vec<Direction>,
+        sampler: impl MultiObjectiveSampler + 'static,
+    ) -> Self {
+        Self {
+            directions,
+            sampler: Arc::new(sampler),
+            completed_trials: Arc::new(RwLock::new(Vec::new())),
+            next_trial_id: AtomicU64::new(0),
+        }
+    }
+
+    /// Returns the optimization directions.
+    #[must_use]
+    pub fn directions(&self) -> &[Direction] {
+        &self.directions
+    }
+
+    /// Returns the number of objectives.
+    #[must_use]
+    pub fn n_objectives(&self) -> usize {
+        self.directions.len()
+    }
+
+    /// Returns the number of completed trials.
+    #[must_use]
+    pub fn n_trials(&self) -> usize {
+        self.completed_trials.read().len()
+    }
+
+    /// Returns all completed trials.
+    #[must_use]
+    pub fn trials(&self) -> Vec<MultiObjectiveTrial> {
+        self.completed_trials.read().clone()
+    }
+
+    /// Return the Pareto-optimal trials (the non-dominated front).
+    ///
+    /// Uses fast non-dominated sorting (Deb et al., 2002) from the
+    /// [`pareto`](crate::pareto) module. Returns an empty vec if no
+    /// trials have completed.
+    #[must_use]
+    pub fn pareto_front(&self) -> Vec<MultiObjectiveTrial> {
+        let trials = self.completed_trials.read();
+        let complete: Vec<_> = trials
+            .iter()
+            .filter(|t| t.state == TrialState::Complete)
+            .collect();
+
+        if complete.is_empty() {
+            return Vec::new();
+        }
+
+        let values: Vec<Vec<f64>> = complete.iter().map(|t| t.values.clone()).collect();
+        let fronts = crate::pareto::fast_non_dominated_sort(&values, &self.directions);
+
+        if fronts.is_empty() {
+            return Vec::new();
+        }
+
+        fronts[0].iter().map(|&i| complete[i].clone()).collect()
+    }
+
+    /// Creates a new trial wired to the study's MO sampler.
+    fn create_trial(&self) -> Trial {
+        let id = self.next_trial_id.fetch_add(1, Ordering::SeqCst);
+
+        let bridge: Arc<dyn Sampler> = Arc::new(MoSamplerBridge {
+            inner: Arc::clone(&self.sampler),
+            history: Arc::clone(&self.completed_trials),
+            directions: self.directions.clone(),
+        });
+
+        // Dummy f64 history — the bridge ignores it.
+        let dummy_history: Arc<RwLock<Vec<CompletedTrial<f64>>>> =
+            Arc::new(RwLock::new(Vec::new()));
+
+        Trial::with_sampler(id, bridge, dummy_history, Arc::new(NopPruner))
+    }
+
+    /// Records a completed trial.
+    fn complete_trial(&self, trial: Trial, values: Vec<f64>) {
+        let mo_trial = trial.into_multi_objective_trial(values, TrialState::Complete);
+        self.completed_trials.write().push(mo_trial);
+    }
+
+    /// Records a failed trial (not stored in history).
+    fn fail_trial(trial: &mut Trial) {
+        trial.set_failed();
+    }
+
+    /// Request a new trial for the ask/tell interface.
+    ///
+    /// After creating the trial, suggest parameters on it, evaluate your
+    /// objective externally, then pass the trial back to [`tell()`](Self::tell).
+    pub fn ask(&self) -> Trial {
+        self.create_trial()
+    }
+
+    /// Report the result of a trial obtained from [`ask()`](Self::ask).
+    ///
+    /// Pass `Ok(values)` for a successful evaluation or `Err(reason)` for a failure.
+    ///
+    /// # Errors
+    ///
+    /// Returns `ObjectiveDimensionMismatch` if the number of values doesn't
+    /// match the number of directions.
+    pub fn tell(
+        &self,
+        mut trial: Trial,
+        result: core::result::Result<Vec<f64>, impl ToString>,
+    ) -> crate::Result<()> {
+        if let Ok(values) = result {
+            if values.len() != self.directions.len() {
+                return Err(crate::Error::ObjectiveDimensionMismatch {
+                    expected: self.directions.len(),
+                    got: values.len(),
+                });
+            }
+            self.complete_trial(trial, values);
+        } else {
+            Self::fail_trial(&mut trial);
+        }
+        Ok(())
+    }
+
+    /// Runs multi-objective optimization for `n_trials` trials.
+    ///
+    /// The objective function must return a `Vec<f64>` with one value per
+    /// objective.
+    ///
+    /// # Errors
+    ///
+    /// Returns `ObjectiveDimensionMismatch` if the objective returns the wrong
+    /// number of values. Returns `NoCompletedTrials` if all trials fail.
+    pub fn optimize<F, E>(&self, n_trials: usize, mut objective: F) -> crate::Result<()>
+    where
+        F: FnMut(&mut Trial) -> core::result::Result<Vec<f64>, E>,
+        E: ToString,
+    {
+        for _ in 0..n_trials {
+            let mut trial = self.create_trial();
+
+            match objective(&mut trial) {
+                Ok(values) => {
+                    if values.len() != self.directions.len() {
+                        return Err(crate::Error::ObjectiveDimensionMismatch {
+                            expected: self.directions.len(),
+                            got: values.len(),
+                        });
+                    }
+                    self.complete_trial(trial, values);
+                }
+                Err(_) => {
+                    Self::fail_trial(&mut trial);
+                }
+            }
+        }
+
+        let has_complete = self
+            .completed_trials
+            .read()
+            .iter()
+            .any(|t| t.state == TrialState::Complete);
+        if !has_complete {
+            return Err(crate::Error::NoCompletedTrials);
+        }
+
+        Ok(())
+    }
+}

src/objective.rs

@@ -0,0 +1,146 @@
+//! The [`Objective`] trait defines what gets optimized.
+//!
+//! # Closures work directly
+//!
+//! Any `Fn(&mut Trial) -> Result<V, E>` closure automatically implements
+//! [`Objective`], so you can pass closures straight to
+//! [`Study::optimize`](crate::Study::optimize):
+//!
+//! ```
+//! use optimizer::prelude::*;
+//!
+//! let study: Study<f64> = Study::new(Direction::Minimize);
+//! let x = FloatParam::new(-10.0, 10.0).name("x");
+//!
+//! study
+//!     .optimize(50, |trial: &mut optimizer::Trial| {
+//!         let v = x.suggest(trial)?;
+//!         Ok::<_, Error>((v - 3.0).powi(2))
+//!     })
+//!     .unwrap();
+//! ```
+//!
+//! # Structs for lifecycle hooks
+//!
+//! For richer control — early stopping or per-trial logging — implement
+//! [`Objective`] on a struct and pass it to the same
+//! [`Study::optimize`](crate::Study::optimize) method:
+//!
+//! ```
+//! use std::ops::ControlFlow;
+//!
+//! use optimizer::Objective;
+//! use optimizer::prelude::*;
+//!
+//! struct QuadraticWithEarlyStopping {
+//!     x: FloatParam,
+//!     target: f64,
+//! }
+//!
+//! impl Objective<f64> for QuadraticWithEarlyStopping {
+//!     type Error = Error;
+//!
+//!     fn evaluate(&self, trial: &mut Trial) -> Result<f64> {
+//!         let v = self.x.suggest(trial)?;
+//!         Ok((v - 3.0).powi(2))
+//!     }
+//!
+//!     fn after_trial(&self, _study: &Study<f64>, trial: &CompletedTrial<f64>) -> ControlFlow<()> {
+//!         if trial.value < self.target {
+//!             ControlFlow::Break(())
+//!         } else {
+//!             ControlFlow::Continue(())
+//!         }
+//!     }
+//! }
+//!
+//! let study: Study<f64> = Study::new(Direction::Minimize);
+//! let obj = QuadraticWithEarlyStopping {
+//!     x: FloatParam::new(-10.0, 10.0).name("x"),
+//!     target: 1.0,
+//! };
+//! study.optimize(200, obj).unwrap();
+//! assert!(study.best_value().unwrap() < 1.0);
+//! ```
+
+use core::ops::ControlFlow;
+
+use crate::sampler::CompletedTrial;
+use crate::study::Study;
+use crate::trial::Trial;
+
+/// Defines an objective function with lifecycle hooks for optimization.
+///
+/// The only required method is [`evaluate`](Objective::evaluate), which
+/// computes the objective value for a given trial. Optional hooks provide
+/// early stopping ([`before_trial`](Objective::before_trial),
+/// [`after_trial`](Objective::after_trial)).
+///
+/// # Closures implement `Objective` automatically
+///
+/// A blanket implementation covers all `Fn(&mut Trial) -> Result<V, E>`
+/// closures, so you can pass closures directly to
+/// [`Study::optimize`](crate::Study::optimize) without wrapping them.
+///
+/// # Thread safety
+///
+/// The async optimization methods (`optimize_async`, `optimize_parallel`)
+/// additionally require `Send + Sync + 'static` on the objective. The
+/// sync `optimize` method has no thread-safety requirements.
+pub trait Objective<V: PartialOrd = f64> {
+    /// The error type returned by [`evaluate`](Objective::evaluate).
+    type Error: ToString + 'static;
+
+    /// Evaluate the objective function for a single trial.
+    ///
+    /// Sample parameters from `trial` via
+    /// [`Parameter::suggest`](crate::parameter::Parameter::suggest) and
+    /// return the objective value. Return `Err(TrialPruned)` to prune a
+    /// trial early.
+    ///
+    /// # Errors
+    ///
+    /// Any error whose type implements `ToString`. Pruning errors
+    /// (`Error::TrialPruned` or `TrialPruned`) are handled specially —
+    /// the trial is recorded as pruned rather than failed.
+    fn evaluate(&self, trial: &mut Trial) -> Result<V, Self::Error>;
+
+    /// Called before each trial is created.
+    ///
+    /// Return `ControlFlow::Break(())` to stop the optimization loop
+    /// before the next trial starts.
+    ///
+    /// Default: always continues.
+    fn before_trial(&self, _study: &Study<V>) -> ControlFlow<()> {
+        ControlFlow::Continue(())
+    }
+
+    /// Called after each **completed** trial (not failed or pruned).
+    ///
+    /// The trial is passed directly as the argument *before* it is pushed
+    /// to storage, so `study.n_trials()` and `study.trials()` do not yet
+    /// include this trial. The trial is always pushed to storage after this
+    /// callback returns, regardless of the return value.
+    ///
+    /// Return `ControlFlow::Break(())` to stop the optimization loop.
+    ///
+    /// Default: always continues.
+    fn after_trial(&self, _study: &Study<V>, _trial: &CompletedTrial<V>) -> ControlFlow<()> {
+        ControlFlow::Continue(())
+    }
+}
+
+/// Blanket implementation: any `Fn(&mut Trial) -> Result<V, E>` is an
+/// `Objective` with no lifecycle hooks.
+impl<F, V, E> Objective<V> for F
+where
+    F: Fn(&mut Trial) -> Result<V, E>,
+    V: PartialOrd,
+    E: ToString + 'static,
+{
+    type Error = E;
+
+    fn evaluate(&self, trial: &mut Trial) -> Result<V, E> {
+        self(trial)
+    }
+}

src/param.rs

@@ -1,18 +1,35 @@
-//! Parameter value storage types.
+//! Raw parameter value storage.
+//!
+//! [`ParamValue`] is the type-erased representation of a sampled parameter.
+//! Users rarely construct `ParamValue` directly — the
+//! [`Parameter::suggest`](crate::parameter::Parameter::suggest) method returns
+//! the already-typed value (e.g., `f64` for [`FloatParam`](crate::parameter::FloatParam)).
+//!
+//! `ParamValue` is useful when inspecting raw trial data via
+//! [`Trial::params`](crate::Trial::params) or
+//! [`CompletedTrial::params`](crate::sampler::CompletedTrial).
 
-/// Represents a sampled parameter value.
+/// A type-erased sampled parameter value.
 ///
-/// This enum stores different parameter value types uniformly.
-/// For categorical parameters, the `Categorical` variant stores
-/// the index into the choices array.
+/// Stores float, integer, or categorical (index) values uniformly.
+/// For categorical parameters the `Categorical` variant stores the
+/// zero-based index into the choices array, not the choice itself.
+///
+/// # Display
+///
+/// `ParamValue` implements [`Display`](core::fmt::Display): floats and
+/// integers print their numeric value, and categoricals print `category(i)`.
 #[derive(Clone, Debug, PartialEq)]
 #[cfg_attr(feature = "serde", derive(serde::Serialize, serde::Deserialize))]
 pub enum ParamValue {
-    /// A floating-point parameter value.
+    /// A floating-point parameter value (from [`FloatParam`](crate::parameter::FloatParam)).
     Float(f64),
-    /// An integer parameter value.
+    /// An integer parameter value (from [`IntParam`](crate::parameter::IntParam)).
     Int(i64),
-    /// A categorical parameter value, stored as an index into the choices array.
+    /// A categorical index into the choices array (from
+    /// [`CategoricalParam`](crate::parameter::CategoricalParam),
+    /// [`BoolParam`](crate::parameter::BoolParam), or
+    /// [`EnumParam`](crate::parameter::EnumParam)).
     Categorical(usize),
 }
 

src/parameter.rs

@@ -1,8 +1,19 @@
-//! Central parameter trait and built-in parameter types.
+//! Parameter trait and five built-in parameter types.
 //!
-//! The [`Parameter`] trait provides a unified way to define parameter types
-//! and suggest values from a [`Trial`]. Built-in implementations
-//! cover floats, integers, categoricals, booleans, and enum types.
+//! The [`Parameter`] trait provides a unified way to define search-space
+//! dimensions and sample values from a [`Trial`]. Five implementations
+//! cover the most common hyperparameter types:
+//!
+//! | Type | Sampled value | Typical use |
+//! |------|---------------|-------------|
+//! | [`FloatParam`] | `f64` | Learning rate, dropout probability |
+//! | [`IntParam`] | `i64` | Layer count, batch size |
+//! | [`CategoricalParam`] | `T: Clone` | Optimizer name, activation function |
+//! | [`BoolParam`] | `bool` | Feature toggle |
+//! | [`EnumParam`] | `T: Categorical` | Typed enum variant selection |
+//!
+//! All five types support `.name()` for a human-readable label and
+//! `.suggest(&mut trial)` as a shorthand for `trial.suggest_param(&param)`.
 //!
 //! # Example
 //!
@@ -14,10 +25,17 @@
 //!
 //! let lr = FloatParam::new(1e-5, 1e-1)
 //!     .log_scale()
+//!     .name("learning_rate")
+//!     .suggest(&mut trial)
+//!     .unwrap();
+//! let layers = IntParam::new(1, 10)
+//!     .name("n_layers")
+//!     .suggest(&mut trial)
+//!     .unwrap();
+//! let dropout = BoolParam::new()
+//!     .name("use_dropout")
 //!     .suggest(&mut trial)
 //!     .unwrap();
-//! let layers = IntParam::new(1, 10).suggest(&mut trial).unwrap();
-//! let dropout = BoolParam::new().suggest(&mut trial).unwrap();
 //! ```
 
 use core::fmt::Debug;
@@ -28,7 +46,7 @@ use crate::distribution::{
     CategoricalDistribution, Distribution, FloatDistribution, IntDistribution,
 };
 use crate::error::{Error, Result};
-use crate::param::ParamValue;
+pub use crate::param::ParamValue;
 use crate::trial::Trial;
 
 static NEXT_PARAM_ID: AtomicU64 = AtomicU64::new(0);
@@ -42,7 +60,7 @@ static NEXT_PARAM_ID: AtomicU64 = AtomicU64::new(0);
 pub struct ParamId(u64);
 
 impl ParamId {
-    /// Creates a new unique `ParamId`.
+    /// Create a new unique `ParamId`.
     pub fn new() -> Self {
         Self(NEXT_PARAM_ID.fetch_add(1, Ordering::Relaxed))
     }
@@ -60,52 +78,67 @@ impl core::fmt::Display for ParamId {
     }
 }
 
-/// A trait for defining parameter types that can be suggested by a [`Trial`].
+/// Define a parameter type that can be suggested by a [`Trial`].
 ///
 /// Implementors specify the distribution to sample from and how to convert
-/// the raw [`ParamValue`] back into a typed value.
+/// the raw [`ParamValue`] back into a typed value. See the five built-in
+/// implementations: [`FloatParam`], [`IntParam`], [`CategoricalParam`],
+/// [`BoolParam`], and [`EnumParam`].
 pub trait Parameter: Debug {
     /// The typed value returned after sampling.
     type Value;
 
-    /// Returns the unique identifier for this parameter.
+    /// Return the unique identifier for this parameter.
     fn id(&self) -> ParamId;
 
-    /// Returns the distribution that this parameter samples from.
+    /// Return the distribution that this parameter samples from.
     fn distribution(&self) -> Distribution;
 
-    /// Converts a raw [`ParamValue`] into the typed value.
+    /// Convert a raw [`ParamValue`] into the typed value.
     ///
     /// # Errors
     ///
-    /// Returns an error if the `ParamValue` variant doesn't match what this parameter expects.
+    /// Return an error if the `ParamValue` variant does not match what this parameter expects.
     fn cast_param_value(&self, param_value: &ParamValue) -> Result<Self::Value>;
 
-    /// Validates the parameter configuration.
+    /// Validate the parameter configuration.
     ///
     /// Called before sampling. The default implementation accepts all configurations.
     ///
     /// # Errors
     ///
-    /// Returns an error if the parameter configuration is invalid.
+    /// Return an error if the parameter configuration is invalid.
     fn validate(&self) -> Result<()> {
         Ok(())
     }
 
-    /// Returns a human-readable label for this parameter.
+    /// Return a human-readable label for this parameter.
     ///
-    /// Defaults to the `Debug` output of the parameter.
+    /// Defaults to the `Debug` output of the parameter. Override with
+    /// the `.name()` builder method on concrete types.
     fn label(&self) -> String {
         format!("{self:?}")
     }
 
-    /// Suggests a value for this parameter from the given trial.
+    /// Suggest a value for this parameter from the given trial.
     ///
     /// This is a convenience method that delegates to [`Trial::suggest_param`].
     ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::Trial;
+    /// use optimizer::parameter::{FloatParam, Parameter};
+    ///
+    /// let mut trial = Trial::new(0);
+    /// let param = FloatParam::new(-5.0, 5.0).name("x");
+    /// let value: f64 = param.suggest(&mut trial).unwrap();
+    /// assert!((-5.0..=5.0).contains(&value));
+    /// ```
+    ///
     /// # Errors
     ///
-    /// Returns an error if validation fails, the parameter conflicts with
+    /// Return an error if validation fails, the parameter conflicts with
     /// a previously suggested parameter of the same id, or sampling fails.
     fn suggest(&self, trial: &mut Trial) -> Result<Self::Value>
     where
@@ -117,7 +150,7 @@ pub trait Parameter: Debug {
 
 /// A floating-point parameter with optional log-scale and step size.
 ///
-/// # Example
+/// # Examples
 ///
 /// ```
 /// use optimizer::Trial;
@@ -128,13 +161,14 @@ pub trait Parameter: Debug {
 /// // Simple range
 /// let x = FloatParam::new(0.0, 1.0).suggest(&mut trial).unwrap();
 ///
-/// // Log-scale
+/// // Log-scale with a human-readable name
 /// let lr = FloatParam::new(1e-5, 1e-1)
 ///     .log_scale()
+///     .name("learning_rate")
 ///     .suggest(&mut trial)
 ///     .unwrap();
 ///
-/// // Stepped
+/// // Stepped (values will be multiples of 0.25)
 /// let step = FloatParam::new(0.0, 1.0)
 ///     .step(0.25)
 ///     .suggest(&mut trial)
@@ -151,7 +185,7 @@ pub struct FloatParam {
 }
 
 impl FloatParam {
-    /// Creates a new float parameter with the given bounds.
+    /// Create a new float parameter sampling uniformly from `[low, high]`.
     #[must_use]
     pub fn new(low: f64, high: f64) -> Self {
         Self {
@@ -164,21 +198,21 @@ impl FloatParam {
         }
     }
 
-    /// Enables log-scale sampling.
+    /// Enable log-scale sampling (bounds must be positive).
     #[must_use]
     pub fn log_scale(mut self) -> Self {
         self.log_scale = true;
         self
     }
 
-    /// Sets a step size for discretized sampling.
+    /// Set a step size for discretized sampling.
     #[must_use]
     pub fn step(mut self, step: f64) -> Self {
         self.step = Some(step);
         self
     }
 
-    /// Sets a human-readable name for this parameter.
+    /// Set a human-readable name for this parameter.
     ///
     /// When set, this name is used as the parameter's label instead of
     /// the default `Debug` output.
@@ -221,6 +255,12 @@ impl Parameter for FloatParam {
     }
 
     fn validate(&self) -> Result<()> {
+        if !self.low.is_finite() || !self.high.is_finite() {
+            return Err(Error::InvalidBounds {
+                low: self.low,
+                high: self.high,
+            });
+        }
         if self.low > self.high {
             return Err(Error::InvalidBounds {
                 low: self.low,
@@ -231,7 +271,7 @@ impl Parameter for FloatParam {
             return Err(Error::InvalidLogBounds);
         }
         if let Some(step) = self.step
-            && step <= 0.0
+            && (!step.is_finite() || step <= 0.0)
         {
             return Err(Error::InvalidStep);
         }
@@ -245,7 +285,7 @@ impl Parameter for FloatParam {
 
 /// An integer parameter with optional log-scale and step size.
 ///
-/// # Example
+/// # Examples
 ///
 /// ```
 /// use optimizer::Trial;
@@ -254,15 +294,19 @@ impl Parameter for FloatParam {
 /// let mut trial = Trial::new(0);
 ///
 /// // Simple range
-/// let n = IntParam::new(1, 10).suggest(&mut trial).unwrap();
+/// let n = IntParam::new(1, 10)
+///     .name("n_layers")
+///     .suggest(&mut trial)
+///     .unwrap();
 ///
 /// // Log-scale
 /// let batch = IntParam::new(1, 1024)
 ///     .log_scale()
+///     .name("batch_size")
 ///     .suggest(&mut trial)
 ///     .unwrap();
 ///
-/// // Stepped
+/// // Stepped (multiples of 32)
 /// let units = IntParam::new(32, 512).step(32).suggest(&mut trial).unwrap();
 /// ```
 #[derive(Clone, Debug)]
@@ -276,7 +320,7 @@ pub struct IntParam {
 }
 
 impl IntParam {
-    /// Creates a new integer parameter with the given bounds.
+    /// Create a new integer parameter sampling uniformly from `[low, high]`.
     #[must_use]
     pub fn new(low: i64, high: i64) -> Self {
         Self {
@@ -289,21 +333,21 @@ impl IntParam {
         }
     }
 
-    /// Enables log-scale sampling.
+    /// Enable log-scale sampling (bounds must be ≥ 1).
     #[must_use]
     pub fn log_scale(mut self) -> Self {
         self.log_scale = true;
         self
     }
 
-    /// Sets a step size for discretized sampling.
+    /// Set a step size for discretized sampling.
     #[must_use]
     pub fn step(mut self, step: i64) -> Self {
         self.step = Some(step);
         self
     }
 
-    /// Sets a human-readable name for this parameter.
+    /// Set a human-readable name for this parameter.
     ///
     /// When set, this name is used as the parameter's label instead of
     /// the default `Debug` output.
@@ -370,7 +414,10 @@ impl Parameter for IntParam {
 
 /// A categorical parameter that selects from a list of choices.
 ///
-/// # Example
+/// The generic type `T` is the element type of the choices vector.
+/// The sampler picks an index and the corresponding element is returned.
+///
+/// # Examples
 ///
 /// ```
 /// use optimizer::Trial;
@@ -378,6 +425,7 @@ impl Parameter for IntParam {
 ///
 /// let mut trial = Trial::new(0);
 /// let opt = CategoricalParam::new(vec!["sgd", "adam", "rmsprop"])
+///     .name("optimizer")
 ///     .suggest(&mut trial)
 ///     .unwrap();
 /// ```
@@ -389,7 +437,7 @@ pub struct CategoricalParam<T: Clone> {
 }
 
 impl<T: Clone> CategoricalParam<T> {
-    /// Creates a new categorical parameter with the given choices.
+    /// Create a new categorical parameter with the given choices.
     #[must_use]
     pub fn new(choices: Vec<T>) -> Self {
         Self {
@@ -399,7 +447,7 @@ impl<T: Clone> CategoricalParam<T> {
         }
     }
 
-    /// Sets a human-readable name for this parameter.
+    /// Set a human-readable name for this parameter.
     ///
     /// When set, this name is used as the parameter's label instead of
     /// the default `Debug` output.
@@ -425,7 +473,11 @@ impl<T: Clone + Debug> Parameter for CategoricalParam<T> {
 
     fn cast_param_value(&self, param_value: &ParamValue) -> Result<T> {
         match param_value {
-            ParamValue::Categorical(index) => Ok(self.choices[*index].clone()),
+            ParamValue::Categorical(index) => self
+                .choices
+                .get(*index)
+                .cloned()
+                .ok_or(Error::Internal("categorical index out of bounds")),
             _ => Err(Error::Internal(
                 "Categorical distribution should return Categorical value",
             )),
@@ -444,16 +496,19 @@ impl<T: Clone + Debug> Parameter for CategoricalParam<T> {
     }
 }
 
-/// A boolean parameter (equivalent to a categorical with `[false, true]`).
+/// A boolean parameter (equivalent to a two-choice categorical: `false` / `true`).
 ///
-/// # Example
+/// # Examples
 ///
 /// ```
 /// use optimizer::Trial;
 /// use optimizer::parameter::{BoolParam, Parameter};
 ///
 /// let mut trial = Trial::new(0);
-/// let dropout = BoolParam::new().suggest(&mut trial).unwrap();
+/// let use_dropout = BoolParam::new()
+///     .name("use_dropout")
+///     .suggest(&mut trial)
+///     .unwrap();
 /// ```
 #[derive(Clone, Debug)]
 pub struct BoolParam {
@@ -462,7 +517,7 @@ pub struct BoolParam {
 }
 
 impl BoolParam {
-    /// Creates a new boolean parameter.
+    /// Create a new boolean parameter.
     #[must_use]
     pub fn new() -> Self {
         Self {
@@ -471,7 +526,7 @@ impl BoolParam {
         }
     }
 
-    /// Sets a human-readable name for this parameter.
+    /// Set a human-readable name for this parameter.
     ///
     /// When set, this name is used as the parameter's label instead of
     /// the default `Debug` output.
@@ -501,7 +556,8 @@ impl Parameter for BoolParam {
 
     fn cast_param_value(&self, param_value: &ParamValue) -> Result<bool> {
         match param_value {
-            ParamValue::Categorical(index) => Ok(*index != 0),
+            ParamValue::Categorical(index) if *index < 2 => Ok(*index != 0),
+            ParamValue::Categorical(_) => Err(Error::Internal("bool index out of bounds")),
             _ => Err(Error::Internal(
                 "Categorical distribution should return Categorical value",
             )),
@@ -513,10 +569,10 @@ impl Parameter for BoolParam {
     }
 }
 
-/// A trait for enum types that can be used as categorical parameters.
+/// Map an enum type to sequential indices for use as a categorical parameter.
 ///
-/// This trait maps enum variants to sequential indices and back. It can be
-/// derived automatically for fieldless enums using `#[derive(Categorical)]`
+/// This trait converts enum variants to sequential indices and back. It can
+/// be derived automatically for fieldless enums using `#[derive(Categorical)]`
 /// when the `derive` feature is enabled.
 ///
 /// # Example
@@ -558,20 +614,24 @@ pub trait Categorical: Sized + Clone {
     /// The number of variants in the enum.
     const N_CHOICES: usize;
 
-    /// Creates an instance from a variant index.
+    /// Create an instance from a variant index.
     ///
     /// # Panics
     ///
     /// Panics if `index >= N_CHOICES`.
     fn from_index(index: usize) -> Self;
 
-    /// Returns the index of this variant.
+    /// Return the index of this variant.
     fn to_index(&self) -> usize;
 }
 
 /// A parameter that selects from the variants of an enum implementing [`Categorical`].
 ///
-/// # Example
+/// Prefer this over [`CategoricalParam`] when the choices map to a Rust enum,
+/// because the returned value is already the correct variant — no string
+/// matching required.
+///
+/// # Examples
 ///
 /// ```
 /// use optimizer::Trial;
@@ -604,7 +664,10 @@ pub trait Categorical: Sized + Clone {
 /// }
 ///
 /// let mut trial = Trial::new(0);
-/// let opt = EnumParam::<Optimizer>::new().suggest(&mut trial).unwrap();
+/// let opt = EnumParam::<Optimizer>::new()
+///     .name("optimizer")
+///     .suggest(&mut trial)
+///     .unwrap();
 /// ```
 #[derive(Clone, Debug)]
 pub struct EnumParam<T: Categorical> {
@@ -614,7 +677,7 @@ pub struct EnumParam<T: Categorical> {
 }
 
 impl<T: Categorical> EnumParam<T> {
-    /// Creates a new enum parameter.
+    /// Create a new enum parameter over all variants of `T`.
     #[must_use]
     pub fn new() -> Self {
         Self {
@@ -624,7 +687,7 @@ impl<T: Categorical> EnumParam<T> {
         }
     }
 
-    /// Sets a human-readable name for this parameter.
+    /// Set a human-readable name for this parameter.
     ///
     /// When set, this name is used as the parameter's label instead of
     /// the default `Debug` output.
@@ -656,7 +719,8 @@ impl<T: Categorical + Debug> Parameter for EnumParam<T> {
 
     fn cast_param_value(&self, param_value: &ParamValue) -> Result<T> {
         match param_value {
-            ParamValue::Categorical(index) => Ok(T::from_index(*index)),
+            ParamValue::Categorical(index) if *index < T::N_CHOICES => Ok(T::from_index(*index)),
+            ParamValue::Categorical(_) => Err(Error::Internal("categorical index out of bounds")),
             _ => Err(Error::Internal(
                 "Categorical distribution should return Categorical value",
             )),
@@ -732,6 +796,30 @@ mod tests {
         assert!(param.validate().is_err());
     }
 
+    #[test]
+    fn float_param_validate_nan() {
+        assert!(FloatParam::new(f64::NAN, 1.0).validate().is_err());
+        assert!(FloatParam::new(0.0, f64::NAN).validate().is_err());
+        assert!(FloatParam::new(f64::NAN, f64::NAN).validate().is_err());
+    }
+
+    #[test]
+    fn float_param_validate_infinity() {
+        assert!(FloatParam::new(f64::INFINITY, 1.0).validate().is_err());
+        assert!(FloatParam::new(0.0, f64::NEG_INFINITY).validate().is_err());
+    }
+
+    #[test]
+    fn float_param_validate_nan_step() {
+        assert!(FloatParam::new(0.0, 1.0).step(f64::NAN).validate().is_err());
+        assert!(
+            FloatParam::new(0.0, 1.0)
+                .step(f64::INFINITY)
+                .validate()
+                .is_err()
+        );
+    }
+
     #[test]
     #[allow(clippy::float_cmp)]
     fn float_param_cast_param_value() {
@@ -835,6 +923,17 @@ mod tests {
         assert!(param.cast_param_value(&ParamValue::Float(1.0)).is_err());
     }
 
+    #[test]
+    fn categorical_param_cast_out_of_bounds() {
+        let param = CategoricalParam::new(vec!["sgd", "adam", "rmsprop"]);
+        assert!(param.cast_param_value(&ParamValue::Categorical(3)).is_err());
+        assert!(
+            param
+                .cast_param_value(&ParamValue::Categorical(usize::MAX))
+                .is_err()
+        );
+    }
+
     #[test]
     fn bool_param_distribution() {
         let param = BoolParam::new();
@@ -852,6 +951,13 @@ mod tests {
         assert!(param.cast_param_value(&ParamValue::Float(1.0)).is_err());
     }
 
+    #[test]
+    fn bool_param_cast_out_of_bounds() {
+        let param = BoolParam::new();
+        assert!(param.cast_param_value(&ParamValue::Categorical(2)).is_err());
+        assert!(param.cast_param_value(&ParamValue::Categorical(5)).is_err());
+    }
+
     #[derive(Clone, Debug, PartialEq)]
     enum TestEnum {
         A,
@@ -903,6 +1009,17 @@ mod tests {
         assert!(param.cast_param_value(&ParamValue::Float(1.0)).is_err());
     }
 
+    #[test]
+    fn enum_param_cast_out_of_bounds() {
+        let param = EnumParam::<TestEnum>::new();
+        assert!(param.cast_param_value(&ParamValue::Categorical(3)).is_err());
+        assert!(
+            param
+                .cast_param_value(&ParamValue::Categorical(usize::MAX))
+                .is_err()
+        );
+    }
+
     #[test]
     fn float_param_suggest_via_trial() {
         let param = FloatParam::new(0.0, 1.0);

src/pareto.rs

@@ -0,0 +1,660 @@
+//! Pareto front analysis utilities for multi-objective optimization.
+//!
+//! In multi-objective optimization there is generally no single best
+//! solution. Instead, the goal is to find the **Pareto front** — the set
+//! of solutions where no objective can be improved without worsening
+//! another. This module provides tools for computing and analyzing Pareto
+//! fronts.
+//!
+//! # Available functions
+//!
+//! | Function | Purpose |
+//! |---|---|
+//! | [`hypervolume`] | Measure the quality of a Pareto front (volume of dominated space) |
+//! | [`non_dominated_sort`] | Rank solutions into successive fronts (front 0, 1, …) |
+//! | [`pareto_front_indices`] | Filter to non-dominated (Pareto-optimal) solutions only |
+//! | [`crowding_distance`] | Measure diversity/spread within a single front |
+//!
+//! # When to use
+//!
+//! - **Evaluating front quality**: Use [`hypervolume`] to compare two
+//!   Pareto fronts — a higher hypervolume indicates a better-quality front.
+//! - **Ranking all solutions**: Use [`non_dominated_sort`] to partition
+//!   solutions into successive fronts, useful for selection in evolutionary
+//!   algorithms.
+//! - **Extracting the best solutions**: Use [`pareto_front_indices`] to get
+//!   only the non-dominated set.
+//! - **Diversity measurement**: Use [`crowding_distance`] to quantify how
+//!   spread out solutions are within a front, which helps maintain diversity.
+//!
+//! Internally, this module also provides the fast non-dominated sorting
+//! algorithm (Deb et al., 2002) used by
+//! [`MultiObjectiveStudy::pareto_front()`](crate::multi_objective::MultiObjectiveStudy::pareto_front)
+//! and [`Nsga2Sampler`](crate::sampler::Nsga2Sampler).
+//!
+//! # Example
+//!
+//! ```
+//! use optimizer::Direction;
+//! use optimizer::pareto::{
+//!     crowding_distance, hypervolume, non_dominated_sort, pareto_front_indices,
+//! };
+//!
+//! let solutions = vec![
+//!     vec![1.0, 5.0], // Pareto-optimal
+//!     vec![5.0, 1.0], // Pareto-optimal
+//!     vec![3.0, 3.0], // Pareto-optimal
+//!     vec![4.0, 4.0], // Dominated by (3, 3)
+//! ];
+//! let dirs = [Direction::Minimize, Direction::Minimize];
+//!
+//! // Non-dominated sorting: front 0 has indices {0, 1, 2}
+//! let fronts = non_dominated_sort(&solutions, &dirs);
+//! assert_eq!(fronts.len(), 2);
+//!
+//! // Pareto front indices (shortcut for fronts[0])
+//! let mut front = pareto_front_indices(&solutions, &dirs);
+//! front.sort();
+//! assert_eq!(front, vec![0, 1, 2]);
+//!
+//! // Hypervolume with reference point (6, 6)
+//! let front_values: Vec<_> = front.iter().map(|&i| solutions[i].clone()).collect();
+//! let hv = hypervolume(&front_values, &[6.0, 6.0], &dirs);
+//! assert!(hv > 0.0);
+//!
+//! // Crowding distance for diversity analysis
+//! let cd = crowding_distance(&front_values, &dirs);
+//! assert!(cd[0].is_infinite()); // boundary solution
+//! ```
+
+use crate::types::Direction;
+
+/// Returns `true` if solution `a` Pareto-dominates solution `b`.
+///
+/// A solution dominates another if it is at least as good in all objectives
+/// and strictly better in at least one, respecting the given directions.
+#[allow(clippy::module_name_repetitions)]
+pub(crate) fn dominates(a: &[f64], b: &[f64], directions: &[Direction]) -> bool {
+    debug_assert_eq!(a.len(), b.len());
+    debug_assert_eq!(a.len(), directions.len());
+
+    let mut strictly_better = false;
+    for ((&av, &bv), dir) in a.iter().zip(b.iter()).zip(directions.iter()) {
+        let better = match dir {
+            Direction::Minimize => av < bv,
+            Direction::Maximize => av > bv,
+        };
+        let worse = match dir {
+            Direction::Minimize => av > bv,
+            Direction::Maximize => av < bv,
+        };
+        if worse {
+            return false;
+        }
+        if better {
+            strictly_better = true;
+        }
+    }
+    strictly_better
+}
+
+/// Constrained dominance: feasible beats infeasible, among infeasible
+/// prefer lower total constraint violation, among feasible use Pareto dominance.
+pub(crate) fn constrained_dominates(
+    a_values: &[f64],
+    b_values: &[f64],
+    a_constraints: &[f64],
+    b_constraints: &[f64],
+    directions: &[Direction],
+) -> bool {
+    let a_feasible = a_constraints.iter().all(|&c| c <= 0.0);
+    let b_feasible = b_constraints.iter().all(|&c| c <= 0.0);
+
+    match (a_feasible, b_feasible) {
+        (true, false) => true,
+        (false, true) => false,
+        (false, false) => {
+            let a_violation: f64 = a_constraints.iter().map(|c| c.max(0.0)).sum();
+            let b_violation: f64 = b_constraints.iter().map(|c| c.max(0.0)).sum();
+            a_violation < b_violation
+        }
+        (true, true) => dominates(a_values, b_values, directions),
+    }
+}
+
+/// Fast non-dominated sorting (Deb et al., 2002).
+///
+/// Returns `Vec<Vec<usize>>` where `fronts[0]` is the Pareto front,
+/// each inner vec contains indices into `values`.
+///
+/// Complexity: O(M * N^2) where M = objectives, N = solutions.
+#[allow(clippy::cast_possible_truncation)]
+pub(crate) fn fast_non_dominated_sort(
+    values: &[Vec<f64>],
+    directions: &[Direction],
+) -> Vec<Vec<usize>> {
+    fast_non_dominated_sort_constrained(values, directions, &[])
+}
+
+/// Fast non-dominated sorting with constraint support.
+///
+/// `constraints` is either empty (no constraints) or has the same length
+/// as `values`, where each entry is the constraint vector for that solution.
+#[allow(clippy::cast_possible_truncation)]
+pub(crate) fn fast_non_dominated_sort_constrained(
+    values: &[Vec<f64>],
+    directions: &[Direction],
+    constraints: &[Vec<f64>],
+) -> Vec<Vec<usize>> {
+    let n = values.len();
+    if n == 0 {
+        return Vec::new();
+    }
+
+    let has_constraints = !constraints.is_empty();
+    let empty_constraints: Vec<f64> = Vec::new();
+
+    // S_p: set of solutions dominated by p
+    let mut dominated_by: Vec<Vec<usize>> = vec![Vec::new(); n];
+    // n_p: domination count for p
+    let mut domination_count: Vec<usize> = vec![0; n];
+
+    for i in 0..n {
+        for j in (i + 1)..n {
+            let (a_c, b_c) = if has_constraints {
+                (&constraints[i], &constraints[j])
+            } else {
+                (&empty_constraints, &empty_constraints)
+            };
+
+            let i_dom_j = if has_constraints {
+                constrained_dominates(&values[i], &values[j], a_c, b_c, directions)
+            } else {
+                dominates(&values[i], &values[j], directions)
+            };
+            let j_dom_i = if has_constraints {
+                constrained_dominates(&values[j], &values[i], b_c, a_c, directions)
+            } else {
+                dominates(&values[j], &values[i], directions)
+            };
+
+            if i_dom_j {
+                dominated_by[i].push(j);
+                domination_count[j] += 1;
+            } else if j_dom_i {
+                dominated_by[j].push(i);
+                domination_count[i] += 1;
+            }
+        }
+    }
+
+    let mut fronts: Vec<Vec<usize>> = Vec::new();
+    let mut current_front: Vec<usize> = (0..n).filter(|&i| domination_count[i] == 0).collect();
+
+    while !current_front.is_empty() {
+        let mut next_front: Vec<usize> = Vec::new();
+        for &p in &current_front {
+            for &q in &dominated_by[p] {
+                domination_count[q] -= 1;
+                if domination_count[q] == 0 {
+                    next_front.push(q);
+                }
+            }
+        }
+        fronts.push(current_front);
+        current_front = next_front;
+    }
+
+    fronts
+}
+
+/// Crowding distance for one front (index-based, internal API).
+///
+/// Boundary solutions get `f64::INFINITY`. Returns one distance value per
+/// solution in the front, in the same order as `front_indices`.
+#[allow(clippy::cast_precision_loss)]
+pub(crate) fn crowding_distance_indexed(front_indices: &[usize], values: &[Vec<f64>]) -> Vec<f64> {
+    let n = front_indices.len();
+    if n <= 2 {
+        return vec![f64::INFINITY; n];
+    }
+
+    let m = values[front_indices[0]].len(); // number of objectives
+    let mut distances = vec![0.0_f64; n];
+
+    // Helper to look up objective value for a front member.
+    let val = |front_pos: usize, obj: usize| -> f64 { values[front_indices[front_pos]][obj] };
+
+    for obj in 0..m {
+        // Sort front positions by this objective
+        let mut sorted: Vec<usize> = (0..n).collect();
+        sorted.sort_by(|&a, &b| {
+            val(a, obj)
+                .partial_cmp(&val(b, obj))
+                .unwrap_or(core::cmp::Ordering::Equal)
+        });
+
+        // Boundary solutions get infinity
+        distances[sorted[0]] = f64::INFINITY;
+        distances[sorted[n - 1]] = f64::INFINITY;
+
+        let range = val(sorted[n - 1], obj) - val(sorted[0], obj);
+        if range > 0.0 {
+            for i in 1..(n - 1) {
+                distances[sorted[i]] += (val(sorted[i + 1], obj) - val(sorted[i - 1], obj)) / range;
+            }
+        }
+    }
+
+    distances
+}
+
+// ---------------------------------------------------------------------------
+// Public API
+// ---------------------------------------------------------------------------
+
+/// Compute the hypervolume indicator of a Pareto front.
+///
+/// The hypervolume is the volume of the objective space dominated by
+/// the Pareto front and bounded by a reference point. A **higher**
+/// hypervolume indicates a better front (closer to the ideal and more
+/// spread out).
+///
+/// Each entry in `front` is one solution's objective values.
+/// `reference_point` should be worse than all front members in every
+/// objective (e.g., the worst acceptable values). Solutions that do
+/// not strictly dominate the reference point are ignored.
+///
+/// Uses recursive slicing for dimensions > 1. Complexity grows with
+/// the number of objectives and front size.
+///
+/// # Panics
+///
+/// Panics (in debug) if dimensions of `front`, `reference_point`, and
+/// `directions` are inconsistent.
+#[must_use]
+#[allow(clippy::cast_precision_loss)]
+pub fn hypervolume(front: &[Vec<f64>], reference_point: &[f64], directions: &[Direction]) -> f64 {
+    if front.is_empty() {
+        return 0.0;
+    }
+    let d = reference_point.len();
+    debug_assert!(front.iter().all(|p| p.len() == d));
+    debug_assert_eq!(d, directions.len());
+
+    // Normalize to minimize-space (negate maximized objectives).
+    let normalized: Vec<Vec<f64>> = front
+        .iter()
+        .map(|p| {
+            p.iter()
+                .zip(directions)
+                .map(|(&v, dir)| match dir {
+                    Direction::Minimize => v,
+                    Direction::Maximize => -v,
+                })
+                .collect()
+        })
+        .collect();
+
+    let ref_norm: Vec<f64> = reference_point
+        .iter()
+        .zip(directions)
+        .map(|(&v, dir)| match dir {
+            Direction::Minimize => v,
+            Direction::Maximize => -v,
+        })
+        .collect();
+
+    // Keep only points strictly dominated by the reference point.
+    let filtered: Vec<Vec<f64>> = normalized
+        .into_iter()
+        .filter(|p| p.iter().zip(&ref_norm).all(|(&pv, &rv)| pv < rv))
+        .collect();
+
+    if filtered.is_empty() {
+        return 0.0;
+    }
+
+    hv_recursive(&filtered, &ref_norm)
+}
+
+/// Recursive hypervolume via slicing on the last objective.
+///
+/// All points are in minimize-space and dominated by `reference`.
+#[allow(clippy::cast_precision_loss)]
+fn hv_recursive(points: &[Vec<f64>], reference: &[f64]) -> f64 {
+    let d = reference.len();
+
+    // Base case: 1-D hypervolume is just the gap from the best point to ref.
+    if d == 1 {
+        let min_val = points.iter().map(|p| p[0]).fold(f64::INFINITY, f64::min);
+        return (reference[0] - min_val).max(0.0);
+    }
+
+    // Single point: hypervolume is the product of gaps.
+    if points.len() == 1 {
+        return points[0]
+            .iter()
+            .zip(reference)
+            .map(|(&p, &r)| (r - p).max(0.0))
+            .product();
+    }
+
+    // Sort by last objective ascending.
+    let mut sorted: Vec<&Vec<f64>> = points.iter().collect();
+    sorted.sort_by(|a, b| {
+        a[d - 1]
+            .partial_cmp(&b[d - 1])
+            .unwrap_or(core::cmp::Ordering::Equal)
+    });
+
+    let sub_ref: Vec<f64> = reference[..d - 1].to_vec();
+    let mut result = 0.0;
+
+    for i in 0..sorted.len() {
+        let height = if i + 1 < sorted.len() {
+            sorted[i + 1][d - 1] - sorted[i][d - 1]
+        } else {
+            reference[d - 1] - sorted[i][d - 1]
+        };
+
+        if height <= 0.0 {
+            continue;
+        }
+
+        // Project points[0..=i] onto the first d-1 dimensions and
+        // keep only the non-dominated subset.
+        let projected: Vec<Vec<f64>> = sorted[..=i].iter().map(|p| p[..d - 1].to_vec()).collect();
+        let non_dom = non_dominated_minimize(&projected);
+
+        if !non_dom.is_empty() {
+            result += height * hv_recursive(&non_dom, &sub_ref);
+        }
+    }
+
+    result
+}
+
+/// Return the non-dominated subset of `points` in minimize-space.
+fn non_dominated_minimize(points: &[Vec<f64>]) -> Vec<Vec<f64>> {
+    let mut result = Vec::new();
+    'outer: for (i, p) in points.iter().enumerate() {
+        for (j, q) in points.iter().enumerate() {
+            if i == j {
+                continue;
+            }
+            // Check if q dominates p (all <=, at least one <).
+            let mut all_leq = true;
+            let mut any_lt = false;
+            for (&qv, &pv) in q.iter().zip(p.iter()) {
+                if qv > pv {
+                    all_leq = false;
+                    break;
+                }
+                if qv < pv {
+                    any_lt = true;
+                }
+            }
+            if all_leq && any_lt {
+                continue 'outer;
+            }
+        }
+        result.push(p.clone());
+    }
+    result
+}
+
+/// Compute non-dominated sorting of a set of solutions.
+///
+/// Return a vec of fronts, where `fronts[0]` is the Pareto front
+/// (non-dominated solutions), `fronts[1]` is the next-best front
+/// (dominated only by front 0), and so on. Each inner vec contains
+/// indices into the original `solutions` slice.
+///
+/// Use the fast non-dominated sorting algorithm from
+/// Deb et al. (2002) with O(M × N²) complexity, where M is the
+/// number of objectives and N is the number of solutions.
+#[must_use]
+pub fn non_dominated_sort(solutions: &[Vec<f64>], directions: &[Direction]) -> Vec<Vec<usize>> {
+    fast_non_dominated_sort(solutions, directions)
+}
+
+/// Filter solutions to return only non-dominated (Pareto-optimal) indices.
+///
+/// Equivalent to `non_dominated_sort(solutions, directions)[0]` but
+/// communicates the intent more clearly. Use this when you only need
+/// the Pareto front and not the full ranking.
+#[must_use]
+pub fn pareto_front_indices(solutions: &[Vec<f64>], directions: &[Direction]) -> Vec<usize> {
+    let fronts = fast_non_dominated_sort(solutions, directions);
+    fronts.into_iter().next().unwrap_or_default()
+}
+
+/// Compute crowding distance for diversity measurement.
+///
+/// Return one distance value per solution in `front` (same order).
+/// Boundary solutions (best/worst in any objective) receive
+/// [`f64::INFINITY`]. Interior solutions get a finite positive value
+/// proportional to the gap between their neighbors in each objective.
+///
+/// Crowding distance is used by NSGA-II to prefer well-spread
+/// solutions when two solutions are in the same front.
+///
+/// `directions` is accepted for API consistency but does not affect
+/// the result, since crowding distance measures spacing regardless of
+/// optimization direction.
+#[must_use]
+#[allow(clippy::cast_precision_loss, clippy::needless_range_loop)]
+pub fn crowding_distance(front: &[Vec<f64>], _directions: &[Direction]) -> Vec<f64> {
+    let n = front.len();
+    if n <= 2 {
+        return vec![f64::INFINITY; n];
+    }
+
+    let m = front[0].len();
+    let mut distances = vec![0.0_f64; n];
+
+    for obj in 0..m {
+        let mut sorted: Vec<usize> = (0..n).collect();
+        sorted.sort_by(|&a, &b| {
+            front[a][obj]
+                .partial_cmp(&front[b][obj])
+                .unwrap_or(core::cmp::Ordering::Equal)
+        });
+
+        distances[sorted[0]] = f64::INFINITY;
+        distances[sorted[n - 1]] = f64::INFINITY;
+
+        let range = front[sorted[n - 1]][obj] - front[sorted[0]][obj];
+        if range > 0.0 {
+            for i in 1..(n - 1) {
+                distances[sorted[i]] +=
+                    (front[sorted[i + 1]][obj] - front[sorted[i - 1]][obj]) / range;
+            }
+        }
+    }
+
+    distances
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_dominates_basic() {
+        let dirs = [Direction::Minimize, Direction::Minimize];
+        assert!(dominates(&[1.0, 1.0], &[2.0, 2.0], &dirs));
+        assert!(!dominates(&[2.0, 2.0], &[1.0, 1.0], &dirs));
+        // Equal does not dominate
+        assert!(!dominates(&[1.0, 1.0], &[1.0, 1.0], &dirs));
+    }
+
+    #[test]
+    fn test_dominates_incomparable() {
+        let dirs = [Direction::Minimize, Direction::Minimize];
+        assert!(!dominates(&[1.0, 3.0], &[3.0, 1.0], &dirs));
+        assert!(!dominates(&[3.0, 1.0], &[1.0, 3.0], &dirs));
+    }
+
+    #[test]
+    fn test_dominates_maximize() {
+        let dirs = [Direction::Maximize, Direction::Minimize];
+        // a = (5, 1) vs b = (3, 2): a is better in both
+        assert!(dominates(&[5.0, 1.0], &[3.0, 2.0], &dirs));
+        assert!(!dominates(&[3.0, 2.0], &[5.0, 1.0], &dirs));
+    }
+
+    #[test]
+    fn test_nds_known() {
+        let values = vec![
+            vec![1.0, 5.0], // front 0
+            vec![5.0, 1.0], // front 0
+            vec![3.0, 3.0], // front 0 (non-dominated)
+            vec![4.0, 4.0], // front 1 (dominated by #2)
+            vec![6.0, 6.0], // front 2
+        ];
+        let dirs = [Direction::Minimize, Direction::Minimize];
+        let fronts = fast_non_dominated_sort(&values, &dirs);
+
+        assert_eq!(fronts.len(), 3);
+        let mut f0 = fronts[0].clone();
+        f0.sort_unstable();
+        assert_eq!(f0, vec![0, 1, 2]);
+        assert_eq!(fronts[1], vec![3]);
+        assert_eq!(fronts[2], vec![4]);
+    }
+
+    #[test]
+    fn test_crowding_indexed_boundaries() {
+        let values = vec![vec![1.0, 5.0], vec![3.0, 3.0], vec![5.0, 1.0]];
+        let front = vec![0, 1, 2];
+        let cd = crowding_distance_indexed(&front, &values);
+        assert!(cd[0].is_infinite());
+        assert!(cd[2].is_infinite());
+        assert!(cd[1].is_finite());
+        assert!(cd[1] > 0.0);
+    }
+
+    // ---- Public API tests ----
+
+    #[test]
+    fn test_hypervolume_2d_minimize() {
+        // Front: (1,3), (2,2), (3,1) with ref (4,4) — all minimize
+        let front = vec![vec![1.0, 3.0], vec![2.0, 2.0], vec![3.0, 1.0]];
+        let dirs = [Direction::Minimize, Direction::Minimize];
+        let hv = hypervolume(&front, &[4.0, 4.0], &dirs);
+        // Strip 1: x=[1,2), h=4-3=1 → area=1
+        // Strip 2: x=[2,3), h=4-2=2 → area=2
+        // Strip 3: x=[3,4], h=4-1=3 → area=3
+        // Total = 6
+        assert!((hv - 6.0).abs() < 1e-10);
+    }
+
+    #[test]
+    fn test_hypervolume_2d_maximize() {
+        // Front: (3,1), (2,2), (1,3) with ref (0,0) — all maximize
+        let front = vec![vec![3.0, 1.0], vec![2.0, 2.0], vec![1.0, 3.0]];
+        let dirs = [Direction::Maximize, Direction::Maximize];
+        let hv = hypervolume(&front, &[0.0, 0.0], &dirs);
+        // In negate-space: points become (-3,-1),(-2,-2),(-1,-3), ref=(0,0)
+        // Same geometry as minimize test above → area = 6
+        assert!((hv - 6.0).abs() < 1e-10);
+    }
+
+    #[test]
+    fn test_hypervolume_single_point() {
+        let front = vec![vec![1.0, 1.0]];
+        let dirs = [Direction::Minimize, Direction::Minimize];
+        let hv = hypervolume(&front, &[3.0, 3.0], &dirs);
+        // Rectangle: (3-1) * (3-1) = 4
+        assert!((hv - 4.0).abs() < 1e-10);
+    }
+
+    #[test]
+    fn test_hypervolume_empty_front() {
+        let front: Vec<Vec<f64>> = vec![];
+        let dirs = [Direction::Minimize];
+        assert!(hypervolume(&front, &[1.0], &dirs).abs() < f64::EPSILON);
+    }
+
+    #[test]
+    fn test_hypervolume_point_at_ref() {
+        // Point not strictly better than ref → contributes nothing
+        let front = vec![vec![5.0, 5.0]];
+        let dirs = [Direction::Minimize, Direction::Minimize];
+        let hv = hypervolume(&front, &[5.0, 5.0], &dirs);
+        assert!(hv.abs() < f64::EPSILON);
+    }
+
+    #[test]
+    fn test_hypervolume_3d() {
+        // Single point in 3D: (1,1,1) with ref (2,2,2)
+        let front = vec![vec![1.0, 1.0, 1.0]];
+        let dirs = [
+            Direction::Minimize,
+            Direction::Minimize,
+            Direction::Minimize,
+        ];
+        let hv = hypervolume(&front, &[2.0, 2.0, 2.0], &dirs);
+        assert!((hv - 1.0).abs() < 1e-10);
+    }
+
+    #[test]
+    fn test_non_dominated_sort_public() {
+        let values = vec![
+            vec![1.0, 5.0],
+            vec![5.0, 1.0],
+            vec![3.0, 3.0],
+            vec![4.0, 4.0],
+        ];
+        let dirs = [Direction::Minimize, Direction::Minimize];
+        let fronts = non_dominated_sort(&values, &dirs);
+        assert_eq!(fronts.len(), 2);
+        let mut f0 = fronts[0].clone();
+        f0.sort_unstable();
+        assert_eq!(f0, vec![0, 1, 2]);
+        assert_eq!(fronts[1], vec![3]);
+    }
+
+    #[test]
+    fn test_pareto_front_indices_basic() {
+        let values = vec![
+            vec![1.0, 5.0],
+            vec![5.0, 1.0],
+            vec![3.0, 3.0],
+            vec![4.0, 4.0],
+        ];
+        let dirs = [Direction::Minimize, Direction::Minimize];
+        let mut idx = pareto_front_indices(&values, &dirs);
+        idx.sort_unstable();
+        assert_eq!(idx, vec![0, 1, 2]);
+    }
+
+    #[test]
+    fn test_pareto_front_indices_empty() {
+        let values: Vec<Vec<f64>> = vec![];
+        let dirs = [Direction::Minimize];
+        assert!(pareto_front_indices(&values, &dirs).is_empty());
+    }
+
+    #[test]
+    fn test_crowding_distance_public() {
+        let front = vec![vec![1.0, 5.0], vec![3.0, 3.0], vec![5.0, 1.0]];
+        let dirs = [Direction::Minimize, Direction::Minimize];
+        let cd = crowding_distance(&front, &dirs);
+        assert!(cd[0].is_infinite());
+        assert!(cd[2].is_infinite());
+        assert!(cd[1].is_finite());
+        assert!(cd[1] > 0.0);
+    }
+
+    #[test]
+    fn test_crowding_distance_single_point() {
+        let front = vec![vec![2.0, 3.0]];
+        let dirs = [Direction::Minimize, Direction::Minimize];
+        let cd = crowding_distance(&front, &dirs);
+        assert_eq!(cd.len(), 1);
+        assert!(cd[0].is_infinite());
+    }
+}

src/pruner/hyperband.rs

@@ -1,3 +1,47 @@
+//! `HyperBand` pruner — adaptive budget scheduling with multiple SHA brackets.
+//!
+//! `HyperBand` addresses the main weakness of
+//! [`SuccessiveHalvingPruner`](super::SuccessiveHalvingPruner): sensitivity to
+//! the `min_resource` setting. It runs multiple Successive Halving brackets in
+//! parallel, each with a different trade-off between the number of trials and
+//! the starting budget:
+//!
+//! - **Bracket 0**: many trials, small starting budget (aggressive early pruning)
+//! - **Bracket `s_max`**: few trials, full budget (no pruning)
+//!
+//! Trials are assigned to brackets in round-robin order. This ensures that
+//! the overall search is robust regardless of how informative early steps are.
+//!
+//! # When to use
+//!
+//! - When you don't know how many epochs/steps are needed before performance
+//!   becomes predictive
+//! - As a drop-in upgrade over [`SuccessiveHalvingPruner`](super::SuccessiveHalvingPruner)
+//!   when you can afford more total trials
+//! - For large-scale hyperparameter searches where compute savings matter most
+//!
+//! # Configuration
+//!
+//! | Option | Default | Description |
+//! |--------|---------|-------------|
+//! | `min_resource` | 1 | Smallest budget for the most aggressive bracket |
+//! | `max_resource` | 81 | Full budget (last rung in every bracket) |
+//! | `reduction_factor` | 3 | At each rung, keep top 1/η trials |
+//! | `direction` | `Minimize` | Optimization direction |
+//!
+//! # Example
+//!
+//! ```
+//! use optimizer::Direction;
+//! use optimizer::pruner::HyperbandPruner;
+//!
+//! let pruner = HyperbandPruner::new()
+//!     .min_resource(1)
+//!     .max_resource(81)
+//!     .reduction_factor(3)
+//!     .direction(Direction::Minimize);
+//! ```
+
 use core::sync::atomic::{AtomicU64, Ordering};
 use std::collections::HashMap;
 use std::sync::Mutex;
@@ -6,7 +50,7 @@ use super::Pruner;
 use crate::sampler::CompletedTrial;
 use crate::types::{Direction, TrialState};
 
-/// Hyperband pruner that manages multiple Successive Halving brackets.
+/// `HyperBand` pruner that manages multiple Successive Halving brackets.
 ///
 /// Hyperband addresses SHA's sensitivity to the `min_resource` choice by
 /// running multiple brackets, each with a different tradeoff between the
@@ -273,7 +317,7 @@ mod tests {
 
         CompletedTrial::with_intermediate_values(
             id,
-            HashMap::<ParamId, crate::ParamValue>::new(),
+            HashMap::<ParamId, crate::parameter::ParamValue>::new(),
             HashMap::new(),
             HashMap::new(),
             0.0,

src/pruner/median.rs

@@ -1,3 +1,39 @@
+//! Median pruner — the recommended default pruner for most use cases.
+//!
+//! At each step, the current trial's intermediate value is compared against
+//! the median of all completed trials' values at the same step. Trials
+//! performing worse than the median are pruned.
+//!
+//! This is a convenience wrapper around [`PercentilePruner`](super::PercentilePruner)
+//! with a fixed percentile of 50%.
+//!
+//! # When to use
+//!
+//! - **Default choice** for any iterative objective (e.g., neural network training)
+//! - Works well when intermediate values are a reasonable proxy for final performance
+//! - Prunes roughly half of unpromising trials, giving a good speed/accuracy balance
+//!
+//! If your intermediate values are noisy, consider [`WilcoxonPruner`](super::WilcoxonPruner)
+//! or wrapping this pruner in a [`PatientPruner`](super::PatientPruner).
+//!
+//! # Configuration
+//!
+//! | Option | Default | Description |
+//! |--------|---------|-------------|
+//! | `n_warmup_steps` | 0 | Skip pruning in the first N steps |
+//! | `n_min_trials` | 1 | Require at least N completed trials before pruning |
+//!
+//! # Example
+//!
+//! ```
+//! use optimizer::Direction;
+//! use optimizer::pruner::MedianPruner;
+//!
+//! let pruner = MedianPruner::new(Direction::Minimize)
+//!     .n_warmup_steps(5)
+//!     .n_min_trials(3);
+//! ```
+
 use super::Pruner;
 use super::percentile::compute_percentile;
 use crate::sampler::CompletedTrial;
@@ -53,8 +89,13 @@ impl MedianPruner {
     }
 
     /// Set the minimum number of completed trials required before pruning.
+    ///
+    /// # Panics
+    ///
+    /// Panics if `n` is 0.
     #[must_use]
     pub fn n_min_trials(mut self, n: usize) -> Self {
+        assert!(n >= 1, "n_min_trials must be >= 1, got {n}");
         self.n_min_trials = n;
         self
     }

src/pruner/mod.rs

@@ -3,6 +3,99 @@
 //! Pruners decide whether to stop (prune) a trial early based on its
 //! intermediate values compared to other trials. This is useful for
 //! discarding unpromising trials before they complete, saving compute.
+//!
+//! # How pruning works
+//!
+//! During optimization, each trial reports intermediate values at discrete
+//! steps (e.g., validation loss after each training epoch). A pruner inspects
+//! these values and compares them against completed trials to decide whether
+//! the current trial should be stopped early.
+//!
+//! The typical flow is:
+//!
+//! 1. Call [`Trial::report`](crate::Trial::report) to record an intermediate value.
+//! 2. Call [`Trial::should_prune`](crate::Trial::should_prune) to check the pruner's decision.
+//! 3. If the pruner says prune, return [`TrialPruned`](crate::TrialPruned) from the objective.
+//!
+//! # Available pruners
+//!
+//! | Pruner | Algorithm | Best for |
+//! |--------|-----------|----------|
+//! | [`MedianPruner`] | Prune below median at each step | General-purpose default |
+//! | [`PercentilePruner`] | Prune below configurable percentile | Tunable aggressiveness |
+//! | [`ThresholdPruner`] | Prune outside fixed bounds | Known divergence limits |
+//! | [`PatientPruner`] | Require N consecutive prune signals | Noisy intermediate values |
+//! | [`SuccessiveHalvingPruner`] | Keep top 1/η fraction at each rung | Budget-aware pruning |
+//! | [`HyperbandPruner`] | Multiple SHA brackets with different budgets | Robust to budget choice |
+//! | [`WilcoxonPruner`] | Statistical signed-rank test vs. best trial | Rigorous noisy pruning |
+//! | [`NopPruner`] | Never prune | Disabling pruning explicitly |
+//!
+//! # When to use pruning
+//!
+//! Pruning is most beneficial when:
+//!
+//! - The objective function has a natural notion of "steps" (e.g., training epochs)
+//! - Early steps are informative about final performance
+//! - Trials are expensive enough that stopping bad ones early saves significant time
+//!
+//! Start with [`MedianPruner`] for most use cases. Switch to [`WilcoxonPruner`]
+//! if your intermediate values are noisy, or to [`HyperbandPruner`] if you want
+//! automatic budget scheduling.
+//!
+//! # Stateful vs stateless pruners
+//!
+//! **Stateless** pruners make their decision purely from the arguments passed
+//! to [`Pruner::should_prune`] — they hold no mutable per-trial state.
+//! [`MedianPruner`], [`PercentilePruner`], [`ThresholdPruner`],
+//! [`WilcoxonPruner`], and [`NopPruner`] are all stateless.
+//!
+//! **Stateful** pruners track information across calls. [`PatientPruner`]
+//! uses `Mutex<HashMap<u64, u64>>` to count consecutive prune signals per
+//! trial. [`HyperbandPruner`] uses `Mutex` and `AtomicU64` for bracket
+//! assignment state. When writing a stateful pruner, wrap mutable state in a
+//! `Mutex` and key it by `trial_id` to keep trials independent.
+//!
+//! # Cold start and warmup
+//!
+//! Two builder parameters control when pruning begins:
+//!
+//! - **`n_warmup_steps`** — skip pruning before step N *within a trial*,
+//!   giving the objective time to stabilize.
+//! - **`n_min_trials`** — require N completed trials before pruning any trial,
+//!   ensuring a meaningful comparison baseline.
+//!
+//! See [`MedianPruner`] for the canonical implementation of both parameters.
+//! Custom pruners should expose similar knobs when applicable.
+//!
+//! # Composing pruners
+//!
+//! [`PatientPruner`] demonstrates the decorator pattern: it wraps any
+//! `Box<dyn Pruner>` and adds patience logic on top. Custom pruners can use
+//! the same pattern to layer multiple pruning conditions — for example,
+//! combining a statistical test with a hard threshold.
+//!
+//! # Thread safety
+//!
+//! The [`Pruner`] trait requires `Send + Sync`.
+//! [`Study`](crate::Study) stores the pruner as `Arc<dyn Pruner>`, so
+//! multiple threads may call [`Pruner::should_prune`] concurrently.
+//!
+//! - **Stateless pruners** satisfy `Send + Sync` automatically.
+//! - **Stateful pruners** should use `std::sync::Mutex` or
+//!   `parking_lot::Mutex` to protect mutable state, keyed by `trial_id`.
+//!
+//! # Testing custom pruners
+//!
+//! Recommended test categories:
+//!
+//! 1. **Never-prune baseline** — empty history and early steps should not
+//!    prune.
+//! 2. **Known-prune scenario** — a clearly worse trial should be pruned.
+//! 3. **Known-keep scenario** — a well-performing trial should survive.
+//! 4. **Warmup respected** — pruning must be suppressed during warmup steps
+//!    and while the minimum trial count has not been reached.
+//! 5. **Per-trial independence** — stateful pruners must not leak state
+//!    between different `trial_id` values.
 
 mod hyperband;
 mod median;
@@ -55,6 +148,54 @@ use crate::sampler::CompletedTrial;
 ///     }
 /// }
 /// ```
+///
+/// A stateful pruner that tracks per-trial state with a `Mutex`:
+///
+/// ```
+/// use std::collections::HashMap;
+/// use std::sync::Mutex;
+/// use optimizer::pruner::Pruner;
+/// use optimizer::sampler::CompletedTrial;
+///
+/// /// Prune after the value worsens for `max_stale` consecutive steps.
+/// struct StalePruner {
+///     max_stale: u64,
+///     // Per-trial: (previous_value, consecutive_stale_count)
+///     state: Mutex<HashMap<u64, (f64, u64)>>,
+/// }
+///
+/// impl StalePruner {
+///     fn new(max_stale: u64) -> Self {
+///         Self { max_stale, state: Mutex::new(HashMap::new()) }
+///     }
+/// }
+///
+/// impl Pruner for StalePruner {
+///     fn should_prune(
+///         &self,
+///         trial_id: u64,
+///         _step: u64,
+///         intermediate_values: &[(u64, f64)],
+///         _completed_trials: &[CompletedTrial],
+///     ) -> bool {
+///         let Some(&(_, current)) = intermediate_values.last() else {
+///             return false;
+///         };
+///         let mut state = self.state.lock().unwrap();
+///         let entry = state.entry(trial_id).or_insert((current, 0));
+///         if current >= entry.0 {
+///             entry.1 += 1;
+///         } else {
+///             entry.1 = 0;
+///         }
+///         entry.0 = current;
+///         entry.1 >= self.max_stale
+///     }
+/// }
+/// ```
+///
+/// See the [module-level documentation](self) for a comprehensive guide
+/// covering warmup, composition, thread safety, and testing.
 pub trait Pruner: Send + Sync {
     /// Decide whether to prune a trial at the given step.
     ///

src/pruner/nop.rs

@@ -1,3 +1,15 @@
+//! No-op pruner — never prune any trial.
+//!
+//! This is the default pruner used when no pruner is configured on a
+//! [`Study`](crate::Study). It unconditionally returns `false` for every
+//! pruning decision, allowing all trials to run to completion.
+//!
+//! # When to use
+//!
+//! - When you want to explicitly disable pruning
+//! - As a baseline to compare against other pruners
+//! - Already used by default — you rarely need to configure this manually
+
 use super::Pruner;
 use crate::sampler::CompletedTrial;
 

src/pruner/patient.rs

@@ -1,3 +1,34 @@
+//! Patient pruner — require consecutive prune signals before actually pruning.
+//!
+//! Wraps any other pruner and adds a patience window: the inner pruner
+//! must recommend pruning for `patience` consecutive steps before the
+//! trial is actually pruned. This prevents premature pruning when
+//! intermediate values are noisy and a single bad step doesn't indicate
+//! a truly bad trial.
+//!
+//! # When to use
+//!
+//! - When your intermediate values have high variance (e.g., mini-batch loss)
+//! - When the inner pruner is too aggressive on its own
+//! - To add robustness to any statistical pruner without changing its threshold
+//!
+//! # Configuration
+//!
+//! | Option | Default | Description |
+//! |--------|---------|-------------|
+//! | `inner` | *(required)* | The underlying pruner to wrap |
+//! | `patience` | *(required)* | Number of consecutive prune signals required |
+//!
+//! # Example
+//!
+//! ```
+//! use optimizer::pruner::{PatientPruner, ThresholdPruner};
+//!
+//! // Only prune after the threshold is exceeded 3 times in a row
+//! let inner = ThresholdPruner::new().upper(100.0);
+//! let pruner = PatientPruner::new(inner, 3);
+//! ```
+
 use std::collections::HashMap;
 use std::sync::Mutex;
 

src/pruner/percentile.rs

@@ -1,3 +1,37 @@
+//! Percentile pruner — prune trials outside the top N% at each step.
+//!
+//! A generalization of [`MedianPruner`](super::MedianPruner) that lets you
+//! control how aggressively to prune. At each step, the current trial's
+//! intermediate value is compared against the given percentile of all
+//! completed trials' values at the same step.
+//!
+//! # When to use
+//!
+//! - When you want finer control over pruning aggressiveness than median pruning
+//! - Lower percentiles (e.g., 25%) are more aggressive — only keep the best quarter
+//! - Higher percentiles (e.g., 75%) are more lenient — keep the top three quarters
+//! - Percentile 50% is equivalent to [`MedianPruner`](super::MedianPruner)
+//!
+//! # Configuration
+//!
+//! | Option | Default | Description |
+//! |--------|---------|-------------|
+//! | `percentile` | *(required)* | Keep trials in the top N% — range `(0, 100)` |
+//! | `n_warmup_steps` | 0 | Skip pruning in the first N steps |
+//! | `n_min_trials` | 1 | Require at least N completed trials before pruning |
+//!
+//! # Example
+//!
+//! ```
+//! use optimizer::Direction;
+//! use optimizer::pruner::PercentilePruner;
+//!
+//! // Keep only the top 25% of trials (aggressive pruning)
+//! let pruner = PercentilePruner::new(25.0, Direction::Minimize)
+//!     .n_warmup_steps(5)
+//!     .n_min_trials(3);
+//! ```
+
 use super::Pruner;
 use crate::sampler::CompletedTrial;
 use crate::types::{Direction, TrialState};
@@ -61,8 +95,13 @@ impl PercentilePruner {
     }
 
     /// Set the minimum number of completed trials required before pruning.
+    ///
+    /// # Panics
+    ///
+    /// Panics if `n` is 0.
     #[must_use]
     pub fn n_min_trials(mut self, n: usize) -> Self {
+        assert!(n >= 1, "n_min_trials must be >= 1, got {n}");
         self.n_min_trials = n;
         self
     }
@@ -123,6 +162,7 @@ impl Pruner for PercentilePruner {
     clippy::cast_sign_loss
 )]
 pub(crate) fn compute_percentile(values: &mut [f64], percentile: f64) -> f64 {
+    assert!(!values.is_empty(), "compute_percentile: empty input");
     values.sort_unstable_by(|a, b| a.partial_cmp(b).unwrap_or(core::cmp::Ordering::Equal));
     let len = values.len();
     if len == 1 {
@@ -199,7 +239,7 @@ mod tests {
 
         CompletedTrial::with_intermediate_values(
             id,
-            HashMap::<ParamId, crate::ParamValue>::new(),
+            HashMap::<ParamId, crate::parameter::ParamValue>::new(),
             HashMap::new(),
             HashMap::new(),
             0.0,

src/pruner/successive_halving.rs

@@ -1,3 +1,53 @@
+//! Successive Halving (SHA) pruner — budget-aware pruning at exponential rungs.
+//!
+//! Trials are evaluated at exponentially-spaced "rungs" (checkpoints). At each
+//! rung, only the top 1/η fraction of trials survive to the next rung. This
+//! is a principled way to allocate compute budget: give many trials a small
+//! budget, then progressively invest more in the best ones.
+//!
+//! For example, with `min_resource=1`, `max_resource=81`, `reduction_factor=3`:
+//!
+//! | Rung | Step | Survivors |
+//! |------|------|-----------|
+//! | 0 | 1 | top 1/3 |
+//! | 1 | 3 | top 1/3 |
+//! | 2 | 9 | top 1/3 |
+//! | 3 | 27 | top 1/3 |
+//! | 4 | 81 | all (full budget) |
+//!
+//! # When to use
+//!
+//! - When your objective has a natural "budget" dimension (epochs, iterations)
+//! - When early performance is a reasonable predictor of final performance
+//! - When you want a principled alternative to median pruning
+//!
+//! If you're unsure about the right `min_resource`, consider
+//! [`HyperbandPruner`](super::HyperbandPruner) which runs multiple brackets
+//! to hedge against that choice.
+//!
+//! # Configuration
+//!
+//! | Option | Default | Description |
+//! |--------|---------|-------------|
+//! | `min_resource` | 1 | Step at which the first rung is placed |
+//! | `max_resource` | 81 | Full budget (final rung, no pruning) |
+//! | `reduction_factor` | 3 | At each rung, keep top 1/η trials |
+//! | `min_early_stopping_rate` | 0 | Skip the first N rungs |
+//! | `direction` | `Minimize` | Optimization direction |
+//!
+//! # Example
+//!
+//! ```
+//! use optimizer::Direction;
+//! use optimizer::pruner::SuccessiveHalvingPruner;
+//!
+//! let pruner = SuccessiveHalvingPruner::new()
+//!     .min_resource(1)
+//!     .max_resource(81)
+//!     .reduction_factor(3)
+//!     .direction(Direction::Minimize);
+//! ```
+
 use super::Pruner;
 use crate::sampler::CompletedTrial;
 use crate::types::{Direction, TrialState};
@@ -235,7 +285,7 @@ mod tests {
 
         CompletedTrial::with_intermediate_values(
             id,
-            HashMap::<ParamId, crate::ParamValue>::new(),
+            HashMap::<ParamId, crate::parameter::ParamValue>::new(),
             HashMap::new(),
             HashMap::new(),
             0.0,

src/pruner/threshold.rs

@@ -1,3 +1,33 @@
+//! Threshold pruner — prune trials whose values fall outside fixed bounds.
+//!
+//! Unlike statistical pruners that compare against other trials, the
+//! threshold pruner uses absolute bounds. Any trial whose latest
+//! intermediate value exceeds the upper bound or falls below the lower
+//! bound is pruned immediately.
+//!
+//! # When to use
+//!
+//! - When you know hard limits for valid intermediate values (e.g., loss should
+//!   never exceed 100.0)
+//! - To catch diverging or NaN-producing trials early
+//! - Often combined with other pruners via [`PatientPruner`](super::PatientPruner)
+//!
+//! # Configuration
+//!
+//! | Option | Default | Description |
+//! |--------|---------|-------------|
+//! | `upper` | `None` | Prune if value exceeds this bound |
+//! | `lower` | `None` | Prune if value falls below this bound |
+//!
+//! # Example
+//!
+//! ```
+//! use optimizer::pruner::ThresholdPruner;
+//!
+//! // Prune if loss exceeds 100.0 or accuracy drops below 0.0
+//! let pruner = ThresholdPruner::new().upper(100.0).lower(0.0);
+//! ```
+
 use super::Pruner;
 use crate::sampler::CompletedTrial;
 

src/pruner/wilcoxon.rs

@@ -1,3 +1,44 @@
+//! Wilcoxon pruner — statistically rigorous pruning for noisy objectives.
+//!
+//! Uses the Wilcoxon signed-rank test to compare the current trial's
+//! intermediate values against the best completed trial at matching steps.
+//! The test accounts for the paired, step-aligned nature of the comparison
+//! and only prunes when the difference is statistically significant.
+//!
+//! This is more principled than [`MedianPruner`](super::MedianPruner) for
+//! noisy objectives because a single bad step won't trigger pruning — the
+//! test considers the full distribution of paired differences.
+//!
+//! # When to use
+//!
+//! - When intermediate values have high variance (e.g., mini-batch loss,
+//!   stochastic reward signals)
+//! - When you want a statistical guarantee that pruned trials are truly worse
+//! - When you have enough steps (at least 6) for a meaningful test
+//!
+//! For less noisy objectives, [`MedianPruner`](super::MedianPruner) is simpler
+//! and often sufficient.
+//!
+//! # Configuration
+//!
+//! | Option | Default | Description |
+//! |--------|---------|-------------|
+//! | `p_value_threshold` | 0.05 | Significance level — lower is more conservative |
+//! | `n_warmup_steps` | 0 | Skip pruning in the first N steps |
+//! | `n_min_trials` | 1 | Require at least N completed trials before pruning |
+//!
+//! # Example
+//!
+//! ```
+//! use optimizer::Direction;
+//! use optimizer::pruner::WilcoxonPruner;
+//!
+//! let pruner = WilcoxonPruner::new(Direction::Minimize)
+//!     .p_value_threshold(0.05)
+//!     .n_warmup_steps(5)
+//!     .n_min_trials(1);
+//! ```
+
 use core::cmp::Ordering;
 
 use super::Pruner;

src/rng_util.rs

@@ -0,0 +1,58 @@
+use crate::distribution::Distribution;
+
+/// Generate a random `f64` in the range `[low, high)`.
+#[inline]
+pub(crate) fn f64_range(rng: &mut fastrand::Rng, low: f64, high: f64) -> f64 {
+    low + rng.f64() * (high - low)
+}
+
+/// Combine a base seed, trial id, and distribution fingerprint into a
+/// deterministic per-call seed using `MurmurHash3`'s 64-bit finalizer.
+#[inline]
+pub(crate) fn mix_seed(base: u64, trial_id: u64, dist_fingerprint: u64) -> u64 {
+    let mut h = base
+        .wrapping_mul(0xff51_afd7_ed55_8ccd)
+        .wrapping_add(trial_id)
+        .wrapping_mul(0xc4ce_b9fe_1a85_ec53)
+        .wrapping_add(dist_fingerprint);
+    h ^= h >> 33;
+    h = h.wrapping_mul(0xff51_afd7_ed55_8ccd);
+    h ^= h >> 33;
+    h = h.wrapping_mul(0xc4ce_b9fe_1a85_ec53);
+    h ^= h >> 33;
+    h
+}
+
+/// Stable `u64` fingerprint for a [`Distribution`], using variant tags and
+/// `f64::to_bits()` for float fields so that distinct distributions within
+/// the same trial produce different RNG streams.
+#[allow(clippy::cast_possible_truncation, clippy::cast_sign_loss)]
+pub(crate) fn distribution_fingerprint(distribution: &Distribution) -> u64 {
+    match distribution {
+        Distribution::Float(d) => {
+            let mut h: u64 = 1;
+            h = h.wrapping_mul(31).wrapping_add(d.low.to_bits());
+            h = h.wrapping_mul(31).wrapping_add(d.high.to_bits());
+            h = h.wrapping_mul(31).wrapping_add(u64::from(d.log_scale));
+            if let Some(step) = d.step {
+                h = h.wrapping_mul(31).wrapping_add(step.to_bits());
+            }
+            h
+        }
+        Distribution::Int(d) => {
+            let mut h: u64 = 2;
+            h = h.wrapping_mul(31).wrapping_add(d.low as u64);
+            h = h.wrapping_mul(31).wrapping_add(d.high as u64);
+            h = h.wrapping_mul(31).wrapping_add(u64::from(d.log_scale));
+            if let Some(step) = d.step {
+                h = h.wrapping_mul(31).wrapping_add(step as u64);
+            }
+            h
+        }
+        Distribution::Categorical(d) => {
+            let mut h: u64 = 3;
+            h = h.wrapping_mul(31).wrapping_add(d.n_choices as u64);
+            h
+        }
+    }
+}

src/sampler/bohb.rs

@@ -1,13 +1,13 @@
 //! BOHB (Bayesian Optimization + `HyperBand`) sampler.
 //!
-//! BOHB combines TPE's model-guided sampling with Hyperband's budget-aware
+//! BOHB combines TPE's model-guided sampling with `HyperBand`'s budget-aware
 //! evaluation. Instead of building one global TPE model, BOHB conditions
 //! its TPE model on trials evaluated at a specific budget level, giving
-//! better-calibrated proposals for each rung of the Hyperband schedule.
+//! better-calibrated proposals for each rung of the `HyperBand` schedule.
 //!
 //! # How it works
 //!
-//! 1. Compute all Hyperband rung steps (budget levels) from the config.
+//! 1. Compute all `HyperBand` rung steps (budget levels) from the config.
 //! 2. On each `sample()` call, scan the history's `intermediate_values`
 //!    to find the **largest budget level** with enough observations
 //!    (`>= min_points_in_model`).
@@ -16,6 +16,26 @@
 //! 4. Delegate to an internal [`TpeSampler`] for the actual sampling.
 //! 5. Fall back to random sampling if no budget level has enough data.
 //!
+//! # When to use
+//!
+//! - You are tuning hyperparameters for models that support early stopping
+//!   (e.g., neural networks with configurable epoch counts).
+//! - You want to combine model-guided search with aggressive pruning of
+//!   unpromising configurations.
+//! - Your objective has a natural "budget" axis (epochs, iterations, data
+//!   fraction) reported via [`Trial::report`](crate::Trial::report).
+//!
+//! Pair `BohbSampler` with [`matching_pruner`](BohbSampler::matching_pruner)
+//! to get a `HyperBandPruner` whose budget schedule is consistent with
+//! the sampler's conditioning levels.
+//!
+//! # Configuration
+//!
+//! - `min_resource` / `max_resource` — budget range (default: 1 … 81)
+//! - `reduction_factor` (η) — successive halving factor (default: 3)
+//! - `min_points_in_model` — observations needed before TPE replaces random (default: 10)
+//! - All [`TpeSamplerBuilder`](super::tpe::TpeSamplerBuilder) options (gamma, seed, etc.)
+//!
 //! # Examples
 //!
 //! ```
@@ -27,7 +47,7 @@
 //! let study: Study<f64> = Study::with_sampler_and_pruner(Direction::Minimize, bohb, pruner);
 //! ```
 //!
-//! Using the builder for custom configuration:
+//! Custom configuration via builder:
 //!
 //! ```
 //! use optimizer::sampler::bohb::BohbSampler;
@@ -50,7 +70,7 @@ use crate::sampler::tpe::TpeSampler;
 use crate::sampler::{CompletedTrial, Sampler};
 use crate::types::Direction;
 
-/// A BOHB sampler that combines TPE with Hyperband budget awareness.
+/// A BOHB sampler that combines TPE with `HyperBand` budget awareness.
 ///
 /// BOHB filters trial history by budget level before delegating to TPE,
 /// so the surrogate model is conditioned on trials evaluated at the same
@@ -58,7 +78,25 @@ use crate::types::Direction;
 /// than using a single global model across all budgets.
 ///
 /// Use [`BohbSampler::matching_pruner`] to create a [`HyperbandPruner`]
-/// with matching parameters.
+/// with matching `HyperBand` parameters.
+///
+/// # Examples
+///
+/// ```
+/// use optimizer::parameter::{FloatParam, Parameter};
+/// use optimizer::sampler::bohb::BohbSampler;
+/// use optimizer::{Direction, Study};
+///
+/// let bohb = BohbSampler::builder()
+///     .min_resource(1)
+///     .max_resource(27)
+///     .reduction_factor(3)
+///     .seed(42)
+///     .build()
+///     .unwrap();
+/// let pruner = bohb.matching_pruner(Direction::Minimize);
+/// let study: Study<f64> = Study::with_sampler_and_pruner(Direction::Minimize, bohb, pruner);
+/// ```
 pub struct BohbSampler {
     min_resource: u64,
     max_resource: u64,
@@ -225,6 +263,14 @@ impl Sampler for BohbSampler {
 
 /// Builder for configuring a [`BohbSampler`].
 ///
+/// # Defaults
+///
+/// - `min_resource`: 1
+/// - `max_resource`: 81
+/// - `reduction_factor`: 3 (η)
+/// - `min_points_in_model`: 10
+/// - TPE: default settings (gamma = 0.25, etc.)
+///
 /// # Examples
 ///
 /// ```

src/sampler/cma_es.rs

@@ -1,15 +1,55 @@
 //! CMA-ES (Covariance Matrix Adaptation Evolution Strategy) sampler.
 //!
-//! CMA-ES maintains a multivariate Gaussian distribution over continuous
-//! parameters and adapts its mean, covariance matrix, and step-size based
-//! on trial rankings. It is one of the most effective derivative-free
-//! optimizers for continuous search spaces.
+//! CMA-ES is a stochastic, population-based optimizer that maintains a
+//! multivariate Gaussian distribution over continuous parameters and adapts
+//! its **mean**, **covariance matrix**, and **step-size** (σ) based on
+//! trial rankings. It is widely regarded as one of the most effective
+//! derivative-free optimizers for continuous search spaces.
 //!
-//! Categorical parameters are sampled uniformly at random (not part of
-//! the CMA-ES vector). If all parameters are categorical, the sampler
-//! falls back to pure random sampling.
+//! # Algorithm overview
 //!
-//! Requires the `cma-es` feature flag.
+//! Each generation:
+//! 1. **Sample** λ (population size) candidates from N(m, σ²C).
+//! 2. **Evaluate** and **rank** the candidates by objective value.
+//! 3. **Update** the mean toward the best μ candidates (weighted recombination).
+//! 4. **Adapt** the covariance matrix C via rank-one and rank-μ updates, and
+//!    adapt σ via cumulative step-size adaptation (CSA).
+//!
+//! Over time the search distribution narrows and rotates to align with the
+//! landscape, efficiently exploiting structure in the objective function.
+//!
+//! # When to use
+//!
+//! - **Continuous parameters only** (float/int). Categorical parameters are
+//!   sampled uniformly at random and do not participate in the CMA-ES model.
+//! - **Moderate dimensionality** — works well up to ~100 continuous dimensions.
+//!   Beyond that, the O(n²) covariance matrix becomes expensive to maintain.
+//! - **Non-separable objectives** — CMA-ES learns parameter correlations
+//!   through the covariance matrix, making it especially effective on
+//!   rotated or ill-conditioned landscapes.
+//! - **Moderate evaluation budgets** — typically needs ≈10×n to 100×n
+//!   evaluations to converge, where n is the number of continuous dimensions.
+//!
+//! For very cheap evaluations in low dimensions (d ≤ 20), consider
+//! [`GpSampler`](super::gp::GpSampler) instead. For high-dimensional
+//! separable problems, TPE may be more efficient.
+//!
+//! # Configuration
+//!
+//! | Option | Default | Description |
+//! |--------|---------|-------------|
+//! | `sigma0` | `avg_range / 4` | Initial step size — controls exploration breadth |
+//! | `population_size` | `4 + ⌊3 ln n⌋` | Candidates per generation (λ) |
+//! | `seed` | random | RNG seed for reproducibility |
+//!
+//! # Feature flag
+//!
+//! Requires the **`cma-es`** feature (adds the `nalgebra` dependency):
+//!
+//! ```toml
+//! [dependencies]
+//! optimizer = { version = "...", features = ["cma-es"] }
+//! ```
 //!
 //! # Examples
 //!
@@ -17,21 +57,28 @@
 //! use optimizer::sampler::cma_es::CmaEsSampler;
 //! use optimizer::{Direction, Study};
 //!
-//! let sampler = CmaEsSampler::with_seed(42);
-//! let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+//! // Minimize a 2-D sphere function with CMA-ES
+//! let sampler = CmaEsSampler::builder()
+//!     .sigma0(1.0)
+//!     .population_size(10)
+//!     .seed(42)
+//!     .build();
+//!
+//! let mut study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
 //! ```
 
 use std::collections::HashMap;
 
 use nalgebra::{DMatrix, DVector};
 use parking_lot::Mutex;
-use rand::rngs::StdRng;
-use rand::{RngExt, SeedableRng};
 
 use crate::distribution::Distribution;
 use crate::param::ParamValue;
+use crate::rng_util;
 use crate::sampler::{CompletedTrial, Sampler};
 
+use super::common::{from_internal, internal_bounds, sample_random};
+
 /// CMA-ES sampler for continuous optimization.
 ///
 /// Adapts a multivariate Gaussian to concentrate around promising regions
@@ -267,7 +314,7 @@ enum Phase {
 /// Top-level mutable state behind the `Mutex`.
 struct CmaEsState {
     /// The RNG used for sampling.
-    rng: StdRng,
+    rng: fastrand::Rng,
     /// User-provided initial sigma (None = auto).
     sigma0: Option<f64>,
     /// User-provided population size (None = auto).
@@ -290,7 +337,7 @@ struct CmaEsState {
 
 impl CmaEsState {
     fn new(sigma0: Option<f64>, user_lambda: Option<usize>, seed: Option<u64>) -> Self {
-        let rng = seed.map_or_else(rand::make_rng, StdRng::seed_from_u64);
+        let rng = seed.map_or_else(fastrand::Rng::new, fastrand::Rng::with_seed);
         Self {
             rng,
             sigma0,
@@ -421,7 +468,7 @@ impl CmaEsAlgorithm {
     /// Generate `lambda` candidate vectors from the current distribution.
     fn generate_candidates(
         &self,
-        rng: &mut StdRng,
+        rng: &mut fastrand::Rng,
         dimensions: &[DimensionInfo],
     ) -> Vec<Candidate> {
         let n = self.constants.n;
@@ -439,7 +486,7 @@ impl CmaEsAlgorithm {
     /// Generate a single candidate from the current distribution.
     fn generate_single_candidate(
         &self,
-        rng: &mut StdRng,
+        rng: &mut fastrand::Rng,
         dimensions: &[DimensionInfo],
         n: usize,
     ) -> Candidate {
@@ -452,7 +499,7 @@ impl CmaEsAlgorithm {
             if !dim.is_continuous
                 && let Distribution::Categorical(cat) = &dim.distribution
             {
-                categorical_values.insert(i, rng.random_range(0..cat.n_choices));
+                categorical_values.insert(i, rng.usize(0..cat.n_choices));
             }
         }
 
@@ -465,7 +512,7 @@ impl CmaEsAlgorithm {
     /// Sample a candidate vector with rejection sampling for bounds.
     fn sample_with_rejection(
         &self,
-        rng: &mut StdRng,
+        rng: &mut fastrand::Rng,
         dimensions: &[DimensionInfo],
         n: usize,
     ) -> DVector<f64> {
@@ -626,111 +673,14 @@ fn clip_to_bounds(x: &mut DVector<f64>, dimensions: &[DimensionInfo]) {
     }
 }
 
-/// Convert an internal-space value back to a `ParamValue`.
-#[allow(clippy::cast_possible_truncation, clippy::cast_precision_loss)]
-fn from_internal(value: f64, distribution: &Distribution) -> ParamValue {
-    match distribution {
-        Distribution::Float(d) => {
-            let v = if d.log_scale { value.exp() } else { value };
-            let v = if let Some(step) = d.step {
-                let k = ((v - d.low) / step).round();
-                d.low + k * step
-            } else {
-                v
-            };
-            ParamValue::Float(v.clamp(d.low, d.high))
-        }
-        Distribution::Int(d) => {
-            let v = if d.log_scale { value.exp() } else { value };
-            let v = if let Some(step) = d.step {
-                let k = ((v - d.low as f64) / step as f64).round() as i64;
-                d.low + k * step
-            } else {
-                v.round() as i64
-            };
-            ParamValue::Int(v.clamp(d.low, d.high))
-        }
-        Distribution::Categorical(_) => {
-            unreachable!("from_internal should not be called for categorical distributions")
-        }
-    }
-}
-
-/// Compute internal-space bounds for a distribution.
-#[allow(clippy::cast_precision_loss)]
-fn internal_bounds(distribution: &Distribution) -> Option<(f64, f64)> {
-    match distribution {
-        Distribution::Float(d) => {
-            if d.log_scale {
-                Some((d.low.ln(), d.high.ln()))
-            } else {
-                Some((d.low, d.high))
-            }
-        }
-        Distribution::Int(d) => {
-            if d.log_scale {
-                Some(((d.low as f64).ln(), (d.high as f64).ln()))
-            } else {
-                Some((d.low as f64, d.high as f64))
-            }
-        }
-        Distribution::Categorical(_) => None,
-    }
-}
-
 /// Sample a value from the standard normal distribution using Box-Muller transform.
-fn sample_standard_normal(rng: &mut StdRng) -> f64 {
+fn sample_standard_normal(rng: &mut fastrand::Rng) -> f64 {
     // Box-Muller transform
-    let u1: f64 = rng.random_range(f64::EPSILON..=1.0);
-    let u2: f64 = rng.random_range(0.0_f64..=core::f64::consts::TAU);
+    let u1: f64 = rng_util::f64_range(rng, f64::EPSILON, 1.0);
+    let u2: f64 = rng_util::f64_range(rng, 0.0_f64, core::f64::consts::TAU);
     (-2.0 * u1.ln()).sqrt() * u2.cos()
 }
 
-/// Sample a categorical value randomly.
-fn sample_random_categorical(rng: &mut StdRng, distribution: &Distribution) -> ParamValue {
-    match distribution {
-        Distribution::Categorical(d) => ParamValue::Categorical(rng.random_range(0..d.n_choices)),
-        _ => unreachable!("sample_random_categorical called with non-categorical distribution"),
-    }
-}
-
-/// Sample a random value for any distribution (used during discovery phase).
-#[allow(clippy::cast_possible_truncation, clippy::cast_precision_loss)]
-fn sample_random(rng: &mut StdRng, distribution: &Distribution) -> ParamValue {
-    match distribution {
-        Distribution::Float(d) => {
-            let value = if d.log_scale {
-                let log_low = d.low.ln();
-                let log_high = d.high.ln();
-                rng.random_range(log_low..=log_high).exp()
-            } else if let Some(step) = d.step {
-                let n_steps = ((d.high - d.low) / step).floor() as i64;
-                let k = rng.random_range(0..=n_steps);
-                d.low + (k as f64) * step
-            } else {
-                rng.random_range(d.low..=d.high)
-            };
-            ParamValue::Float(value)
-        }
-        Distribution::Int(d) => {
-            let value = if d.log_scale {
-                let log_low = (d.low as f64).ln();
-                let log_high = (d.high as f64).ln();
-                let raw = rng.random_range(log_low..=log_high).exp().round() as i64;
-                raw.clamp(d.low, d.high)
-            } else if let Some(step) = d.step {
-                let n_steps = (d.high - d.low) / step;
-                let k = rng.random_range(0..=n_steps);
-                d.low + k * step
-            } else {
-                rng.random_range(d.low..=d.high)
-            };
-            ParamValue::Int(value)
-        }
-        Distribution::Categorical(d) => ParamValue::Categorical(rng.random_range(0..d.n_choices)),
-    }
-}
-
 // ---------------------------------------------------------------------------
 // Sampler trait implementation
 // ---------------------------------------------------------------------------
@@ -811,7 +761,7 @@ fn finalize_discovery(state: &mut CmaEsState) {
 
 /// Generate candidates that are purely categorical (no continuous dims).
 fn generate_pure_categorical_candidates(
-    rng: &mut StdRng,
+    rng: &mut fastrand::Rng,
     dimensions: &[DimensionInfo],
     lambda: usize,
 ) -> Vec<Candidate> {
@@ -820,7 +770,7 @@ fn generate_pure_categorical_candidates(
             let mut categorical_values = HashMap::new();
             for (i, dim) in dimensions.iter().enumerate() {
                 if let Distribution::Categorical(cat) = &dim.distribution {
-                    categorical_values.insert(i, rng.random_range(0..cat.n_choices));
+                    categorical_values.insert(i, rng.usize(0..cat.n_choices));
                 }
             }
             Candidate {
@@ -875,7 +825,7 @@ fn sample_active(
         if let Some(&cat_idx) = candidate.categorical_values.get(&dim_idx) {
             ParamValue::Categorical(cat_idx)
         } else {
-            sample_random_categorical(&mut state.rng, distribution)
+            sample_random(&mut state.rng, distribution)
         }
     }
 }
@@ -913,7 +863,7 @@ fn generate_overflow_candidate(state: &mut CmaEsState) -> Candidate {
         let mut categorical_values = HashMap::new();
         for (i, dim) in state.dimensions.iter().enumerate() {
             if let Distribution::Categorical(cat) = &dim.distribution {
-                categorical_values.insert(i, state.rng.random_range(0..cat.n_choices));
+                categorical_values.insert(i, state.rng.usize(0..cat.n_choices));
             }
         }
         return Candidate {

src/sampler/common.rs

@@ -0,0 +1,134 @@
+//! Shared distribution-level utilities used across multiple samplers.
+
+use crate::distribution::Distribution;
+use crate::param::ParamValue;
+use crate::rng_util;
+
+/// Compute internal-space bounds for a distribution.
+#[allow(clippy::cast_precision_loss)]
+pub(crate) fn internal_bounds(distribution: &Distribution) -> Option<(f64, f64)> {
+    match distribution {
+        Distribution::Float(d) => {
+            if d.log_scale {
+                if d.low <= 0.0 || d.high <= 0.0 {
+                    return None;
+                }
+                Some((d.low.ln(), d.high.ln()))
+            } else {
+                Some((d.low, d.high))
+            }
+        }
+        Distribution::Int(d) => {
+            if d.log_scale {
+                if d.low < 1 {
+                    return None;
+                }
+                Some(((d.low as f64).ln(), (d.high as f64).ln()))
+            } else {
+                Some((d.low as f64, d.high as f64))
+            }
+        }
+        Distribution::Categorical(_) => None,
+    }
+}
+
+/// Convert an internal-space value back to a `ParamValue`.
+#[allow(clippy::cast_possible_truncation, clippy::cast_precision_loss)]
+pub(crate) fn from_internal(value: f64, distribution: &Distribution) -> ParamValue {
+    match distribution {
+        Distribution::Float(d) => {
+            let v = if d.log_scale { value.exp() } else { value };
+            let v = if let Some(step) = d.step {
+                let k = ((v - d.low) / step).round();
+                d.low + k * step
+            } else {
+                v
+            };
+            ParamValue::Float(v.clamp(d.low, d.high))
+        }
+        Distribution::Int(d) => {
+            let v = if d.log_scale { value.exp() } else { value };
+            let v = if let Some(step) = d.step {
+                let k = ((v - d.low as f64) / step as f64).round() as i64;
+                d.low.saturating_add(k.saturating_mul(step))
+            } else {
+                v.round() as i64
+            };
+            ParamValue::Int(v.clamp(d.low, d.high))
+        }
+        Distribution::Categorical(_) => {
+            unreachable!("from_internal should not be called for categorical distributions")
+        }
+    }
+}
+
+/// Convert a `ParamValue` to its internal-space representation.
+#[allow(clippy::cast_precision_loss, dead_code)]
+pub(crate) fn to_internal(value: &ParamValue, distribution: &Distribution) -> f64 {
+    match (value, distribution) {
+        (ParamValue::Float(v), Distribution::Float(d)) => {
+            if d.log_scale {
+                v.ln()
+            } else {
+                *v
+            }
+        }
+        (ParamValue::Int(v), Distribution::Int(d)) => {
+            if d.log_scale {
+                (*v as f64).ln()
+            } else {
+                *v as f64
+            }
+        }
+        _ => 0.0,
+    }
+}
+
+/// Sample a random value for any distribution.
+#[allow(clippy::cast_possible_truncation, clippy::cast_precision_loss)]
+pub(crate) fn sample_random(rng: &mut fastrand::Rng, distribution: &Distribution) -> ParamValue {
+    match distribution {
+        Distribution::Float(d) => {
+            let value = if d.log_scale {
+                let log_low = d.low.ln();
+                let log_high = d.high.ln();
+                let v = rng_util::f64_range(rng, log_low, log_high).exp();
+                if let Some(step) = d.step {
+                    let k = ((v - d.low) / step).round();
+                    (d.low + k * step).clamp(d.low, d.high)
+                } else {
+                    v
+                }
+            } else if let Some(step) = d.step {
+                let n_steps = ((d.high - d.low) / step).floor() as i64;
+                let k = rng.i64(0..=n_steps);
+                d.low + (k as f64) * step
+            } else {
+                rng_util::f64_range(rng, d.low, d.high)
+            };
+            ParamValue::Float(value)
+        }
+        Distribution::Int(d) => {
+            let value = if d.log_scale {
+                let log_low = (d.low as f64).ln();
+                let log_high = (d.high as f64).ln();
+                let v = rng_util::f64_range(rng, log_low, log_high).exp();
+                let raw = if let Some(step) = d.step {
+                    let k = ((v - d.low as f64) / step as f64).round() as i64;
+                    d.low.saturating_add(k.saturating_mul(step))
+                } else {
+                    v.round() as i64
+                };
+                raw.clamp(d.low, d.high)
+            } else if let Some(step) = d.step {
+                let n_steps = (d.high - d.low) / step;
+                let k = rng.i64(0..=n_steps);
+                d.low + k * step
+            } else {
+                rng.i64(d.low..=d.high)
+            };
+            ParamValue::Int(value)
+        }
+        Distribution::Categorical(d) => ParamValue::Categorical(rng.usize(0..d.n_choices)),
+    }
+}

src/sampler/de.rs

@@ -0,0 +1,926 @@
+//! Differential Evolution (DE) sampler.
+//!
+//! DE is a population-based metaheuristic that maintains a pool of candidate
+//! solutions and creates new candidates through **mutation** (combining
+//! difference vectors of existing members) and **binomial crossover**. A
+//! trial vector replaces its parent only if it achieves a better objective
+//! value, guaranteeing monotonic improvement of the population.
+//!
+//! # Algorithm overview
+//!
+//! Each generation, for every population member *xᵢ*:
+//! 1. **Mutation** — create a mutant vector *v* from other population
+//!    members using the selected [`DEStrategy`]:
+//!    - `Rand1`:  `v = x_r1 + F * (x_r2 - x_r3)`
+//!    - `Best1`:  `v = x_best + F * (x_r1 - x_r2)`
+//!    - `CurrentToBest1`:  `v = x_i + F * (x_best - x_i) + F * (x_r1 - x_r2)`
+//! 2. **Crossover** — create a trial vector *u* by mixing *v* and *xᵢ*
+//!    dimension-by-dimension with probability CR.
+//! 3. **Selection** — replace *xᵢ* with *u* if `f(u) ≤ f(xᵢ)`.
+//!
+//! # When to use
+//!
+//! - **Continuous parameters** (float/int). Categorical parameters are
+//!   sampled uniformly at random and do not participate in DE.
+//! - **Moderate to large search spaces** — DE scales better than GP-based
+//!   methods to higher dimensions, though it may need more evaluations.
+//! - **Multi-modal landscapes** — the `Rand1` strategy maintains diversity
+//!   and avoids premature convergence.
+//! - **No feature flags required** — DE is available with default features.
+//!
+//! For non-separable problems in moderate dimensions, consider
+//! `CmaEsSampler` which learns parameter
+//! correlations. For expensive functions with few dimensions, consider
+//! `GpSampler`.
+//!
+//! # Configuration
+//!
+//! | Option | Default | Description |
+//! |--------|---------|-------------|
+//! | `population_size` | `max(10n, 15)` | Candidates per generation |
+//! | `mutation_factor` (F) | 0.8 | Differential amplification — higher = more exploration |
+//! | `crossover_rate` (CR) | 0.9 | Probability of taking a dimension from the mutant |
+//! | `strategy` | `Rand1` | Mutation strategy (see [`DEStrategy`]) |
+//! | `seed` | random | RNG seed for reproducibility |
+//!
+//! # Examples
+//!
+//! ```
+//! use optimizer::sampler::de::{DESampler, DEStrategy};
+//! use optimizer::{Direction, Study};
+//!
+//! // Minimize with DE using the Best1 strategy for faster convergence
+//! let sampler = DESampler::builder()
+//!     .mutation_factor(0.7)
+//!     .crossover_rate(0.9)
+//!     .strategy(DEStrategy::Best1)
+//!     .population_size(20)
+//!     .seed(42)
+//!     .build();
+//!
+//! let mut study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+//! ```
+
+use std::collections::HashMap;
+
+use parking_lot::Mutex;
+
+use crate::distribution::Distribution;
+use crate::param::ParamValue;
+use crate::rng_util;
+use crate::sampler::{CompletedTrial, Sampler};
+
+use super::common::{from_internal, internal_bounds, sample_random};
+
+/// Differential Evolution mutation strategy.
+///
+/// Controls how mutant vectors are created from the current population.
+#[derive(Clone, Copy, Debug, Default)]
+pub enum DEStrategy {
+    /// DE/rand/1: `v = x_r1 + F * (x_r2 - x_r3)`
+    ///
+    /// The most robust strategy. Uses three random population members.
+    #[default]
+    Rand1,
+    /// DE/best/1: `v = x_best + F * (x_r1 - x_r2)`
+    ///
+    /// Greedier strategy that biases toward the current best solution.
+    Best1,
+    /// DE/current-to-best/1: `v = x_i + F * (x_best - x_i) + F * (x_r1 - x_r2)`
+    ///
+    /// Balances exploration and exploitation by blending the current
+    /// individual with the best.
+    CurrentToBest1,
+}
+
+/// Differential Evolution sampler for continuous global optimization.
+///
+/// Maintains a population of candidate solutions. New candidates are
+/// created by combining (mutating + crossing over) existing members.
+///
+/// # Examples
+///
+/// ```
+/// use optimizer::sampler::de::DESampler;
+/// use optimizer::{Direction, Study};
+///
+/// // Default configuration
+/// let study: Study<f64> = Study::with_sampler(Direction::Minimize, DESampler::new());
+///
+/// // With seed for reproducibility
+/// let study: Study<f64> = Study::with_sampler(Direction::Minimize, DESampler::with_seed(42));
+///
+/// // Custom configuration via builder
+/// use optimizer::sampler::de::DEStrategy;
+/// let sampler = DESampler::builder()
+///     .mutation_factor(0.8)
+///     .crossover_rate(0.9)
+///     .strategy(DEStrategy::Best1)
+///     .population_size(30)
+///     .seed(42)
+///     .build();
+/// let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+/// ```
+pub struct DESampler {
+    state: Mutex<State>,
+}
+
+impl DESampler {
+    /// Creates a new DE sampler with default settings and a random seed.
+    #[must_use]
+    pub fn new() -> Self {
+        Self {
+            state: Mutex::new(State::new(None, 0.8, 0.9, DEStrategy::Rand1, None)),
+        }
+    }
+
+    /// Creates a new DE sampler with a fixed seed for reproducibility.
+    #[must_use]
+    pub fn with_seed(seed: u64) -> Self {
+        Self {
+            state: Mutex::new(State::new(None, 0.8, 0.9, DEStrategy::Rand1, Some(seed))),
+        }
+    }
+
+    /// Creates a builder for configuring a `DESampler`.
+    #[must_use]
+    pub fn builder() -> DESamplerBuilder {
+        DESamplerBuilder::new()
+    }
+}
+
+impl Default for DESampler {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+/// Builder for configuring a [`DESampler`].
+///
+/// All options have sensible defaults:
+/// - `population_size`: `max(10 * n_dims, 15)` (auto-computed from parameter count)
+/// - `mutation_factor` (F): 0.8
+/// - `crossover_rate` (CR): 0.9
+/// - `strategy`: `Rand1`
+/// - `seed`: random
+///
+/// # Examples
+///
+/// ```
+/// use optimizer::sampler::de::{DESamplerBuilder, DEStrategy};
+///
+/// let sampler = DESamplerBuilder::new()
+///     .mutation_factor(0.5)
+///     .crossover_rate(0.7)
+///     .strategy(DEStrategy::CurrentToBest1)
+///     .population_size(20)
+///     .seed(42)
+///     .build();
+/// ```
+#[derive(Debug, Clone)]
+pub struct DESamplerBuilder {
+    population_size: Option<usize>,
+    mutation_factor: f64,
+    crossover_rate: f64,
+    strategy: DEStrategy,
+    seed: Option<u64>,
+}
+
+impl Default for DESamplerBuilder {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl DESamplerBuilder {
+    /// Creates a new builder with default settings.
+    #[must_use]
+    pub fn new() -> Self {
+        Self {
+            population_size: None,
+            mutation_factor: 0.8,
+            crossover_rate: 0.9,
+            strategy: DEStrategy::Rand1,
+            seed: None,
+        }
+    }
+
+    /// Sets the population size.
+    ///
+    /// Number of candidate solutions maintained across generations.
+    /// Larger populations improve robustness but require more evaluations
+    /// per generation.
+    ///
+    /// Default: `max(10 * n_continuous_dims, 15)`.
+    #[must_use]
+    pub fn population_size(mut self, size: usize) -> Self {
+        self.population_size = Some(size);
+        self
+    }
+
+    /// Sets the mutation factor (F).
+    ///
+    /// Controls the amplification of differential variation.
+    /// Typical values are in `[0.5, 1.0]`. Higher values increase
+    /// exploration; lower values favor exploitation.
+    ///
+    /// Default: 0.8.
+    #[must_use]
+    pub fn mutation_factor(mut self, f: f64) -> Self {
+        self.mutation_factor = f;
+        self
+    }
+
+    /// Sets the crossover rate (CR).
+    ///
+    /// Probability of each dimension being taken from the mutant vector
+    /// rather than the parent. Typical values are in `[0.7, 1.0]`.
+    ///
+    /// Default: 0.9.
+    #[must_use]
+    pub fn crossover_rate(mut self, cr: f64) -> Self {
+        self.crossover_rate = cr;
+        self
+    }
+
+    /// Sets the mutation strategy.
+    ///
+    /// Default: [`DEStrategy::Rand1`].
+    #[must_use]
+    pub fn strategy(mut self, strategy: DEStrategy) -> Self {
+        self.strategy = strategy;
+        self
+    }
+
+    /// Sets the random seed for reproducibility.
+    #[must_use]
+    pub fn seed(mut self, seed: u64) -> Self {
+        self.seed = Some(seed);
+        self
+    }
+
+    /// Builds the configured [`DESampler`].
+    #[must_use]
+    pub fn build(self) -> DESampler {
+        DESampler {
+            state: Mutex::new(State::new(
+                self.population_size,
+                self.mutation_factor,
+                self.crossover_rate,
+                self.strategy,
+                self.seed,
+            )),
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Internal types
+// ---------------------------------------------------------------------------
+
+/// Describes how a parameter dimension maps into the DE internal vector.
+#[derive(Clone, Debug)]
+struct DimensionInfo {
+    /// The distribution for this dimension (stored for decoding).
+    distribution: Distribution,
+    /// Whether this dimension participates in DE (Float/Int = true, Categorical = false).
+    is_continuous: bool,
+    /// Internal-space bounds for continuous dimensions: `(low, high)`.
+    /// For log-scale parameters these are in log-space.
+    bounds: Option<(f64, f64)>,
+}
+
+/// A candidate solution produced by mutation + crossover.
+#[derive(Clone, Debug)]
+struct Candidate {
+    /// Internal-space vector (only continuous dimensions).
+    x: Vec<f64>,
+    /// Values for categorical dimensions (index in `dimensions` -> categorical index).
+    categorical_values: HashMap<usize, usize>,
+    /// Index of the population member this candidate competes against.
+    target_idx: usize,
+}
+
+/// Tracks per-trial sampling progress.
+#[derive(Clone, Debug)]
+struct TrialProgress {
+    /// Index of the candidate assigned to this trial.
+    candidate_idx: usize,
+    /// Next dimension to return for this trial.
+    next_dim: usize,
+}
+
+/// Phase of the DE state machine.
+enum Phase {
+    /// Discovering the search space structure (first trial).
+    Discovery,
+    /// Active sampling and evolving.
+    Active,
+}
+
+/// Top-level mutable state behind the `Mutex`.
+struct State {
+    /// The RNG used for sampling.
+    rng: fastrand::Rng,
+    /// User-provided population size (None = auto).
+    user_population_size: Option<usize>,
+    /// Mutation factor (F).
+    mutation_factor: f64,
+    /// Crossover rate (CR).
+    crossover_rate: f64,
+    /// Mutation strategy.
+    strategy: DEStrategy,
+    /// Current phase.
+    phase: Phase,
+    /// Discovered dimension info (populated during discovery).
+    dimensions: Vec<DimensionInfo>,
+    /// Last `trial_id` seen during discovery.
+    discovery_trial_id: Option<u64>,
+
+    // --- Population state ---
+    /// Current population (internal-space vectors, continuous dims only).
+    population: Vec<Vec<f64>>,
+    /// Categorical values for each population member.
+    population_categorical: Vec<HashMap<usize, usize>>,
+    /// Objective values for the current population.
+    population_values: Vec<f64>,
+    /// Index of the best population member.
+    best_idx: usize,
+    /// Whether the initial population has been evaluated.
+    initialized: bool,
+    /// Effective population size (resolved after discovery).
+    population_size: usize,
+
+    // --- Current generation ---
+    /// Current generation's candidates.
+    candidates: Vec<Candidate>,
+    /// Mapping from `trial_id` to its progress.
+    trial_progress: HashMap<u64, TrialProgress>,
+    /// Number of candidates assigned so far in the current generation.
+    assigned_count: usize,
+    /// Trial IDs assigned in the current generation.
+    generation_trial_ids: Vec<u64>,
+}
+
+impl State {
+    fn new(
+        user_population_size: Option<usize>,
+        mutation_factor: f64,
+        crossover_rate: f64,
+        strategy: DEStrategy,
+        seed: Option<u64>,
+    ) -> Self {
+        let rng = seed.map_or_else(fastrand::Rng::new, fastrand::Rng::with_seed);
+        Self {
+            rng,
+            user_population_size,
+            mutation_factor,
+            crossover_rate,
+            strategy,
+            phase: Phase::Discovery,
+            dimensions: Vec::new(),
+            discovery_trial_id: None,
+            population: Vec::new(),
+            population_categorical: Vec::new(),
+            population_values: Vec::new(),
+            best_idx: 0,
+            initialized: false,
+            population_size: 0,
+            candidates: Vec::new(),
+            trial_progress: HashMap::new(),
+            assigned_count: 0,
+            generation_trial_ids: Vec::new(),
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Helpers
+// ---------------------------------------------------------------------------
+
+/// Sample a random value in internal space for a continuous dimension.
+fn sample_random_internal(rng: &mut fastrand::Rng, bounds: (f64, f64)) -> f64 {
+    rng_util::f64_range(rng, bounds.0, bounds.1)
+}
+
+/// Clamp a value to the given bounds.
+fn clamp_to_bounds(value: f64, bounds: Option<(f64, f64)>) -> f64 {
+    if let Some((lo, hi)) = bounds {
+        value.clamp(lo, hi)
+    } else {
+        value
+    }
+}
+
+// ---------------------------------------------------------------------------
+// DE algorithm
+// ---------------------------------------------------------------------------
+
+/// Select `count` distinct random indices from `0..n`, all different from `exclude`.
+fn select_random_indices(
+    rng: &mut fastrand::Rng,
+    n: usize,
+    count: usize,
+    exclude: &[usize],
+) -> Vec<usize> {
+    let mut selected = Vec::with_capacity(count);
+    while selected.len() < count {
+        let idx = rng.usize(0..n);
+        if !exclude.contains(&idx) && !selected.contains(&idx) {
+            selected.push(idx);
+        }
+    }
+    selected
+}
+
+/// Generate trial vectors (mutation + crossover) for the current population.
+fn generate_trial_vectors(state: &mut State) -> Vec<Candidate> {
+    let n_continuous = state.dimensions.iter().filter(|d| d.is_continuous).count();
+    let pop_size = state.population_size;
+
+    let mut candidates = Vec::with_capacity(pop_size);
+
+    for i in 0..pop_size {
+        // Mutation
+        let mutant = create_mutant_with_rng(state, i, n_continuous);
+
+        // Crossover (binomial)
+        let j_rand = state.rng.usize(0..n_continuous.max(1));
+        let trial_x: Vec<f64> = if n_continuous > 0 {
+            (0..n_continuous)
+                .map(|j| {
+                    let use_mutant = j == j_rand || state.rng.f64() < state.crossover_rate;
+                    let val = if use_mutant {
+                        mutant[j]
+                    } else {
+                        state.population[i][j]
+                    };
+                    // Clamp to bounds
+                    let dim_bounds = continuous_dim_bounds(&state.dimensions, j);
+                    clamp_to_bounds(val, dim_bounds)
+                })
+                .collect()
+        } else {
+            Vec::new()
+        };
+
+        // Categorical: randomly sample (DE doesn't optimize categoricals)
+        let mut categorical_values = HashMap::new();
+        for (dim_idx, dim) in state.dimensions.iter().enumerate() {
+            if !dim.is_continuous
+                && let Distribution::Categorical(cat) = &dim.distribution
+            {
+                categorical_values.insert(dim_idx, state.rng.usize(0..cat.n_choices));
+            }
+        }
+
+        candidates.push(Candidate {
+            x: trial_x,
+            categorical_values,
+            target_idx: i,
+        });
+    }
+
+    candidates
+}
+
+/// Create a mutant vector, consuming RNG from state.
+fn create_mutant_with_rng(state: &mut State, target_idx: usize, n_continuous: usize) -> Vec<f64> {
+    if n_continuous == 0 {
+        return Vec::new();
+    }
+
+    let pop = &state.population;
+    let best_idx = state.best_idx;
+    let f = state.mutation_factor;
+    let pop_size = state.population_size;
+
+    match state.strategy {
+        DEStrategy::Rand1 => {
+            let indices = select_random_indices(&mut state.rng, pop_size, 3, &[target_idx]);
+            let (r1, r2, r3) = (indices[0], indices[1], indices[2]);
+            (0..n_continuous)
+                .map(|j| pop[r1][j] + f * (pop[r2][j] - pop[r3][j]))
+                .collect()
+        }
+        DEStrategy::Best1 => {
+            let indices = select_random_indices(&mut state.rng, pop_size, 2, &[target_idx]);
+            let (r1, r2) = (indices[0], indices[1]);
+            (0..n_continuous)
+                .map(|j| pop[best_idx][j] + f * (pop[r1][j] - pop[r2][j]))
+                .collect()
+        }
+        DEStrategy::CurrentToBest1 => {
+            let indices = select_random_indices(&mut state.rng, pop_size, 2, &[target_idx]);
+            let (r1, r2) = (indices[0], indices[1]);
+            (0..n_continuous)
+                .map(|j| {
+                    pop[target_idx][j]
+                        + f * (pop[best_idx][j] - pop[target_idx][j])
+                        + f * (pop[r1][j] - pop[r2][j])
+                })
+                .collect()
+        }
+    }
+}
+
+/// Get the bounds for the j-th continuous dimension.
+fn continuous_dim_bounds(
+    dimensions: &[DimensionInfo],
+    continuous_idx: usize,
+) -> Option<(f64, f64)> {
+    let mut ci = 0;
+    for dim in dimensions {
+        if dim.is_continuous {
+            if ci == continuous_idx {
+                return dim.bounds;
+            }
+            ci += 1;
+        }
+    }
+    None
+}
+
+/// Generate the initial random population.
+fn generate_initial_population(state: &mut State) -> Vec<Candidate> {
+    let n_continuous = state.dimensions.iter().filter(|d| d.is_continuous).count();
+
+    let mut candidates = Vec::with_capacity(state.population_size);
+
+    for i in 0..state.population_size {
+        let x: Vec<f64> = if n_continuous > 0 {
+            let mut v = Vec::with_capacity(n_continuous);
+            for dim in &state.dimensions {
+                if dim.is_continuous {
+                    let val = if let Some(bounds) = dim.bounds {
+                        sample_random_internal(&mut state.rng, bounds)
+                    } else {
+                        0.0
+                    };
+                    v.push(val);
+                }
+            }
+            v
+        } else {
+            Vec::new()
+        };
+
+        let mut categorical_values = HashMap::new();
+        for (dim_idx, dim) in state.dimensions.iter().enumerate() {
+            if !dim.is_continuous
+                && let Distribution::Categorical(cat) = &dim.distribution
+            {
+                categorical_values.insert(dim_idx, state.rng.usize(0..cat.n_choices));
+            }
+        }
+
+        candidates.push(Candidate {
+            x,
+            categorical_values,
+            target_idx: i,
+        });
+    }
+
+    candidates
+}
+
+// ---------------------------------------------------------------------------
+// Sampler trait implementation
+// ---------------------------------------------------------------------------
+
+impl Sampler for DESampler {
+    #[allow(clippy::cast_precision_loss)]
+    fn sample(
+        &self,
+        distribution: &Distribution,
+        trial_id: u64,
+        history: &[CompletedTrial],
+    ) -> ParamValue {
+        let mut state = self.state.lock();
+
+        match &state.phase {
+            Phase::Discovery => sample_discovery(&mut state, distribution, trial_id),
+            Phase::Active => sample_active(&mut state, distribution, trial_id, history),
+        }
+    }
+}
+
+/// Handle sampling during the discovery phase.
+fn sample_discovery(state: &mut State, distribution: &Distribution, trial_id: u64) -> ParamValue {
+    // Check if this is a new trial (discovery phase ended for previous trial)
+    if let Some(prev_id) = state.discovery_trial_id
+        && trial_id != prev_id
+    {
+        // First trial is done; we know the search space. Initialize DE.
+        finalize_discovery(state);
+        return sample_active(state, distribution, trial_id, &[]);
+    }
+
+    // Record this trial_id
+    state.discovery_trial_id = Some(trial_id);
+
+    // Record this dimension
+    let is_continuous = !matches!(distribution, Distribution::Categorical(_));
+    let bounds = internal_bounds(distribution);
+    state.dimensions.push(DimensionInfo {
+        distribution: distribution.clone(),
+        is_continuous,
+        bounds,
+    });
+
+    // Sample randomly for the discovery trial
+    sample_random(&mut state.rng, distribution)
+}
+
+/// Finalize discovery and transition to the active phase.
+#[allow(clippy::cast_precision_loss)]
+fn finalize_discovery(state: &mut State) {
+    let n_continuous = state.dimensions.iter().filter(|d| d.is_continuous).count();
+
+    // Resolve population size
+    state.population_size = state
+        .user_population_size
+        .unwrap_or_else(|| (10 * n_continuous).max(15));
+
+    // Ensure population size is at least 4 (DE needs distinct random indices)
+    state.population_size = state.population_size.max(4);
+
+    // Generate initial random population
+    state.candidates = generate_initial_population(state);
+    state.assigned_count = 0;
+    state.generation_trial_ids.clear();
+    state.trial_progress.clear();
+    state.phase = Phase::Active;
+}
+
+/// Handle sampling during the active phase.
+fn sample_active(
+    state: &mut State,
+    distribution: &Distribution,
+    trial_id: u64,
+    history: &[CompletedTrial],
+) -> ParamValue {
+    // Check if we need to process completed trials and start a new generation
+    maybe_update_generation(state, history);
+
+    // Assign a candidate to this trial if not yet done
+    if !state.trial_progress.contains_key(&trial_id) {
+        assign_candidate(state, trial_id);
+    }
+
+    let progress = state.trial_progress.get_mut(&trial_id).unwrap();
+    let dim_idx = progress.next_dim;
+    progress.next_dim += 1;
+
+    // Safety check
+    if dim_idx >= state.dimensions.len() {
+        return sample_random(&mut state.rng, distribution);
+    }
+
+    let candidate = &state.candidates[progress.candidate_idx];
+    let dim_info = &state.dimensions[dim_idx];
+
+    if dim_info.is_continuous {
+        // Map from overall dimension index to continuous index
+        let ci = state.dimensions[..dim_idx]
+            .iter()
+            .filter(|d| d.is_continuous)
+            .count();
+        if ci < candidate.x.len() {
+            from_internal(candidate.x[ci], &dim_info.distribution)
+        } else {
+            sample_random(&mut state.rng, distribution)
+        }
+    } else {
+        // Categorical: use pre-sampled value
+        if let Some(&cat_idx) = candidate.categorical_values.get(&dim_idx) {
+            ParamValue::Categorical(cat_idx)
+        } else {
+            sample_random(&mut state.rng, distribution)
+        }
+    }
+}
+
+/// Assign a candidate to a trial.
+fn assign_candidate(state: &mut State, trial_id: u64) {
+    let candidate_idx = if state.assigned_count < state.candidates.len() {
+        let idx = state.assigned_count;
+        state.assigned_count += 1;
+        idx
+    } else {
+        // Overflow: generate an extra random candidate
+        let n_continuous = state.dimensions.iter().filter(|d| d.is_continuous).count();
+        let x: Vec<f64> = (0..n_continuous)
+            .map(|j| {
+                let bounds = continuous_dim_bounds(&state.dimensions, j);
+                if let Some(b) = bounds {
+                    sample_random_internal(&mut state.rng, b)
+                } else {
+                    0.0
+                }
+            })
+            .collect();
+        let mut categorical_values = HashMap::new();
+        for (dim_idx, dim) in state.dimensions.iter().enumerate() {
+            if !dim.is_continuous
+                && let Distribution::Categorical(cat) = &dim.distribution
+            {
+                categorical_values.insert(dim_idx, state.rng.usize(0..cat.n_choices));
+            }
+        }
+        state.candidates.push(Candidate {
+            x,
+            categorical_values,
+            target_idx: 0, // overflow candidates don't compete
+        });
+        let idx = state.candidates.len() - 1;
+        state.assigned_count = state.candidates.len();
+        idx
+    };
+
+    state.trial_progress.insert(
+        trial_id,
+        TrialProgress {
+            candidate_idx,
+            next_dim: 0,
+        },
+    );
+    state.generation_trial_ids.push(trial_id);
+}
+
+/// Check if we should process completed trials and start a new generation.
+fn maybe_update_generation(state: &mut State, history: &[CompletedTrial]) {
+    let pop_size = state.population_size;
+
+    // Only update when at least pop_size candidates have been assigned
+    if state.generation_trial_ids.len() < pop_size {
+        return;
+    }
+
+    // Check if the first pop_size trial IDs are all completed
+    let trial_ids: Vec<u64> = state
+        .generation_trial_ids
+        .iter()
+        .take(pop_size)
+        .copied()
+        .collect();
+    let history_map: HashMap<u64, f64> = history.iter().map(|t| (t.id, t.value)).collect();
+
+    let all_completed = trial_ids.iter().all(|id| history_map.contains_key(id));
+    if !all_completed {
+        return;
+    }
+
+    let n_continuous = state.dimensions.iter().filter(|d| d.is_continuous).count();
+
+    if state.initialized {
+        // Subsequent generations: selection
+        perform_selection(state, &trial_ids, &history_map);
+    } else {
+        // First generation: store as initial population
+        initialize_population(state, &trial_ids, &history_map, n_continuous);
+    }
+
+    // Generate next generation's trial vectors
+    state.candidates = if state.initialized && n_continuous > 0 {
+        generate_trial_vectors(state)
+    } else {
+        generate_initial_population(state)
+    };
+    state.assigned_count = 0;
+    state.generation_trial_ids.clear();
+    state.trial_progress.clear();
+}
+
+/// Initialize the population from the first generation's results.
+fn initialize_population(
+    state: &mut State,
+    trial_ids: &[u64],
+    history_map: &HashMap<u64, f64>,
+    _n_continuous: usize,
+) {
+    state.population.clear();
+    state.population_categorical.clear();
+    state.population_values.clear();
+
+    let mut best_value = f64::INFINITY;
+    let mut best_idx = 0;
+
+    for (i, &trial_id) in trial_ids.iter().enumerate() {
+        let progress = &state.trial_progress[&trial_id];
+        let candidate = &state.candidates[progress.candidate_idx];
+        let value = history_map[&trial_id];
+
+        state.population.push(candidate.x.clone());
+        state
+            .population_categorical
+            .push(candidate.categorical_values.clone());
+        state.population_values.push(value);
+
+        if value < best_value {
+            best_value = value;
+            best_idx = i;
+        }
+    }
+
+    state.best_idx = best_idx;
+    state.initialized = true;
+}
+
+/// Perform DE selection: replace parent if trial vector is better.
+fn perform_selection(state: &mut State, trial_ids: &[u64], history_map: &HashMap<u64, f64>) {
+    for &trial_id in trial_ids {
+        let progress = &state.trial_progress[&trial_id];
+        let candidate = &state.candidates[progress.candidate_idx];
+        let trial_value = history_map[&trial_id];
+        let target_idx = candidate.target_idx;
+
+        if target_idx < state.population_size && trial_value <= state.population_values[target_idx]
+        {
+            state.population[target_idx] = candidate.x.clone();
+            state.population_categorical[target_idx] = candidate.categorical_values.clone();
+            state.population_values[target_idx] = trial_value;
+        }
+    }
+
+    // Update best index
+    let mut best_value = f64::INFINITY;
+    let mut best_idx = 0;
+    for (i, &val) in state.population_values.iter().enumerate() {
+        if val < best_value {
+            best_value = val;
+            best_idx = i;
+        }
+    }
+    state.best_idx = best_idx;
+}
+
+#[cfg(test)]
+#[allow(clippy::cast_possible_truncation, clippy::cast_precision_loss)]
+mod tests {
+    use super::*;
+    use crate::distribution::FloatDistribution;
+
+    #[test]
+    fn test_de_sampler_basic_float() {
+        let sampler = DESampler::with_seed(42);
+        let dist = Distribution::Float(FloatDistribution {
+            low: -5.0,
+            high: 5.0,
+            log_scale: false,
+            step: None,
+        });
+
+        // Sample many values and check bounds
+        for i in 0..100 {
+            let value = sampler.sample(&dist, i, &[]);
+            if let ParamValue::Float(v) = value {
+                assert!(
+                    (-5.0..=5.0).contains(&v),
+                    "value {v} out of bounds at trial {i}"
+                );
+            } else {
+                panic!("Expected Float value");
+            }
+        }
+    }
+
+    #[test]
+    fn test_de_sampler_reproducibility() {
+        let dist = Distribution::Float(FloatDistribution {
+            low: 0.0,
+            high: 1.0,
+            log_scale: false,
+            step: None,
+        });
+
+        let sample_values = |seed: u64| {
+            let sampler = DESampler::with_seed(seed);
+            (0..20)
+                .map(|i| sampler.sample(&dist, i, &[]))
+                .collect::<Vec<_>>()
+        };
+
+        let v1 = sample_values(42);
+        let v2 = sample_values(42);
+        assert_eq!(v1, v2, "same seed should produce same results");
+
+        let v3 = sample_values(99);
+        assert_ne!(v1, v3, "different seeds should produce different results");
+    }
+
+    #[test]
+    fn test_de_strategy_default() {
+        assert!(matches!(DEStrategy::default(), DEStrategy::Rand1));
+    }
+
+    #[test]
+    fn test_builder_defaults() {
+        let builder = DESamplerBuilder::new();
+        assert!(builder.population_size.is_none());
+        assert!((builder.mutation_factor - 0.8).abs() < f64::EPSILON);
+        assert!((builder.crossover_rate - 0.9).abs() < f64::EPSILON);
+        assert!(matches!(builder.strategy, DEStrategy::Rand1));
+        assert!(builder.seed.is_none());
+    }
+}

src/sampler/genetic.rs

@@ -0,0 +1,537 @@
+//! Shared types and genetic operators for evolutionary multi-objective samplers.
+//!
+//! This module extracts common functionality used by NSGA-II, NSGA-III, and MOEA/D:
+//! candidate management, discovery/active phase logic, SBX crossover,
+//! polynomial mutation, and Das-Dennis reference point generation.
+
+use std::collections::HashMap;
+
+use crate::distribution::Distribution;
+use crate::multi_objective::MultiObjectiveTrial;
+use crate::param::ParamValue;
+use crate::rng_util;
+
+/// Describes a parameter dimension discovered during the first trial.
+#[derive(Clone, Debug)]
+pub(crate) struct DimensionInfo {
+    pub distribution: Distribution,
+}
+
+/// A candidate solution: one value per dimension.
+#[derive(Clone, Debug)]
+pub(crate) struct Candidate {
+    pub params: Vec<ParamValue>,
+}
+
+/// Tracks per-trial sampling progress (which candidate, which dimension next).
+#[derive(Clone, Debug)]
+pub(crate) struct TrialProgress {
+    pub candidate_idx: usize,
+    pub next_dim: usize,
+}
+
+/// Phase of an evolutionary sampler.
+pub(crate) enum Phase {
+    /// First trial reveals parameter dimensions.
+    Discovery,
+    /// Evolutionary optimisation.
+    Active,
+}
+
+/// Common state shared by all evolutionary multi-objective samplers.
+pub(crate) struct EvolutionaryState {
+    pub rng: fastrand::Rng,
+    pub phase: Phase,
+    pub dimensions: Vec<DimensionInfo>,
+    pub population_size: usize,
+    pub candidates: Vec<Candidate>,
+    pub trial_progress: HashMap<u64, TrialProgress>,
+    pub assigned_count: usize,
+    pub generation_trial_ids: Vec<u64>,
+    pub discovery_trial_id: Option<u64>,
+    pub generation: usize,
+}
+
+impl EvolutionaryState {
+    pub(crate) fn new(seed: Option<u64>) -> Self {
+        let rng = seed.map_or_else(fastrand::Rng::new, fastrand::Rng::with_seed);
+        Self {
+            rng,
+            phase: Phase::Discovery,
+            dimensions: Vec::new(),
+            population_size: 4,
+            candidates: Vec::new(),
+            trial_progress: HashMap::new(),
+            assigned_count: 0,
+            generation_trial_ids: Vec::new(),
+            discovery_trial_id: None,
+            generation: 0,
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Discovery phase helpers
+// ---------------------------------------------------------------------------
+
+/// Handle sampling during the discovery phase.
+///
+/// Returns `Some(value)` if the discovery phase handled the sample,
+/// or `None` if it transitioned to active phase and the caller should
+/// generate candidates and sample from them.
+pub(crate) fn sample_discovery(
+    evo: &mut EvolutionaryState,
+    distribution: &Distribution,
+    trial_id: u64,
+) -> Option<ParamValue> {
+    if let Some(prev_id) = evo.discovery_trial_id
+        && trial_id != prev_id
+    {
+        // A new trial arrived — transition to active phase
+        return None;
+    }
+
+    evo.discovery_trial_id = Some(trial_id);
+    evo.dimensions.push(DimensionInfo {
+        distribution: distribution.clone(),
+    });
+
+    Some(sample_random(&mut evo.rng, distribution))
+}
+
+/// Compute population size from dimensions and optional user override.
+#[allow(
+    clippy::cast_precision_loss,
+    clippy::cast_possible_truncation,
+    clippy::cast_sign_loss
+)]
+pub(crate) fn compute_population_size(
+    n_dims: usize,
+    user_pop_size: Option<usize>,
+    minimum: usize,
+) -> usize {
+    user_pop_size
+        .unwrap_or_else(|| (4.0 + 3.0 * (n_dims as f64).ln().max(0.0)).floor() as usize)
+        .max(minimum)
+}
+
+/// Transition from discovery to active phase.
+pub(crate) fn finalize_discovery(evo: &mut EvolutionaryState, user_pop_size: Option<usize>) {
+    evo.population_size = compute_population_size(evo.dimensions.len(), user_pop_size, 4);
+    evo.phase = Phase::Active;
+}
+
+/// Generate `population_size` random candidates.
+pub(crate) fn generate_random_candidates(evo: &mut EvolutionaryState) {
+    let pop = evo.population_size;
+    evo.candidates = (0..pop)
+        .map(|_| {
+            let params: Vec<ParamValue> = evo
+                .dimensions
+                .iter()
+                .map(|d| sample_random(&mut evo.rng, &d.distribution))
+                .collect();
+            Candidate { params }
+        })
+        .collect();
+    evo.assigned_count = 0;
+    evo.generation_trial_ids.clear();
+    evo.trial_progress.clear();
+}
+
+/// Assign a candidate to a trial and return the next dimension value.
+pub(crate) fn sample_from_candidate(evo: &mut EvolutionaryState, trial_id: u64) -> ParamValue {
+    if !evo.trial_progress.contains_key(&trial_id) {
+        let candidate_idx = if evo.assigned_count < evo.candidates.len() {
+            let idx = evo.assigned_count;
+            evo.assigned_count += 1;
+            idx
+        } else {
+            // Overflow: generate a random candidate
+            let params: Vec<ParamValue> = evo
+                .dimensions
+                .iter()
+                .map(|d| sample_random(&mut evo.rng, &d.distribution))
+                .collect();
+            evo.candidates.push(Candidate { params });
+            let idx = evo.candidates.len() - 1;
+            evo.assigned_count = evo.candidates.len();
+            idx
+        };
+
+        evo.trial_progress.insert(
+            trial_id,
+            TrialProgress {
+                candidate_idx,
+                next_dim: 0,
+            },
+        );
+        evo.generation_trial_ids.push(trial_id);
+    }
+
+    let progress = evo.trial_progress.get_mut(&trial_id).unwrap();
+    let dim_idx = progress.next_dim;
+    progress.next_dim += 1;
+
+    if dim_idx >= evo.dimensions.len() {
+        return sample_random(&mut evo.rng, &evo.dimensions.last().unwrap().distribution);
+    }
+
+    evo.candidates[progress.candidate_idx].params[dim_idx].clone()
+}
+
+/// Extract parameter values from a trial, ordered by dimension index.
+pub(crate) fn extract_trial_params(
+    trial: &MultiObjectiveTrial,
+    dimensions: &[DimensionInfo],
+    rng: &mut fastrand::Rng,
+) -> Vec<ParamValue> {
+    let mut param_pairs: Vec<_> = trial.params.iter().collect();
+    param_pairs.sort_by_key(|(id, _)| *id);
+
+    dimensions
+        .iter()
+        .enumerate()
+        .map(|(dim_idx, dim_info)| {
+            if dim_idx < param_pairs.len() {
+                param_pairs[dim_idx].1.clone()
+            } else {
+                sample_random(rng, &dim_info.distribution)
+            }
+        })
+        .collect()
+}
+
+/// Install new offspring as the next generation's candidates.
+pub(crate) fn advance_generation(evo: &mut EvolutionaryState, offspring: Vec<Candidate>) {
+    evo.candidates = offspring;
+    evo.assigned_count = 0;
+    evo.generation_trial_ids.clear();
+    evo.trial_progress.clear();
+    evo.generation += 1;
+}
+
+/// Check if the current generation is fully evaluated and return the
+/// evaluated trials if so.
+pub(crate) fn collect_evaluated_generation<'a>(
+    evo: &EvolutionaryState,
+    history: &'a [MultiObjectiveTrial],
+) -> Option<Vec<&'a MultiObjectiveTrial>> {
+    let pop_size = evo.population_size;
+
+    if evo.generation_trial_ids.len() < pop_size {
+        return None;
+    }
+
+    let gen_ids: Vec<u64> = evo
+        .generation_trial_ids
+        .iter()
+        .take(pop_size)
+        .copied()
+        .collect();
+    let history_map: HashMap<u64, &MultiObjectiveTrial> =
+        history.iter().map(|t| (t.id, t)).collect();
+
+    if !gen_ids.iter().all(|id| history_map.contains_key(id)) {
+        return None;
+    }
+
+    Some(
+        gen_ids
+            .iter()
+            .filter_map(|id| history_map.get(id).copied())
+            .collect(),
+    )
+}
+
+// ---------------------------------------------------------------------------
+// Genetic operators
+// ---------------------------------------------------------------------------
+
+/// SBX crossover for continuous params, uniform crossover for categorical.
+pub(crate) fn crossover(
+    rng: &mut fastrand::Rng,
+    parent1: &[ParamValue],
+    parent2: &[ParamValue],
+    dimensions: &[DimensionInfo],
+    crossover_prob: f64,
+    eta: f64,
+) -> (Vec<ParamValue>, Vec<ParamValue>) {
+    let n = parent1.len();
+    let mut child1 = parent1.to_vec();
+    let mut child2 = parent2.to_vec();
+
+    let u: f64 = rng_util::f64_range(rng, 0.0, 1.0);
+    if u > crossover_prob {
+        return (child1, child2);
+    }
+
+    for i in 0..n {
+        match (&parent1[i], &parent2[i], &dimensions[i].distribution) {
+            (ParamValue::Float(p1), ParamValue::Float(p2), Distribution::Float(d)) => {
+                if (p1 - p2).abs() < 1e-14 {
+                    continue;
+                }
+                let (c1, c2) = sbx_crossover_f64(rng, *p1, *p2, d.low, d.high, eta);
+                child1[i] = ParamValue::Float(c1);
+                child2[i] = ParamValue::Float(c2);
+            }
+            (ParamValue::Int(p1), ParamValue::Int(p2), Distribution::Int(d)) => {
+                if p1 == p2 {
+                    continue;
+                }
+                #[allow(clippy::cast_precision_loss)]
+                let (c1, c2) = sbx_crossover_f64(
+                    rng,
+                    *p1 as f64,
+                    *p2 as f64,
+                    d.low as f64,
+                    d.high as f64,
+                    eta,
+                );
+                #[allow(clippy::cast_possible_truncation)]
+                {
+                    child1[i] = ParamValue::Int((c1.round() as i64).clamp(d.low, d.high));
+                    child2[i] = ParamValue::Int((c2.round() as i64).clamp(d.low, d.high));
+                }
+            }
+            (ParamValue::Categorical(_), ParamValue::Categorical(_), _) => {
+                if rng_util::f64_range(rng, 0.0, 1.0) < 0.5 {
+                    core::mem::swap(&mut child1[i], &mut child2[i]);
+                }
+            }
+            _ => {}
+        }
+    }
+
+    (child1, child2)
+}
+
+/// SBX crossover for a single float dimension.
+pub(crate) fn sbx_crossover_f64(
+    rng: &mut fastrand::Rng,
+    p1: f64,
+    p2: f64,
+    low: f64,
+    high: f64,
+    eta: f64,
+) -> (f64, f64) {
+    let u: f64 = rng_util::f64_range(rng, 0.0, 1.0);
+
+    let beta = if u <= 0.5 {
+        (2.0 * u).powf(1.0 / (eta + 1.0))
+    } else {
+        (1.0 / (2.0 * (1.0 - u))).powf(1.0 / (eta + 1.0))
+    };
+
+    let c1 = 0.5 * ((1.0 + beta) * p1 + (1.0 - beta) * p2);
+    let c2 = 0.5 * ((1.0 - beta) * p1 + (1.0 + beta) * p2);
+
+    (c1.clamp(low, high), c2.clamp(low, high))
+}
+
+/// Polynomial mutation for each dimension.
+#[allow(clippy::cast_precision_loss)]
+pub(crate) fn mutate(
+    rng: &mut fastrand::Rng,
+    individual: &mut [ParamValue],
+    dimensions: &[DimensionInfo],
+    eta: f64,
+) {
+    let n = individual.len();
+    if n == 0 {
+        return;
+    }
+    let mutation_prob = 1.0 / n as f64;
+
+    for (i, value) in individual.iter_mut().enumerate() {
+        if rng_util::f64_range(rng, 0.0, 1.0) >= mutation_prob {
+            continue;
+        }
+
+        match (value, &dimensions[i].distribution) {
+            (v @ ParamValue::Float(_), Distribution::Float(d)) => {
+                let ParamValue::Float(x) = *v else {
+                    unreachable!();
+                };
+                let mutated = polynomial_mutation_f64(rng, x, d.low, d.high, eta);
+                *v = ParamValue::Float(mutated);
+            }
+            (v @ ParamValue::Int(_), Distribution::Int(d)) => {
+                let ParamValue::Int(x) = *v else {
+                    unreachable!();
+                };
+                #[allow(clippy::cast_possible_truncation)]
+                {
+                    let mutated =
+                        polynomial_mutation_f64(rng, x as f64, d.low as f64, d.high as f64, eta);
+                    *v = ParamValue::Int((mutated.round() as i64).clamp(d.low, d.high));
+                }
+            }
+            (v @ ParamValue::Categorical(_), Distribution::Categorical(d)) => {
+                *v = ParamValue::Categorical(rng.usize(0..d.n_choices));
+            }
+            _ => {}
+        }
+    }
+}
+
+/// Polynomial mutation for a single float value.
+pub(crate) fn polynomial_mutation_f64(
+    rng: &mut fastrand::Rng,
+    x: f64,
+    low: f64,
+    high: f64,
+    eta: f64,
+) -> f64 {
+    let u: f64 = rng_util::f64_range(rng, 0.0, 1.0);
+    let range = high - low;
+    if range <= 0.0 {
+        return x;
+    }
+
+    let delta1 = (x - low) / range;
+    let delta2 = (high - x) / range;
+
+    let delta_q = if u < 0.5 {
+        let xy = 1.0 - delta1;
+        let val = 2.0 * u + (1.0 - 2.0 * u) * xy.powf(eta + 1.0);
+        val.powf(1.0 / (eta + 1.0)) - 1.0
+    } else {
+        let xy = 1.0 - delta2;
+        let val = 2.0 * (1.0 - u) + 2.0 * (u - 0.5) * xy.powf(eta + 1.0);
+        1.0 - val.powf(1.0 / (eta + 1.0))
+    };
+
+    (x + delta_q * range).clamp(low, high)
+}
+
+pub(crate) use super::common::sample_random;
+
+// ---------------------------------------------------------------------------
+// Das-Dennis reference point generation
+// ---------------------------------------------------------------------------
+
+/// Generate Das-Dennis (simplex-lattice) reference points.
+///
+/// Returns `C(H + M - 1, M - 1)` uniformly spaced points on the
+/// `M`-dimensional unit simplex, where `M = n_objectives` and
+/// `H = divisions`.
+pub(crate) fn das_dennis(n_objectives: usize, divisions: usize) -> Vec<Vec<f64>> {
+    let mut points = Vec::new();
+    let mut point = vec![0.0_f64; n_objectives];
+    das_dennis_recursive(
+        n_objectives,
+        divisions,
+        0,
+        divisions,
+        &mut point,
+        &mut points,
+    );
+    points
+}
+
+#[allow(clippy::cast_precision_loss)]
+fn das_dennis_recursive(
+    n_objectives: usize,
+    divisions: usize,
+    depth: usize,
+    remaining: usize,
+    current: &mut Vec<f64>,
+    result: &mut Vec<Vec<f64>>,
+) {
+    if depth == n_objectives - 1 {
+        current[depth] = remaining as f64 / divisions as f64;
+        result.push(current.clone());
+        return;
+    }
+
+    for i in 0..=remaining {
+        current[depth] = i as f64 / divisions as f64;
+        das_dennis_recursive(
+            n_objectives,
+            divisions,
+            depth + 1,
+            remaining - i,
+            current,
+            result,
+        );
+    }
+}
+
+/// Choose the number of divisions for Das-Dennis to get close to a target
+/// population size.
+///
+/// The number of reference points is `C(H + M - 1, M - 1)`. This function
+/// finds the smallest `H` such that the number of points >= `target_pop`.
+pub(crate) fn auto_divisions(n_objectives: usize, target_pop: usize) -> usize {
+    let m = n_objectives;
+    for h in 1..200 {
+        let n_points = n_combinations(h + m - 1, m - 1);
+        if n_points >= target_pop {
+            return h;
+        }
+    }
+    12
+}
+
+/// Compute `C(n, k)` = n! / (k! * (n-k)!).
+fn n_combinations(n: usize, k: usize) -> usize {
+    if k > n {
+        return 0;
+    }
+    let k = k.min(n - k);
+    let mut result: usize = 1;
+    for i in 0..k {
+        result = result.saturating_mul(n - i) / (i + 1);
+    }
+    result
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_das_dennis_2d() {
+        let points = das_dennis(2, 4);
+        // C(4+1, 1) = 5 points
+        assert_eq!(points.len(), 5);
+        for p in &points {
+            let sum: f64 = p.iter().sum();
+            assert!((sum - 1.0).abs() < 1e-10, "point {p:?} doesn't sum to 1");
+        }
+    }
+
+    #[test]
+    fn test_das_dennis_3d() {
+        let points = das_dennis(3, 4);
+        // C(4+2, 2) = 15 points
+        assert_eq!(points.len(), 15);
+        for p in &points {
+            let sum: f64 = p.iter().sum();
+            assert!((sum - 1.0).abs() < 1e-10);
+        }
+    }
+
+    #[test]
+    fn test_auto_divisions() {
+        // For 2 objectives targeting 10 points: H=9 gives C(10,1)=10
+        let h = auto_divisions(2, 10);
+        let n = n_combinations(h + 1, 1);
+        assert!(n >= 10);
+
+        // For 3 objectives targeting ~91 points: H=12 gives C(14,2)=91
+        let h3 = auto_divisions(3, 91);
+        let n3 = n_combinations(h3 + 2, 2);
+        assert!(n3 >= 91);
+    }
+
+    #[test]
+    fn test_n_combinations() {
+        assert_eq!(n_combinations(5, 2), 10);
+        assert_eq!(n_combinations(4, 0), 1);
+        assert_eq!(n_combinations(4, 4), 1);
+        assert_eq!(n_combinations(6, 3), 20);
+    }
+}

src/sampler/gp.rs

@@ -0,0 +1,790 @@
+//! Gaussian Process (GP) sampler with Expected Improvement acquisition.
+//!
+//! A classical Bayesian optimization sampler that builds a Gaussian Process
+//! surrogate model with a **Matérn 5/2 kernel** (with ARD lengthscales) and
+//! selects the next trial by maximizing the **Expected Improvement (EI)**
+//! acquisition function. Best suited for small, expensive evaluations in
+//! low-dimensional continuous spaces (d ≤ 20).
+//!
+//! # Algorithm overview
+//!
+//! 1. **Startup phase** — the first `n_startup_trials` trials are sampled
+//!    uniformly at random to build an initial dataset.
+//! 2. **Fit GP** — training observations are standardized (zero mean, unit
+//!    variance) and a GP with Matérn 5/2 kernel is fitted via Cholesky
+//!    decomposition. ARD lengthscales are set to the per-dimension standard
+//!    deviation of the training inputs.
+//! 3. **Maximize EI** — `n_candidates` random points are evaluated under
+//!    the GP posterior and the point with the highest Expected Improvement
+//!    is returned as the next trial.
+//!
+//! The GP uses at most 100 training points (the most recent ones) to keep
+//! the O(n³) fitting cost manageable.
+//!
+//! # When to use
+//!
+//! - **Expensive objective functions** where every evaluation is costly
+//!   (e.g. physical experiments, large simulations). The GP surrogate
+//!   amortizes this cost by making fewer evaluations.
+//! - **Low-dimensional continuous spaces** — typically d ≤ 20. Beyond that,
+//!   the GP becomes unreliable and alternatives like
+//!   [`CmaEsSampler`](super::cma_es::CmaEsSampler) or
+//!   [`TpeSampler`](super::tpe::TpeSampler) are preferable.
+//! - **Smooth, low-noise objectives** — the GP assumes smoothness through
+//!   the Matérn 5/2 kernel. Very noisy objectives require increasing
+//!   `noise_variance`.
+//!
+//! Categorical parameters are sampled uniformly at random and do not
+//! participate in the GP model. If all parameters are categorical, the
+//! sampler falls back to pure random sampling.
+//!
+//! # Configuration
+//!
+//! | Option | Default | Description |
+//! |--------|---------|-------------|
+//! | `n_startup_trials` | 10 | Random trials before GP-guided sampling begins |
+//! | `n_candidates` | 1000 | Random candidates for EI maximization |
+//! | `noise_variance` | 1e-6 | Observation noise added to kernel diagonal |
+//! | `seed` | random | RNG seed for reproducibility |
+//!
+//! # Feature flag
+//!
+//! Requires the **`gp`** feature (adds the `nalgebra` dependency):
+//!
+//! ```toml
+//! [dependencies]
+//! optimizer = { version = "...", features = ["gp"] }
+//! ```
+//!
+//! # Examples
+//!
+//! ```
+//! use optimizer::sampler::gp::GpSampler;
+//! use optimizer::{Direction, Study};
+//!
+//! // Minimize an expensive function with GP-based Bayesian optimization
+//! let sampler = GpSampler::builder()
+//!     .n_startup_trials(5)
+//!     .n_candidates(500)
+//!     .seed(42)
+//!     .build();
+//!
+//! let mut study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+//! ```
+
+use std::collections::HashMap;
+
+use nalgebra::DMatrix;
+use parking_lot::Mutex;
+
+use crate::distribution::Distribution;
+use crate::param::ParamValue;
+use crate::rng_util;
+use crate::sampler::{CompletedTrial, Sampler};
+
+use super::common::{from_internal, internal_bounds, sample_random, to_internal};
+
+// ---------------------------------------------------------------------------
+// Public API
+// ---------------------------------------------------------------------------
+
+/// Gaussian Process sampler for Bayesian optimization.
+///
+/// Uses a GP surrogate with Matérn 5/2 kernel and Expected Improvement
+/// acquisition to guide sampling toward promising regions of the search
+/// space. Best suited for continuous (float/int) parameters in low
+/// dimensions (up to ~20).
+///
+/// # Examples
+///
+/// ```
+/// use optimizer::sampler::gp::GpSampler;
+/// use optimizer::{Direction, Study};
+///
+/// // Default configuration
+/// let study: Study<f64> = Study::with_sampler(Direction::Minimize, GpSampler::new());
+///
+/// // With seed for reproducibility
+/// let study: Study<f64> = Study::with_sampler(Direction::Minimize, GpSampler::with_seed(42));
+///
+/// // Custom configuration via builder
+/// let sampler = GpSampler::builder()
+///     .n_startup_trials(15)
+///     .n_candidates(2000)
+///     .noise_variance(1e-4)
+///     .seed(42)
+///     .build();
+/// let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+/// ```
+pub struct GpSampler {
+    state: Mutex<GpState>,
+}
+
+impl GpSampler {
+    /// Creates a new GP sampler with a random seed.
+    #[must_use]
+    pub fn new() -> Self {
+        Self {
+            state: Mutex::new(GpState::new(None, None, None, None)),
+        }
+    }
+
+    /// Creates a new GP sampler with a fixed seed for reproducibility.
+    #[must_use]
+    pub fn with_seed(seed: u64) -> Self {
+        Self {
+            state: Mutex::new(GpState::new(None, None, None, Some(seed))),
+        }
+    }
+
+    /// Creates a builder for configuring a `GpSampler`.
+    #[must_use]
+    pub fn builder() -> GpSamplerBuilder {
+        GpSamplerBuilder::new()
+    }
+}
+
+impl Default for GpSampler {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+/// Builder for configuring a [`GpSampler`].
+///
+/// All options have sensible defaults:
+/// - `n_startup_trials`: 10
+/// - `n_candidates`: 1000
+/// - `noise_variance`: 1e-6
+/// - `seed`: random
+///
+/// # Examples
+///
+/// ```
+/// use optimizer::sampler::gp::GpSamplerBuilder;
+///
+/// let sampler = GpSamplerBuilder::new()
+///     .n_startup_trials(15)
+///     .n_candidates(2000)
+///     .noise_variance(1e-4)
+///     .seed(42)
+///     .build();
+/// ```
+#[derive(Debug, Clone, Default)]
+pub struct GpSamplerBuilder {
+    n_startup_trials: Option<usize>,
+    n_candidates: Option<usize>,
+    noise_variance: Option<f64>,
+    seed: Option<u64>,
+}
+
+impl GpSamplerBuilder {
+    /// Creates a new builder with default settings.
+    #[must_use]
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    /// Sets the number of random trials before GP-guided sampling begins.
+    ///
+    /// Default: 10.
+    #[must_use]
+    pub fn n_startup_trials(mut self, n: usize) -> Self {
+        self.n_startup_trials = Some(n);
+        self
+    }
+
+    /// Sets the number of random candidate points for acquisition optimization.
+    ///
+    /// More candidates improve the quality of the acquisition maximum
+    /// at the cost of more GP predictions per trial.
+    ///
+    /// Default: 1000.
+    #[must_use]
+    pub fn n_candidates(mut self, n: usize) -> Self {
+        self.n_candidates = Some(n);
+        self
+    }
+
+    /// Sets the observation noise variance added to the kernel diagonal.
+    ///
+    /// Controls the assumed noise level. Larger values make the GP smoother.
+    ///
+    /// Default: 1e-6 (near-noiseless).
+    #[must_use]
+    pub fn noise_variance(mut self, v: f64) -> Self {
+        self.noise_variance = Some(v);
+        self
+    }
+
+    /// Sets the random seed for reproducibility.
+    #[must_use]
+    pub fn seed(mut self, seed: u64) -> Self {
+        self.seed = Some(seed);
+        self
+    }
+
+    /// Builds the configured [`GpSampler`].
+    #[must_use]
+    pub fn build(self) -> GpSampler {
+        GpSampler {
+            state: Mutex::new(GpState::new(
+                self.n_startup_trials,
+                self.n_candidates,
+                self.noise_variance,
+                self.seed,
+            )),
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Internal types
+// ---------------------------------------------------------------------------
+
+/// Default number of random startup trials before GP kicks in.
+const DEFAULT_N_STARTUP: usize = 10;
+/// Default number of candidate points for EI optimization.
+const DEFAULT_N_CANDIDATES: usize = 1000;
+/// Default observation noise variance.
+const DEFAULT_NOISE_VAR: f64 = 1e-6;
+
+/// Describes how a parameter dimension maps into the GP internal vector.
+#[derive(Clone, Debug)]
+struct DimensionInfo {
+    distribution: Distribution,
+    is_continuous: bool,
+    bounds: Option<(f64, f64)>,
+}
+
+/// Tracks per-trial sampling progress.
+#[derive(Clone, Debug)]
+struct TrialProgress {
+    /// The candidate values for each dimension.
+    values: Vec<ParamValue>,
+    /// Next dimension to return.
+    next_dim: usize,
+}
+
+/// Phase of the GP state machine.
+enum GpPhase {
+    /// Discovering the search space (first trial).
+    Discovery,
+    /// Steady-state sampling.
+    Active,
+}
+
+/// A fitted GP model ready for predictions.
+struct GpModel {
+    /// Cholesky factor L of K + σ²I.
+    cholesky: nalgebra::linalg::Cholesky<f64, nalgebra::Dyn>,
+    /// α = (K + σ²I)^{-1} y.
+    alpha: nalgebra::DVector<f64>,
+    /// Training inputs (each row is a data point, normalized to [0, 1]).
+    x_train: Vec<Vec<f64>>,
+    /// ARD lengthscales per dimension.
+    lengthscales: Vec<f64>,
+    /// Signal variance.
+    signal_var: f64,
+    /// Mean of original y values (for un-standardization, unused but kept for diagnostics).
+    _y_mean: f64,
+    /// Std dev of original y values (unused but kept for diagnostics).
+    _y_std: f64,
+    /// Best observed (standardized) y.
+    f_best: f64,
+}
+
+/// Top-level mutable state behind the `Mutex`.
+struct GpState {
+    rng: fastrand::Rng,
+    n_startup_trials: usize,
+    n_candidates: usize,
+    noise_variance: f64,
+    phase: GpPhase,
+    dimensions: Vec<DimensionInfo>,
+    trial_progress: HashMap<u64, TrialProgress>,
+    discovery_trial_id: Option<u64>,
+}
+
+impl GpState {
+    fn new(
+        n_startup: Option<usize>,
+        n_candidates: Option<usize>,
+        noise_var: Option<f64>,
+        seed: Option<u64>,
+    ) -> Self {
+        let rng = seed.map_or_else(fastrand::Rng::new, fastrand::Rng::with_seed);
+        Self {
+            rng,
+            n_startup_trials: n_startup.unwrap_or(DEFAULT_N_STARTUP),
+            n_candidates: n_candidates.unwrap_or(DEFAULT_N_CANDIDATES),
+            noise_variance: noise_var.unwrap_or(DEFAULT_NOISE_VAR),
+            phase: GpPhase::Discovery,
+            dimensions: Vec::new(),
+            trial_progress: HashMap::new(),
+            discovery_trial_id: None,
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Matérn 5/2 kernel
+// ---------------------------------------------------------------------------
+
+/// Matérn 5/2 kernel with ARD lengthscales.
+///
+/// `k(x1, x2) = σ² (1 + √5 r + 5/3 r²) exp(-√5 r)`
+/// where `r = sqrt(Σ ((x1_i - x2_i) / l_i)²)`
+fn matern52(x1: &[f64], x2: &[f64], lengthscales: &[f64], signal_var: f64) -> f64 {
+    let mut r_sq = 0.0;
+    for i in 0..x1.len() {
+        let diff = (x1[i] - x2[i]) / lengthscales[i];
+        r_sq += diff * diff;
+    }
+    let r = r_sq.sqrt();
+    let sqrt5_r = SQRT_5 * r;
+    signal_var * (1.0 + sqrt5_r + 5.0 / 3.0 * r_sq) * (-sqrt5_r).exp()
+}
+
+/// Build the kernel matrix `K + σ²I`.
+fn kernel_matrix(
+    x: &[Vec<f64>],
+    lengthscales: &[f64],
+    signal_var: f64,
+    noise_var: f64,
+) -> DMatrix<f64> {
+    let n = x.len();
+    DMatrix::from_fn(n, n, |i, j| {
+        let k = matern52(&x[i], &x[j], lengthscales, signal_var);
+        if i == j { k + noise_var } else { k }
+    })
+}
+
+/// Compute the kernel vector k(x*, X) for a test point.
+fn kernel_vector(
+    x_star: &[f64],
+    x_train: &[Vec<f64>],
+    lengthscales: &[f64],
+    signal_var: f64,
+) -> nalgebra::DVector<f64> {
+    nalgebra::DVector::from_fn(x_train.len(), |i, _| {
+        matern52(x_star, &x_train[i], lengthscales, signal_var)
+    })
+}
+
+/// Precomputed √5 constant.
+const SQRT_5: f64 = 2.236_213_562_373_095;
+
+// ---------------------------------------------------------------------------
+// GP fitting and prediction
+// ---------------------------------------------------------------------------
+
+/// Fit a GP model to the training data.
+///
+/// Returns `None` if fitting fails (e.g. Cholesky decomposition failure).
+#[allow(clippy::cast_precision_loss)]
+fn fit_gp(x_train: &[Vec<f64>], y_train: &[f64], noise_var: f64) -> Option<GpModel> {
+    let n = y_train.len();
+    if n == 0 {
+        return None;
+    }
+
+    // Standardize y
+    let y_mean = y_train.iter().sum::<f64>() / n as f64;
+    let y_var = if n > 1 {
+        y_train.iter().map(|&y| (y - y_mean).powi(2)).sum::<f64>() / (n - 1) as f64
+    } else {
+        1.0
+    };
+    let y_std = y_var.sqrt().max(1e-10);
+    let y_standardized: Vec<f64> = y_train.iter().map(|&y| (y - y_mean) / y_std).collect();
+
+    let f_best = y_standardized.iter().copied().fold(f64::INFINITY, f64::min);
+
+    // ARD lengthscales: per-dimension std dev of training X, clamped
+    let d = if x_train.is_empty() {
+        0
+    } else {
+        x_train[0].len()
+    };
+    let lengthscales: Vec<f64> = (0..d)
+        .map(|j| {
+            let vals: Vec<f64> = x_train.iter().map(|x| x[j]).collect();
+            let mean_j = vals.iter().sum::<f64>() / n as f64;
+            let var_j = vals.iter().map(|&v| (v - mean_j).powi(2)).sum::<f64>() / n as f64;
+            var_j.sqrt().max(0.01)
+        })
+        .collect();
+
+    // Signal variance = 1.0 (data is standardized)
+    let signal_var = 1.0;
+
+    let k = kernel_matrix(x_train, &lengthscales, signal_var, noise_var);
+    let cholesky = nalgebra::linalg::Cholesky::new(k)?;
+
+    // α = (K + σ²I)^{-1} y
+    let y_vec = nalgebra::DVector::from_column_slice(&y_standardized);
+    let alpha = cholesky.solve(&y_vec);
+
+    Some(GpModel {
+        cholesky,
+        alpha,
+        x_train: x_train.to_vec(),
+        lengthscales,
+        signal_var,
+        _y_mean: y_mean,
+        _y_std: y_std,
+        f_best,
+    })
+}
+
+/// Predict mean and standard deviation at a test point.
+fn predict(model: &GpModel, x: &[f64]) -> (f64, f64) {
+    let k_star = kernel_vector(x, &model.x_train, &model.lengthscales, model.signal_var);
+
+    // Mean: k*^T α
+    let mean = k_star.dot(&model.alpha);
+
+    // Variance: k(x*, x*) - k*^T (K + σ²I)^{-1} k*
+    let k_self = model.signal_var;
+    let v = model.cholesky.solve(&k_star);
+    let var = (k_self - k_star.dot(&v)).max(0.0);
+
+    (mean, var.sqrt())
+}
+
+// ---------------------------------------------------------------------------
+// Normal distribution helpers (Abramowitz-Stegun approximation)
+// ---------------------------------------------------------------------------
+
+/// Standard normal PDF.
+fn norm_pdf(x: f64) -> f64 {
+    const INV_SQRT_2PI: f64 = 0.398_942_280_401_432_7;
+    INV_SQRT_2PI * (-0.5 * x * x).exp()
+}
+
+/// Standard normal CDF (Abramowitz-Stegun rational approximation).
+fn norm_cdf(x: f64) -> f64 {
+    // Hart approximation (higher precision than basic A&S)
+    if x < -8.0 {
+        return 0.0;
+    }
+    if x > 8.0 {
+        return 1.0;
+    }
+
+    let abs_x = x.abs();
+    let t = 1.0 / (1.0 + 0.231_641_9 * abs_x);
+    let t2 = t * t;
+    let t3 = t2 * t;
+    let t4 = t3 * t;
+    let t5 = t4 * t;
+
+    let poly = 0.319_381_530 * t - 0.356_563_782 * t2 + 1.781_477_937 * t3 - 1.821_255_978 * t4
+        + 1.330_274_429 * t5;
+    let pdf = norm_pdf(abs_x);
+    let cdf = 1.0 - pdf * poly;
+
+    if x >= 0.0 { cdf } else { 1.0 - cdf }
+}
+
+// ---------------------------------------------------------------------------
+// Expected Improvement
+// ---------------------------------------------------------------------------
+
+/// Compute Expected Improvement at a point.
+///
+/// `EI(x) = (f_best - mean) Φ(z) + std φ(z)`
+/// where `z = (f_best - mean) / std`
+fn expected_improvement(mean: f64, std: f64, f_best: f64) -> f64 {
+    if std < 1e-12 {
+        return (f_best - mean).max(0.0);
+    }
+    let z = (f_best - mean) / std;
+    let improvement = (f_best - mean) * norm_cdf(z) + std * norm_pdf(z);
+    improvement.max(0.0)
+}
+
+// ---------------------------------------------------------------------------
+// Acquisition optimization
+// ---------------------------------------------------------------------------
+
+/// Find the point in [0, 1]^d that maximizes EI via multi-start random search.
+fn optimize_acquisition(
+    model: &GpModel,
+    n_dims: usize,
+    n_candidates: usize,
+    rng: &mut fastrand::Rng,
+) -> Vec<f64> {
+    let mut best_ei = f64::NEG_INFINITY;
+    let mut best_x = vec![0.5; n_dims];
+
+    for _ in 0..n_candidates {
+        let x: Vec<f64> = (0..n_dims)
+            .map(|_| rng_util::f64_range(rng, 0.0, 1.0))
+            .collect();
+        let (mean, std) = predict(model, &x);
+        let ei = expected_improvement(mean, std, model.f_best);
+        if ei > best_ei {
+            best_ei = ei;
+            best_x = x;
+        }
+    }
+
+    best_x
+}
+
+// ---------------------------------------------------------------------------
+// Data preprocessing helpers
+// ---------------------------------------------------------------------------
+
+/// Convert an internal-space value to normalized [0, 1] using bounds.
+fn to_normalized(value: f64, lo: f64, hi: f64) -> f64 {
+    if (hi - lo).abs() < 1e-15 {
+        0.5
+    } else {
+        (value - lo) / (hi - lo)
+    }
+}
+
+/// Convert a normalized [0, 1] value back to internal space.
+fn from_normalized(value: f64, lo: f64, hi: f64) -> f64 {
+    lo + value * (hi - lo)
+}
+
+// ---------------------------------------------------------------------------
+// Extract training data from history
+// ---------------------------------------------------------------------------
+
+/// Maximum number of training points to use for the GP.
+/// Caps computational cost at O(`MAX_TRAIN_POINTS`^3) per trial.
+const MAX_TRAIN_POINTS: usize = 100;
+
+/// Establish a deterministic mapping from dimension index to `ParamId`
+/// using the first trial in history.
+///
+/// Matches dimensions to params by distribution equality, consuming
+/// matched params to correctly handle duplicate distributions.
+fn establish_param_mapping(
+    trial: &CompletedTrial,
+    dimensions: &[DimensionInfo],
+) -> Vec<Option<crate::parameter::ParamId>> {
+    use crate::parameter::ParamId;
+
+    let mut available: Vec<(ParamId, &Distribution)> =
+        trial.distributions.iter().map(|(id, d)| (*id, d)).collect();
+    // Sort for deterministic matching order
+    available.sort_by_key(|(id, _)| *id);
+
+    let mut mapping = Vec::with_capacity(dimensions.len());
+    for dim in dimensions {
+        let pos = available.iter().position(|(_, d)| **d == dim.distribution);
+        if let Some(pos) = pos {
+            mapping.push(Some(available.remove(pos).0));
+        } else {
+            mapping.push(None);
+        }
+    }
+    mapping
+}
+
+/// Build normalized training data from completed trials.
+///
+/// Returns `(x_train, y_train)` where x values are normalized to [0, 1]
+/// per dimension using the bounds from `dimensions`. Only continuous
+/// dimensions are included. Uses at most [`MAX_TRAIN_POINTS`] most recent
+/// trials.
+#[allow(clippy::cast_precision_loss)]
+fn build_training_data(
+    history: &[CompletedTrial],
+    dimensions: &[DimensionInfo],
+) -> (Vec<Vec<f64>>, Vec<f64>) {
+    if history.is_empty() {
+        return (Vec::new(), Vec::new());
+    }
+
+    // Use only the most recent trials to cap GP fitting cost
+    let start = history.len().saturating_sub(MAX_TRAIN_POINTS);
+    let recent = &history[start..];
+
+    // Establish dimension → ParamId mapping from the first trial
+    let param_mapping = establish_param_mapping(&recent[0], dimensions);
+
+    let continuous_indices: Vec<usize> = dimensions
+        .iter()
+        .enumerate()
+        .filter(|(_, d)| d.is_continuous)
+        .map(|(i, _)| i)
+        .collect();
+
+    let mut x_train = Vec::with_capacity(recent.len());
+    let mut y_train = Vec::with_capacity(recent.len());
+
+    for trial in recent {
+        let mut x_row = Vec::with_capacity(continuous_indices.len());
+        let mut valid = true;
+
+        for &dim_idx in &continuous_indices {
+            let dim_info = &dimensions[dim_idx];
+            if let Some(param_id) = param_mapping[dim_idx] {
+                if let Some(param_val) = trial.params.get(&param_id) {
+                    let internal = to_internal(param_val, &dim_info.distribution);
+                    let (lo, hi) = dim_info.bounds.unwrap_or((0.0, 1.0));
+                    x_row.push(to_normalized(internal, lo, hi));
+                } else {
+                    valid = false;
+                    break;
+                }
+            } else {
+                valid = false;
+                break;
+            }
+        }
+
+        if valid && x_row.len() == continuous_indices.len() {
+            x_train.push(x_row);
+            y_train.push(trial.value);
+        }
+    }
+
+    (x_train, y_train)
+}
+
+// ---------------------------------------------------------------------------
+// Sampler trait implementation
+// ---------------------------------------------------------------------------
+
+impl Sampler for GpSampler {
+    #[allow(clippy::cast_precision_loss, clippy::cast_possible_truncation)]
+    fn sample(
+        &self,
+        distribution: &Distribution,
+        trial_id: u64,
+        history: &[CompletedTrial],
+    ) -> ParamValue {
+        let mut state = self.state.lock();
+
+        match &state.phase {
+            GpPhase::Discovery => sample_discovery(&mut state, distribution, trial_id),
+            GpPhase::Active => sample_active(&mut state, distribution, trial_id, history),
+        }
+    }
+}
+
+/// Handle sampling during the discovery phase.
+fn sample_discovery(state: &mut GpState, distribution: &Distribution, trial_id: u64) -> ParamValue {
+    // A new trial_id means discovery is done
+    if let Some(prev_id) = state.discovery_trial_id
+        && trial_id != prev_id
+    {
+        finalize_discovery(state);
+        return sample_active(state, distribution, trial_id, &[]);
+    }
+
+    state.discovery_trial_id = Some(trial_id);
+
+    let is_continuous = !matches!(distribution, Distribution::Categorical(_));
+    let bounds = internal_bounds(distribution);
+    state.dimensions.push(DimensionInfo {
+        distribution: distribution.clone(),
+        is_continuous,
+        bounds,
+    });
+
+    sample_random(&mut state.rng, distribution)
+}
+
+/// Finalize discovery and transition to the active phase.
+fn finalize_discovery(state: &mut GpState) {
+    state.phase = GpPhase::Active;
+    state.trial_progress.clear();
+}
+
+/// Handle sampling during the active phase.
+fn sample_active(
+    state: &mut GpState,
+    distribution: &Distribution,
+    trial_id: u64,
+    history: &[CompletedTrial],
+) -> ParamValue {
+    // If this trial already has progress, return the next pre-computed value
+    if let Some(progress) = state.trial_progress.get_mut(&trial_id) {
+        let dim_idx = progress.next_dim;
+        progress.next_dim += 1;
+        if dim_idx < progress.values.len() {
+            return progress.values[dim_idx].clone();
+        }
+        // Extra dimension not seen during discovery
+        return sample_random(&mut state.rng, distribution);
+    }
+
+    // New trial: compute all dimension values at once
+    let n_continuous = state.dimensions.iter().filter(|d| d.is_continuous).count();
+    let use_gp = n_continuous > 0 && history.len() >= state.n_startup_trials;
+
+    let values = if use_gp {
+        compute_gp_candidate(state, history)
+    } else {
+        compute_random_candidate(state)
+    };
+
+    let first_value = values
+        .first()
+        .cloned()
+        .unwrap_or_else(|| sample_random(&mut state.rng, distribution));
+
+    state.trial_progress.insert(
+        trial_id,
+        TrialProgress {
+            values,
+            next_dim: 1,
+        },
+    );
+
+    first_value
+}
+
+/// Compute a candidate using the GP model.
+fn compute_gp_candidate(state: &mut GpState, history: &[CompletedTrial]) -> Vec<ParamValue> {
+    let (x_train, y_train) = build_training_data(history, &state.dimensions);
+
+    // Try to fit GP; fall back to random if it fails
+    let model = fit_gp(&x_train, &y_train, state.noise_variance);
+
+    let n_continuous = state.dimensions.iter().filter(|d| d.is_continuous).count();
+
+    let normalized_candidate = if let Some(ref model) = model {
+        optimize_acquisition(model, n_continuous, state.n_candidates, &mut state.rng)
+    } else {
+        // GP fitting failed; use random
+        (0..n_continuous)
+            .map(|_| rng_util::f64_range(&mut state.rng, 0.0, 1.0))
+            .collect()
+    };
+
+    // Convert normalized candidate back to parameter values
+    let mut values = Vec::with_capacity(state.dimensions.len());
+    let mut ci = 0; // continuous dimension index
+
+    for dim in &state.dimensions {
+        if dim.is_continuous {
+            let (lo, hi) = dim.bounds.unwrap_or((0.0, 1.0));
+            let internal_val = from_normalized(normalized_candidate[ci], lo, hi);
+            values.push(from_internal(internal_val, &dim.distribution));
+            ci += 1;
+        } else {
+            values.push(sample_random(&mut state.rng, &dim.distribution));
+        }
+    }
+
+    values
+}
+
+/// Compute a random candidate for all dimensions.
+fn compute_random_candidate(state: &mut GpState) -> Vec<ParamValue> {
+    state
+        .dimensions
+        .iter()
+        .map(|dim| sample_random(&mut state.rng, &dim.distribution))
+        .collect()
+}

src/sampler/grid.rs

@@ -1,7 +1,39 @@
-//! Grid search sampler implementation.
+//! Grid search sampler — exhaustive evaluation of discretized parameter spaces.
 //!
-//! `GridSearchSampler` performs exhaustive grid search over the parameter space,
-//! systematically evaluating all combinations of discretized parameter values.
+//! [`GridSampler`] divides each parameter range into a fixed number of
+//! evenly spaced points (or uses the explicit step size when defined) and
+//! evaluates them sequentially. This guarantees complete coverage of the
+//! search grid at the cost of scaling exponentially with the number of
+//! parameters.
+//!
+//! # When to use
+//!
+//! - **Small, discrete spaces** — when you have a handful of categorical or
+//!   integer parameters and want to evaluate every combination.
+//! - **Reproducibility** — grid search is fully deterministic with no random
+//!   component.
+//! - **Benchmarking** — compare grid search results against adaptive samplers
+//!   to measure their benefit.
+//!
+//! Avoid grid search for high-dimensional or large continuous spaces;
+//! prefer [`TpeSampler`](super::tpe::TpeSampler) or
+//! [`RandomSampler`](super::random::RandomSampler) instead.
+//!
+//! # Configuration
+//!
+//! | Option | Default | Description |
+//! |---|---|---|
+//! | `n_points_per_param` | 10 | Points per continuous parameter (ignored when `step` is set) |
+//!
+//! # Example
+//!
+//! ```
+//! use optimizer::prelude::*;
+//! use optimizer::sampler::grid::GridSampler;
+//!
+//! let sampler = GridSampler::builder().n_points_per_param(5).build();
+//! let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+//! ```
 
 use std::collections::HashMap;
 
@@ -295,48 +327,48 @@ struct GridState {
     grids: HashMap<String, CachedGrid>,
 }
 
-/// A grid search sampler that exhaustively evaluates all grid points.
+/// Exhaustive grid search sampler.
 ///
-/// `GridSearchSampler` divides the parameter space into a grid and systematically
-/// samples each point. This is useful when you want to evaluate all combinations
-/// of parameter values, especially for discrete or small parameter spaces.
+/// Divide each parameter range into evenly spaced points and evaluate them
+/// sequentially. When a parameter has an explicit `step` size, the step grid
+/// is used instead of auto-discretization.
 ///
-/// # Grid Exhaustion
+/// Grid state is tracked **per distribution key** (bounds + step + log-scale).
+/// Parameters with identical distributions share the same grid counter, so
+/// use distinct ranges when multiple parameters span the same domain.
 ///
-/// The sampler tracks its position in the grid for each distribution independently.
-/// When all grid points for a distribution have been sampled, subsequent calls to
-/// `sample()` for that distribution will **panic** with the message:
-/// `"GridSearchSampler: all grid points exhausted"`.
+/// # Grid exhaustion
 ///
-/// To avoid panics, use [`is_exhausted()`](Self::is_exhausted) to check if all
-/// points have been sampled before calling `sample()`. You can also use
-/// [`grid_size()`](Self::grid_size) to determine the total number of grid points
-/// that will be sampled.
+/// When all grid points for a distribution have been sampled, the next
+/// `sample()` call for that distribution **panics**. Use
+/// [`is_exhausted()`](Self::is_exhausted) to check before sampling, or
+/// set `n_points_per_param` high enough to cover the planned number of
+/// trials.
 ///
-/// # Thread Safety
+/// # Thread safety
 ///
-/// `GridSearchSampler` is thread-safe (`Send + Sync`) and uses internal locking
-/// to ensure safe concurrent access to grid state.
+/// `GridSearchSampler` is `Send + Sync` and uses internal locking for
+/// safe concurrent access.
 ///
 /// # Examples
 ///
 /// ```
-/// use optimizer::sampler::grid::GridSearchSampler;
+/// use optimizer::sampler::grid::GridSampler;
 ///
-/// // Create with default settings (10 points per parameter)
-/// let sampler = GridSearchSampler::new();
+/// // Default: 10 points per parameter
+/// let sampler = GridSampler::new();
 ///
-/// // Create with custom settings using the builder
-/// let sampler = GridSearchSampler::builder().n_points_per_param(20).build();
+/// // Custom grid density
+/// let sampler = GridSampler::builder().n_points_per_param(20).build();
 /// ```
-pub struct GridSearchSampler {
+pub struct GridSampler {
     /// Number of grid points per parameter (used when auto-discretizing).
     n_points_per_param: usize,
     /// Thread-safe internal state for tracking grid positions.
     state: Mutex<GridState>,
 }
 
-impl GridSearchSampler {
+impl GridSampler {
     /// Creates a new grid search sampler with default settings.
     ///
     /// Default settings:
@@ -354,9 +386,9 @@ impl GridSearchSampler {
     /// # Examples
     ///
     /// ```
-    /// use optimizer::sampler::grid::GridSearchSampler;
+    /// use optimizer::sampler::grid::GridSampler;
     ///
-    /// let sampler = GridSearchSampler::builder().n_points_per_param(20).build();
+    /// let sampler = GridSampler::builder().n_points_per_param(20).build();
     /// ```
     #[must_use]
     pub fn builder() -> GridSearchSamplerBuilder {
@@ -364,13 +396,13 @@ impl GridSearchSampler {
     }
 }
 
-impl Default for GridSearchSampler {
+impl Default for GridSampler {
     fn default() -> Self {
         Self::new()
     }
 }
 
-impl GridSearchSampler {
+impl GridSampler {
     /// Returns `true` if all grid points for all tracked distributions have been sampled.
     ///
     /// A distribution is considered exhausted when its `current_index` equals the number
@@ -384,9 +416,9 @@ impl GridSearchSampler {
     /// # Examples
     ///
     /// ```
-    /// use optimizer::sampler::grid::GridSearchSampler;
+    /// use optimizer::sampler::grid::GridSampler;
     ///
-    /// let sampler = GridSearchSampler::new();
+    /// let sampler = GridSampler::new();
     /// // Initially exhausted (no distributions tracked yet)
     /// assert!(sampler.is_exhausted());
     /// ```
@@ -414,9 +446,9 @@ impl GridSearchSampler {
     /// # Examples
     ///
     /// ```
-    /// use optimizer::sampler::grid::GridSearchSampler;
+    /// use optimizer::sampler::grid::GridSampler;
     ///
-    /// let sampler = GridSearchSampler::new();
+    /// let sampler = GridSampler::new();
     /// // No distributions tracked yet
     /// assert_eq!(sampler.grid_size(), 0);
     /// ```
@@ -427,7 +459,7 @@ impl GridSearchSampler {
     }
 }
 
-/// Builder for configuring a [`GridSearchSampler`].
+/// Builder for configuring a [`GridSampler`].
 ///
 /// # Examples
 ///
@@ -479,7 +511,7 @@ impl GridSearchSamplerBuilder {
         self
     }
 
-    /// Builds the configured [`GridSearchSampler`].
+    /// Builds the configured [`GridSampler`].
     ///
     /// # Examples
     ///
@@ -491,8 +523,8 @@ impl GridSearchSamplerBuilder {
     ///     .build();
     /// ```
     #[must_use]
-    pub fn build(self) -> GridSearchSampler {
-        GridSearchSampler {
+    pub fn build(self) -> GridSampler {
+        GridSampler {
             n_points_per_param: self.n_points_per_param,
             state: Mutex::new(GridState::default()),
         }
@@ -534,7 +566,7 @@ fn distribution_key(dist: &Distribution) -> String {
     }
 }
 
-impl Sampler for GridSearchSampler {
+impl Sampler for GridSampler {
     fn sample(
         &self,
         distribution: &Distribution,
@@ -569,7 +601,7 @@ impl Sampler for GridSearchSampler {
         // Check if all points have been exhausted
         assert!(
             cached.current_index < cached.points.len(),
-            "GridSearchSampler: all grid points exhausted"
+            "GridSampler: all grid points exhausted"
         );
 
         // Get the current grid point and advance the index
@@ -844,7 +876,7 @@ mod tests {
 
     #[test]
     fn test_sampler_exhausts_after_expected_samples() {
-        let sampler = GridSearchSampler::new();
+        let sampler = GridSampler::new();
         let dist = Distribution::Categorical(CategoricalDistribution { n_choices: 3 });
 
         // Sample all 3 points
@@ -858,7 +890,7 @@ mod tests {
 
     #[test]
     fn test_sampler_exhaustion_with_int_distribution() {
-        let sampler = GridSearchSampler::builder().n_points_per_param(5).build();
+        let sampler = GridSampler::builder().n_points_per_param(5).build();
         let dist = Distribution::Int(IntDistribution {
             low: 0,
             high: 100,
@@ -876,9 +908,9 @@ mod tests {
     }
 
     #[test]
-    #[should_panic(expected = "GridSearchSampler: all grid points exhausted")]
+    #[should_panic(expected = "GridSampler: all grid points exhausted")]
     fn test_sampler_panics_after_exhaustion() {
-        let sampler = GridSearchSampler::new();
+        let sampler = GridSampler::new();
         let dist = Distribution::Categorical(CategoricalDistribution { n_choices: 2 });
 
         // Sample all 2 points
@@ -893,14 +925,14 @@ mod tests {
 
     #[test]
     fn test_is_exhausted_before_sampling() {
-        let sampler = GridSearchSampler::new();
+        let sampler = GridSampler::new();
         // Newly created sampler is vacuously exhausted (no distributions tracked)
         assert!(sampler.is_exhausted());
     }
 
     #[test]
     fn test_is_exhausted_during_sampling() {
-        let sampler = GridSearchSampler::new();
+        let sampler = GridSampler::new();
         let dist = Distribution::Categorical(CategoricalDistribution { n_choices: 3 });
 
         // After first sample, not exhausted
@@ -918,7 +950,7 @@ mod tests {
 
     #[test]
     fn test_is_exhausted_multiple_distributions() {
-        let sampler = GridSearchSampler::new();
+        let sampler = GridSampler::new();
         // Use different n_choices so they have different distribution keys
         let dist1 = Distribution::Categorical(CategoricalDistribution { n_choices: 2 });
         let dist2 = Distribution::Categorical(CategoricalDistribution { n_choices: 3 });
@@ -944,7 +976,7 @@ mod tests {
 
     #[test]
     fn test_builder_default() {
-        let sampler = GridSearchSampler::builder().build();
+        let sampler = GridSampler::builder().build();
         let dist = Distribution::Float(FloatDistribution {
             low: 0.0,
             high: 1.0,
@@ -961,7 +993,7 @@ mod tests {
 
     #[test]
     fn test_builder_custom_n_points() {
-        let sampler = GridSearchSampler::builder().n_points_per_param(3).build();
+        let sampler = GridSampler::builder().n_points_per_param(3).build();
         let dist = Distribution::Float(FloatDistribution {
             low: 0.0,
             high: 1.0,
@@ -979,7 +1011,7 @@ mod tests {
 
     #[test]
     fn test_new_default() {
-        let sampler = GridSearchSampler::new();
+        let sampler = GridSampler::new();
         let dist = Distribution::Float(FloatDistribution {
             low: 0.0,
             high: 1.0,
@@ -999,8 +1031,8 @@ mod tests {
     #[test]
     fn test_reproducibility_same_grid_order() {
         // Two samplers with the same configuration should produce the same grid order
-        let sampler1 = GridSearchSampler::builder().n_points_per_param(5).build();
-        let sampler2 = GridSearchSampler::builder().n_points_per_param(5).build();
+        let sampler1 = GridSampler::builder().n_points_per_param(5).build();
+        let sampler2 = GridSampler::builder().n_points_per_param(5).build();
 
         let dist = Distribution::Float(FloatDistribution {
             low: 0.0,
@@ -1019,8 +1051,8 @@ mod tests {
 
     #[test]
     fn test_reproducibility_int_distribution() {
-        let sampler1 = GridSearchSampler::new();
-        let sampler2 = GridSearchSampler::new();
+        let sampler1 = GridSampler::new();
+        let sampler2 = GridSampler::new();
 
         let dist = Distribution::Int(IntDistribution {
             low: 0,
@@ -1041,8 +1073,8 @@ mod tests {
 
     #[test]
     fn test_reproducibility_categorical() {
-        let sampler1 = GridSearchSampler::new();
-        let sampler2 = GridSearchSampler::new();
+        let sampler1 = GridSampler::new();
+        let sampler2 = GridSampler::new();
 
         let dist = Distribution::Categorical(CategoricalDistribution { n_choices: 4 });
 
@@ -1059,13 +1091,13 @@ mod tests {
 
     #[test]
     fn test_grid_size_empty() {
-        let sampler = GridSearchSampler::new();
+        let sampler = GridSampler::new();
         assert_eq!(sampler.grid_size(), 0);
     }
 
     #[test]
     fn test_grid_size_single_distribution() {
-        let sampler = GridSearchSampler::builder().n_points_per_param(5).build();
+        let sampler = GridSampler::builder().n_points_per_param(5).build();
         let dist = Distribution::Float(FloatDistribution {
             low: 0.0,
             high: 1.0,
@@ -1083,7 +1115,7 @@ mod tests {
 
     #[test]
     fn test_grid_size_multiple_distributions() {
-        let sampler = GridSearchSampler::builder().n_points_per_param(3).build();
+        let sampler = GridSampler::builder().n_points_per_param(3).build();
         let dist1 = Distribution::Float(FloatDistribution {
             low: 0.0,
             high: 1.0,

src/sampler/mod.rs

@@ -1,9 +1,193 @@
 //! Sampler trait and implementations for parameter sampling.
+//!
+//! A sampler generates parameter values for each trial. It receives a
+//! [`Distribution`] describing the parameter space, a monotonically increasing
+//! `trial_id`, and the list of all [`CompletedTrial`]s so far, and returns a
+//! [`ParamValue`] that matches the distribution variant.
+//!
+//! # Available samplers
+//!
+//! ## Single-objective
+//!
+//! | Sampler | Algorithm | Best for |
+//! |---------|-----------|----------|
+//! | [`RandomSampler`] | Uniform independent sampling | Baselines, startup phases |
+//! | [`TpeSampler`] | Tree-Parzen Estimator | General-purpose Bayesian optimization |
+//! | [`TpeSampler`] (multivariate) | Multivariate TPE with tree-structured Parzen | Correlated parameters |
+//! | [`GridSampler`] | Exhaustive grid evaluation | Small discrete spaces |
+//! | [`SobolSampler`]\* | Quasi-random Sobol sequences | Uniform coverage without model |
+//! | [`CmaEsSampler`]\* | Covariance Matrix Adaptation | Continuous, non-separable problems |
+//! | [`GpSampler`]\* | Gaussian Process with EI | Expensive, low-dimensional functions |
+//! | [`DESampler`] | Differential Evolution | Population-based, multi-modal landscapes |
+//! | [`BohbSampler`] | Bayesian Optimization + `HyperBand` | Combined sampling and pruning |
+//!
+//! \*Requires a feature flag (`sobol`, `cma-es`, or `gp`).
+//!
+//! ## Multi-objective
+//!
+//! | Sampler | Algorithm | Best for |
+//! |---------|-----------|----------|
+//! | [`Nsga2Sampler`] | NSGA-II | General multi-objective with 2-3 objectives |
+//! | [`Nsga3Sampler`] | NSGA-III | Many-objective (4+ objectives) |
+//! | [`MoeadSampler`] | MOEA/D with decomposition | Structured Pareto front exploration |
+//! | [`MotpeSampler`] | Multi-objective TPE | Bayesian multi-objective |
+//!
+//! # Implementing a custom sampler
+//!
+//! Implement the [`Sampler`] trait with its single method:
+//!
+//! ```rust
+//! use optimizer::sampler::{Sampler, CompletedTrial};
+//! use optimizer::distribution::Distribution;
+//! use optimizer::param::ParamValue;
+//!
+//! /// A sampler that always picks the midpoint of each distribution.
+//! struct MidpointSampler;
+//!
+//! impl Sampler for MidpointSampler {
+//!     fn sample(
+//!         &self,
+//!         distribution: &Distribution,
+//!         _trial_id: u64,
+//!         _history: &[CompletedTrial],
+//!     ) -> ParamValue {
+//!         match distribution {
+//!             Distribution::Float(fd) => {
+//!                 ParamValue::Float((fd.low + fd.high) / 2.0)
+//!             }
+//!             Distribution::Int(id) => {
+//!                 ParamValue::Int((id.low + id.high) / 2)
+//!             }
+//!             Distribution::Categorical(cd) => {
+//!                 ParamValue::Categorical(cd.n_choices / 2)
+//!             }
+//!         }
+//!     }
+//! }
+//! ```
+//!
+//! The arguments to [`Sampler::sample`]:
+//!
+//! - **`distribution`** — a [`Distribution::Float`], [`Distribution::Int`], or
+//!   [`Distribution::Categorical`] that describes the parameter bounds,
+//!   log-scale flag, and optional step size.
+//! - **`trial_id`** — a monotonically increasing identifier. Useful for
+//!   deterministic RNG seeding (see [Stateless vs stateful samplers]).
+//! - **`history`** — all completed trials so far. May be empty on the first
+//!   trial. Model-based samplers use this to guide future sampling.
+//! - **Return value** — the [`ParamValue`] variant *must* match the
+//!   distribution variant (`Float` → `ParamValue::Float`, etc.).
+//!
+//! [Stateless vs stateful samplers]: #stateless-vs-stateful-samplers
+//!
+//! # Stateless vs stateful samplers
+//!
+//! **Stateless** samplers derive all randomness from a deterministic function
+//! of `seed + trial_id + distribution`. They use an [`AtomicU64`] call-sequence
+//! counter to disambiguate multiple calls within the same trial, but need no
+//! `Mutex`. See [`RandomSampler`] and [`TpeSampler`] for this pattern.
+//!
+//! **Stateful** samplers maintain mutable state (e.g. a population pool)
+//! across calls. Wrap mutable state in `parking_lot::Mutex<State>` and lock
+//! for the duration of [`Sampler::sample`]. See [`DESampler`] and
+//! [`GridSampler`] for this pattern.
+//!
+//! [`AtomicU64`]: core::sync::atomic::AtomicU64
+//!
+//! # Cold start handling
+//!
+//! Model-based samplers need completed trials before their surrogate model is
+//! useful. The standard pattern is to check `history.len() < n_startup_trials`
+//! and fall back to random sampling during the startup phase. Expose this as a
+//! builder parameter so users can tune the trade-off between exploration and
+//! exploitation. See [`TpeSampler`] for a reference implementation.
+//!
+//! # Reading trial history
+//!
+//! The `history` slice contains only completed trials (never pending ones).
+//! Common operations:
+//!
+//! - **Extract a parameter value:**
+//!   `trial.params.get(&param_id)` returns `Option<&ParamValue>`.
+//! - **Find the best trial:**
+//!   `history.iter().min_by(|a, b| a.value.partial_cmp(&b.value).unwrap())`.
+//! - **Filter by state:**
+//!   `history.iter().filter(|t| t.state == TrialState::Complete)`.
+//! - **Check feasibility:**
+//!   `trial.is_feasible()` returns `true` when all constraints are ≤ 0.
+//!
+//! # Thread safety
+//!
+//! The [`Sampler`] trait requires `Send + Sync`. [`Study`](crate::Study) stores
+//! the sampler as `Arc<dyn Sampler>`, so multiple threads may call
+//! [`Sampler::sample`] concurrently.
+//!
+//! - **Stateless:** `AtomicU64` counters satisfy `Send + Sync` without locking.
+//! - **Stateful:** use `parking_lot::Mutex` (the crate convention) or
+//!   `std::sync::Mutex` to protect mutable state.
+//!
+//! # Testing custom samplers
+//!
+//! Recommended test categories:
+//!
+//! 1. **Bounds compliance** — sample many values and assert they fall within
+//!    the distribution range.
+//! 2. **Step / log-scale correctness** — verify that discretized and
+//!    log-scaled distributions produce valid values.
+//! 3. **Reproducibility** — the same seed must produce the same output.
+//! 4. **History sensitivity** — model-based samplers should produce different
+//!    (better) samples as history grows.
+//! 5. **Empty history** — `sample()` must not panic when `history` is empty.
+//!
+//! # Using a custom sampler with Study
+//!
+//! ```rust
+//! use optimizer::{Direction, Study};
+//! use optimizer::sampler::{Sampler, CompletedTrial};
+//! use optimizer::distribution::Distribution;
+//! use optimizer::param::ParamValue;
+//!
+//! struct MySampler;
+//! impl Sampler for MySampler {
+//!     fn sample(
+//!         &self,
+//!         distribution: &Distribution,
+//!         _trial_id: u64,
+//!         _history: &[CompletedTrial],
+//!     ) -> ParamValue {
+//!         match distribution {
+//!             Distribution::Float(fd) => ParamValue::Float(fd.low),
+//!             Distribution::Int(id) => ParamValue::Int(id.low),
+//!             Distribution::Categorical(_) => ParamValue::Categorical(0),
+//!         }
+//!     }
+//! }
+//!
+//! let study: Study<f64> = Study::with_sampler(Direction::Minimize, MySampler);
+//! ```
+//!
+//! The sampler is wrapped in `Arc<dyn Sampler>` internally.
+//!
+//! # Reference implementations
+//!
+//! - [`RandomSampler`] — simplest sampler; stateless, ignores history.
+//! - [`TpeSampler`] — model-based with cold start fallback.
+//! - [`DESampler`] — stateful, population-based.
+//! - [`GridSampler`] — deterministic, exhaustive search.
 
 pub mod bohb;
 #[cfg(feature = "cma-es")]
 pub mod cma_es;
+pub(crate) mod common;
+pub mod de;
+pub(crate) mod genetic;
+#[cfg(feature = "gp")]
+pub mod gp;
 pub mod grid;
+pub mod moead;
+pub mod motpe;
+pub mod nsga2;
+pub mod nsga3;
 pub mod random;
 #[cfg(feature = "sobol")]
 pub mod sobol;
@@ -11,6 +195,22 @@ pub mod tpe;
 
 use std::collections::HashMap;
 
+pub use bohb::BohbSampler;
+#[cfg(feature = "cma-es")]
+pub use cma_es::CmaEsSampler;
+pub use de::{DESampler, DEStrategy};
+#[cfg(feature = "gp")]
+pub use gp::GpSampler;
+pub use grid::GridSampler;
+pub use moead::{Decomposition, MoeadSampler};
+pub use motpe::MotpeSampler;
+pub use nsga2::Nsga2Sampler;
+pub use nsga3::Nsga3Sampler;
+pub use random::RandomSampler;
+#[cfg(feature = "sobol")]
+pub use sobol::SobolSampler;
+pub use tpe::TpeSampler;
+
 use crate::distribution::Distribution;
 use crate::param::ParamValue;
 use crate::parameter::{ParamId, Parameter};
@@ -96,13 +296,9 @@ impl<V> CompletedTrial<V> {
     /// Looks up the parameter by its unique id and casts the stored
     /// [`ParamValue`] to the parameter's typed value.
     ///
-    /// Returns `None` if the parameter was not used in this trial.
-    ///
-    /// # Panics
-    ///
-    /// Panics if the stored value is incompatible with the parameter type
-    /// (e.g., a `Float` value stored for an `IntParam`). This indicates
-    /// a bug in the program, not a runtime error.
+    /// Returns `None` if the parameter was not used in this trial or if
+    /// the stored value is incompatible with the parameter type (e.g., a
+    /// `Float` value stored for an `IntParam`).
     ///
     /// # Examples
     ///
@@ -114,7 +310,7 @@ impl<V> CompletedTrial<V> {
     /// let x = FloatParam::new(-10.0, 10.0);
     ///
     /// study
-    ///     .optimize(5, |trial| {
+    ///     .optimize(5, |trial: &mut optimizer::Trial| {
     ///         let val = x.suggest(trial)?;
     ///         Ok::<_, optimizer::Error>(val * val)
     ///     })
@@ -125,11 +321,9 @@ impl<V> CompletedTrial<V> {
     /// assert!((-10.0..=10.0).contains(&x_val));
     /// ```
     pub fn get<P: Parameter>(&self, param: &P) -> Option<P::Value> {
-        self.params.get(&param.id()).map(|v| {
-            param
-                .cast_param_value(v)
-                .expect("parameter type mismatch: stored value incompatible with parameter")
-        })
+        self.params
+            .get(&param.id())
+            .and_then(|v| param.cast_param_value(v).ok())
     }
 
     /// Returns `true` if all constraints are satisfied (values <= 0.0).
@@ -151,6 +345,74 @@ impl<V> CompletedTrial<V> {
     pub fn user_attrs(&self) -> &HashMap<String, AttrValue> {
         &self.user_attrs
     }
+
+    /// Validates that all floating-point fields are finite (not NaN or
+    /// Infinity).
+    ///
+    /// Checks distribution bounds, parameter values, constraints, and
+    /// intermediate values.  Returns a description of the first invalid
+    /// field found, or `Ok(())` if everything is valid.
+    ///
+    /// # Errors
+    ///
+    /// Returns a `String` describing the first non-finite value found.
+    pub fn validate(&self) -> core::result::Result<(), String> {
+        for (id, dist) in &self.distributions {
+            if let Distribution::Float(fd) = dist {
+                if !fd.low.is_finite() {
+                    return Err(format!(
+                        "trial {}: float distribution for param {id} has non-finite low bound ({})",
+                        self.id, fd.low
+                    ));
+                }
+                if !fd.high.is_finite() {
+                    return Err(format!(
+                        "trial {}: float distribution for param {id} has non-finite high bound ({})",
+                        self.id, fd.high
+                    ));
+                }
+                if let Some(step) = fd.step
+                    && !step.is_finite()
+                {
+                    return Err(format!(
+                        "trial {}: float distribution for param {id} has non-finite step ({step})",
+                        self.id
+                    ));
+                }
+            }
+        }
+
+        for (id, pv) in &self.params {
+            if let ParamValue::Float(v) = pv
+                && !v.is_finite()
+            {
+                return Err(format!(
+                    "trial {}: param {id} has non-finite float value ({v})",
+                    self.id
+                ));
+            }
+        }
+
+        for (i, &c) in self.constraints.iter().enumerate() {
+            if !c.is_finite() {
+                return Err(format!(
+                    "trial {}: constraint[{i}] is non-finite ({c})",
+                    self.id
+                ));
+            }
+        }
+
+        for &(step, v) in &self.intermediate_values {
+            if !v.is_finite() {
+                return Err(format!(
+                    "trial {}: intermediate value at step {step} is non-finite ({v})",
+                    self.id
+                ));
+            }
+        }
+
+        Ok(())
+    }
 }
 
 /// A pending (running) trial with its parameters and distributions, but no objective value yet.
@@ -193,6 +455,50 @@ impl PendingTrial {
 /// Samplers are responsible for generating parameter values based on
 /// the distribution and historical trial data. The trait requires
 /// `Send + Sync` to support concurrent and async optimization.
+///
+/// # Implementing a custom sampler
+///
+/// ```
+/// use optimizer::sampler::{Sampler, CompletedTrial};
+/// use optimizer::distribution::Distribution;
+/// use optimizer::param::ParamValue;
+///
+/// struct NoisySampler {
+///     noise_scale: f64,
+///     seed: u64,
+/// }
+///
+/// impl Sampler for NoisySampler {
+///     fn sample(
+///         &self,
+///         distribution: &Distribution,
+///         trial_id: u64,
+///         history: &[CompletedTrial],
+///     ) -> ParamValue {
+///         // Find the best value seen so far, or fall back to the midpoint
+///         match distribution {
+///             Distribution::Float(fd) => {
+///                 let center = if history.is_empty() {
+///                     (fd.low + fd.high) / 2.0
+///                 } else {
+///                     history.iter()
+///                         .filter_map(|t| t.params.values().next())
+///                         .filter_map(|v| if let ParamValue::Float(f) = v { Some(*f) } else { None })
+///                         .next()
+///                         .unwrap_or((fd.low + fd.high) / 2.0)
+///                 };
+///                 let noise = (trial_id as f64 * 0.1).sin() * self.noise_scale;
+///                 ParamValue::Float(center + noise)
+///             }
+///             Distribution::Int(id) => ParamValue::Int((id.low + id.high) / 2),
+///             Distribution::Categorical(cd) => ParamValue::Categorical(trial_id as usize % cd.n_choices),
+///         }
+///     }
+/// }
+/// ```
+///
+/// See the [module-level documentation](self) for a comprehensive guide
+/// covering cold start handling, thread safety patterns, and testing.
 pub trait Sampler: Send + Sync {
     /// Samples a parameter value from the given distribution.
     ///

src/sampler/moead.rs

@@ -0,0 +1,672 @@
+//! MOEA/D (Multi-Objective Evolutionary Algorithm based on Decomposition) sampler.
+//!
+//! MOEA/D takes a fundamentally different approach from Pareto-based
+//! algorithms like NSGA-II/III. It **decomposes** the multi-objective
+//! problem into a set of scalar subproblems using evenly distributed
+//! weight vectors (Das-Dennis points), then solves them collaboratively
+//! through **neighborhood-based mating and replacement**.
+//!
+//! # Algorithm
+//!
+//! 1. **Decompose** — generate weight vectors on the unit simplex and
+//!    assign one scalar subproblem per weight vector.
+//! 2. **Build neighborhoods** — for each subproblem, find its T nearest
+//!    neighbors by Euclidean distance between weight vectors.
+//! 3. **Mate from neighborhood** — select parents from the neighborhood
+//!    of each subproblem and produce offspring via SBX crossover +
+//!    polynomial mutation.
+//! 4. **Scalarize and update** — evaluate offspring using a scalarization
+//!    function and update neighboring subproblems if the offspring improves
+//!    their scalar value.
+//! 5. **Update ideal point** — track the best value seen per objective.
+//!
+//! # Scalarization methods
+//!
+//! | Method | Formula | Best for |
+//! |--------|---------|----------|
+//! | [`Tchebycheff`](Decomposition::Tchebycheff) (default) | `max(wᵢ * \|fᵢ - zᵢ*\|)` | General purpose, handles non-convex fronts |
+//! | [`WeightedSum`](Decomposition::WeightedSum) | `Σ(wᵢ * fᵢ)` | Convex Pareto fronts only |
+//! | [`Pbi`](Decomposition::Pbi) | `d₁ + θ * d₂` | Fine-grained convergence/diversity control |
+//!
+//! # When to use
+//!
+//! - Problems where you want **evenly distributed** solutions along the
+//!   Pareto front (one solution per weight direction).
+//! - Many-objective optimization (3+ objectives) — scales well because
+//!   each subproblem is a simple scalar optimization.
+//! - Problems with **non-convex** Pareto fronts (use Tchebycheff or PBI).
+//! - When you need explicit control over the trade-off distribution via
+//!   weight vectors.
+//!
+//! For Pareto-based approaches, see
+//! [`Nsga2Sampler`](super::nsga2::Nsga2Sampler) (crowding distance) or
+//! [`Nsga3Sampler`](super::nsga3::Nsga3Sampler) (reference-point niching).
+//!
+//! # Configuration
+//!
+//! | Parameter | Builder method | Default |
+//! |-----------|---------------|---------|
+//! | Population size | [`population_size`](MoeadSamplerBuilder::population_size) | Number of Das-Dennis weight vectors |
+//! | Neighborhood size (T) | [`neighborhood_size`](MoeadSamplerBuilder::neighborhood_size) | `min(20, pop_size)` |
+//! | Decomposition method | [`decomposition`](MoeadSamplerBuilder::decomposition) | Tchebycheff |
+//! | Crossover probability | [`crossover_prob`](MoeadSamplerBuilder::crossover_prob) | 1.0 |
+//! | SBX distribution index | [`crossover_eta`](MoeadSamplerBuilder::crossover_eta) | 20.0 |
+//! | Mutation distribution index | [`mutation_eta`](MoeadSamplerBuilder::mutation_eta) | 20.0 |
+//! | Random seed | [`seed`](MoeadSamplerBuilder::seed) | random |
+//!
+//! # Examples
+//!
+//! ```
+//! use optimizer::Direction;
+//! use optimizer::multi_objective::MultiObjectiveStudy;
+//! use optimizer::parameter::{FloatParam, Parameter};
+//! use optimizer::sampler::moead::MoeadSampler;
+//!
+//! let sampler = MoeadSampler::with_seed(42);
+//! let study =
+//!     MultiObjectiveStudy::with_sampler(vec![Direction::Minimize, Direction::Minimize], sampler);
+//!
+//! let x = FloatParam::new(0.0, 1.0);
+//! study
+//!     .optimize(100, |trial: &mut optimizer::Trial| {
+//!         let xv = x.suggest(trial)?;
+//!         Ok::<_, optimizer::Error>(vec![xv, 1.0 - xv])
+//!     })
+//!     .unwrap();
+//! ```
+
+use parking_lot::Mutex;
+
+use super::genetic::{
+    self, Candidate, EvolutionaryState, Phase, advance_generation, auto_divisions,
+    collect_evaluated_generation, crossover, das_dennis, extract_trial_params,
+    generate_random_candidates, mutate, sample_from_candidate, sample_random,
+};
+use crate::distribution::Distribution;
+use crate::multi_objective::MultiObjectiveTrial;
+use crate::param::ParamValue;
+use crate::types::Direction;
+
+/// Decomposition (scalarization) method for [`MoeadSampler`].
+///
+/// Control how multi-objective values are reduced to a single scalar
+/// for each subproblem. The default is [`Tchebycheff`](Self::Tchebycheff),
+/// which handles both convex and non-convex Pareto fronts.
+#[derive(Debug, Clone, Default)]
+pub enum Decomposition {
+    /// Weighted sum: `Σ(wᵢ * fᵢ)`.
+    ///
+    /// Simplest method but can only find solutions on convex regions
+    /// of the Pareto front.
+    WeightedSum,
+    /// Tchebycheff: `max(wᵢ * |fᵢ - zᵢ*|)`.
+    ///
+    /// Handles non-convex Pareto fronts. The most commonly used
+    /// decomposition method (default).
+    #[default]
+    Tchebycheff,
+    /// Penalty-based Boundary Intersection: `d₁ + θ * d₂`.
+    ///
+    /// Provides fine-grained control over the convergence/diversity
+    /// balance via the penalty parameter `theta`. Higher `theta`
+    /// favors solutions closer to the weight direction.
+    Pbi {
+        /// Penalty parameter controlling the balance between convergence
+        /// and diversity. Default: 5.0.
+        theta: f64,
+    },
+}
+
+/// MOEA/D sampler for multi-objective optimization.
+///
+/// Decompose a multi-objective problem into scalar subproblems using
+/// weight vectors and solve them collaboratively via neighborhood-based
+/// mating. Supports [`Tchebycheff`](Decomposition::Tchebycheff),
+/// [`WeightedSum`](Decomposition::WeightedSum), and
+/// [`Pbi`](Decomposition::Pbi) scalarization.
+///
+/// Create with [`MoeadSampler::new`], [`MoeadSampler::with_seed`], or
+/// [`MoeadSampler::builder`] for full configuration.
+pub struct MoeadSampler {
+    state: Mutex<MoeadState>,
+}
+
+impl MoeadSampler {
+    /// Creates a new MOEA/D sampler with a random seed.
+    #[must_use]
+    pub fn new() -> Self {
+        Self {
+            state: Mutex::new(MoeadState::new(MoeadConfig::default(), None)),
+        }
+    }
+
+    /// Creates a new MOEA/D sampler with a fixed seed.
+    #[must_use]
+    pub fn with_seed(seed: u64) -> Self {
+        Self {
+            state: Mutex::new(MoeadState::new(MoeadConfig::default(), Some(seed))),
+        }
+    }
+
+    /// Creates a builder for configuring a `MoeadSampler`.
+    #[must_use]
+    pub fn builder() -> MoeadSamplerBuilder {
+        MoeadSamplerBuilder::default()
+    }
+}
+
+impl Default for MoeadSampler {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+/// Builder for [`MoeadSampler`].
+#[derive(Debug, Clone, Default)]
+pub struct MoeadSamplerBuilder {
+    population_size: Option<usize>,
+    neighborhood_size: Option<usize>,
+    decomposition: Decomposition,
+    crossover_prob: Option<f64>,
+    crossover_eta: Option<f64>,
+    mutation_eta: Option<f64>,
+    seed: Option<u64>,
+}
+
+impl MoeadSamplerBuilder {
+    /// Sets the population size. If unset, equals the number of
+    /// Das-Dennis weight vectors.
+    #[must_use]
+    pub fn population_size(mut self, size: usize) -> Self {
+        self.population_size = Some(size);
+        self
+    }
+
+    /// Sets the neighborhood size (T). Default: `min(20, pop_size)`.
+    #[must_use]
+    pub fn neighborhood_size(mut self, size: usize) -> Self {
+        self.neighborhood_size = Some(size);
+        self
+    }
+
+    /// Sets the decomposition method. Default: Tchebycheff.
+    #[must_use]
+    pub fn decomposition(mut self, decomp: Decomposition) -> Self {
+        self.decomposition = decomp;
+        self
+    }
+
+    /// Sets the crossover probability. Default: 1.0.
+    #[must_use]
+    pub fn crossover_prob(mut self, prob: f64) -> Self {
+        self.crossover_prob = Some(prob);
+        self
+    }
+
+    /// Sets the SBX distribution index. Default: 20.0.
+    #[must_use]
+    pub fn crossover_eta(mut self, eta: f64) -> Self {
+        self.crossover_eta = Some(eta);
+        self
+    }
+
+    /// Sets the polynomial mutation distribution index. Default: 20.0.
+    #[must_use]
+    pub fn mutation_eta(mut self, eta: f64) -> Self {
+        self.mutation_eta = Some(eta);
+        self
+    }
+
+    /// Sets the random seed for reproducibility.
+    #[must_use]
+    pub fn seed(mut self, seed: u64) -> Self {
+        self.seed = Some(seed);
+        self
+    }
+
+    /// Builds the configured [`MoeadSampler`].
+    #[must_use]
+    pub fn build(self) -> MoeadSampler {
+        let config = MoeadConfig {
+            user_population_size: self.population_size,
+            neighborhood_size: self.neighborhood_size,
+            decomposition: self.decomposition,
+            crossover_prob: self.crossover_prob.unwrap_or(1.0),
+            crossover_eta: self.crossover_eta.unwrap_or(20.0),
+            mutation_eta: self.mutation_eta.unwrap_or(20.0),
+        };
+        MoeadSampler {
+            state: Mutex::new(MoeadState::new(config, self.seed)),
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Internal types
+// ---------------------------------------------------------------------------
+
+#[derive(Debug, Clone)]
+struct MoeadConfig {
+    user_population_size: Option<usize>,
+    neighborhood_size: Option<usize>,
+    decomposition: Decomposition,
+    crossover_prob: f64,
+    crossover_eta: f64,
+    mutation_eta: f64,
+}
+
+impl Default for MoeadConfig {
+    fn default() -> Self {
+        Self {
+            user_population_size: None,
+            neighborhood_size: None,
+            decomposition: Decomposition::default(),
+            crossover_prob: 1.0,
+            crossover_eta: 20.0,
+            mutation_eta: 20.0,
+        }
+    }
+}
+
+struct MoeadState {
+    evo: EvolutionaryState,
+    config: MoeadConfig,
+    /// Weight vectors (Das-Dennis), one per subproblem.
+    weight_vectors: Vec<Vec<f64>>,
+    /// Neighborhoods: for each subproblem, indices of T nearest weight vectors.
+    neighborhoods: Vec<Vec<usize>>,
+    /// Ideal point z* (best per-objective in minimize-space).
+    ideal_point: Vec<f64>,
+    /// Current population's objective values in minimize-space (one per subproblem).
+    population_values: Vec<Vec<f64>>,
+    /// Current population's parameter vectors (one per subproblem).
+    population_params: Vec<Vec<ParamValue>>,
+    /// Whether the MOEA/D state has been initialized.
+    initialized: bool,
+}
+
+impl MoeadState {
+    fn new(config: MoeadConfig, seed: Option<u64>) -> Self {
+        Self {
+            evo: EvolutionaryState::new(seed),
+            config,
+            weight_vectors: Vec::new(),
+            neighborhoods: Vec::new(),
+            ideal_point: Vec::new(),
+            population_values: Vec::new(),
+            population_params: Vec::new(),
+            initialized: false,
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// MultiObjectiveSampler implementation
+// ---------------------------------------------------------------------------
+
+impl crate::multi_objective::MultiObjectiveSampler for MoeadSampler {
+    fn sample(
+        &self,
+        distribution: &Distribution,
+        trial_id: u64,
+        history: &[MultiObjectiveTrial],
+        directions: &[Direction],
+    ) -> ParamValue {
+        let mut state = self.state.lock();
+
+        match &state.evo.phase {
+            Phase::Discovery => {
+                if let Some(value) =
+                    genetic::sample_discovery(&mut state.evo, distribution, trial_id)
+                {
+                    return value;
+                }
+                // Transitioned to active phase
+                initialize_moead(&mut state, directions);
+                generate_random_candidates(&mut state.evo);
+                sample_from_candidate(&mut state.evo, trial_id)
+            }
+            Phase::Active => {
+                maybe_generate_new_generation(&mut state, history, directions);
+                sample_from_candidate(&mut state.evo, trial_id)
+            }
+        }
+    }
+}
+
+/// Initialize MOEA/D: weight vectors, neighborhoods, ideal point.
+fn initialize_moead(state: &mut MoeadState, directions: &[Direction]) {
+    let n_obj = directions.len();
+
+    // Generate weight vectors
+    let divisions = auto_divisions(n_obj, state.config.user_population_size.unwrap_or(100));
+    state.weight_vectors = das_dennis(n_obj, divisions);
+
+    let pop_size = state
+        .config
+        .user_population_size
+        .unwrap_or(state.weight_vectors.len())
+        .max(4);
+
+    // Trim or pad weight vectors to match population size
+    state.weight_vectors.truncate(pop_size);
+    while state.weight_vectors.len() < pop_size {
+        // Duplicate random existing weight vectors
+        let idx = state.evo.rng.usize(0..state.weight_vectors.len());
+        let w = state.weight_vectors[idx].clone();
+        state.weight_vectors.push(w);
+    }
+
+    // Compute neighborhoods
+    let t = state
+        .config
+        .neighborhood_size
+        .unwrap_or_else(|| 20.min(pop_size));
+    let t = t.min(pop_size);
+    state.neighborhoods = compute_neighborhoods(&state.weight_vectors, t);
+
+    state.evo.population_size = pop_size;
+    state.evo.phase = Phase::Active;
+    state.ideal_point = vec![f64::INFINITY; n_obj];
+    state.initialized = true;
+}
+
+/// Compute T-nearest neighborhoods by Euclidean distance between weight vectors.
+fn compute_neighborhoods(weights: &[Vec<f64>], t: usize) -> Vec<Vec<usize>> {
+    let n = weights.len();
+    weights
+        .iter()
+        .map(|wi| {
+            let mut distances: Vec<(usize, f64)> = (0..n)
+                .map(|j| {
+                    let d: f64 = wi
+                        .iter()
+                        .zip(&weights[j])
+                        .map(|(&a, &b)| (a - b).powi(2))
+                        .sum::<f64>()
+                        .sqrt();
+                    (j, d)
+                })
+                .collect();
+            distances.sort_by(|a, b| a.1.partial_cmp(&b.1).unwrap_or(core::cmp::Ordering::Equal));
+            distances.into_iter().take(t).map(|(idx, _)| idx).collect()
+        })
+        .collect()
+}
+
+/// Convert values to minimize-space.
+fn to_minimize_space(values: &[f64], directions: &[Direction]) -> Vec<f64> {
+    values
+        .iter()
+        .zip(directions)
+        .map(|(&v, d)| match d {
+            Direction::Minimize => v,
+            Direction::Maximize => -v,
+        })
+        .collect()
+}
+
+fn maybe_generate_new_generation(
+    state: &mut MoeadState,
+    history: &[MultiObjectiveTrial],
+    directions: &[Direction],
+) {
+    if state.evo.candidates.is_empty() {
+        generate_random_candidates(&mut state.evo);
+        return;
+    }
+
+    if let Some(evaluated) = collect_evaluated_generation(&state.evo, history) {
+        let offspring = moead_generate_offspring(state, &evaluated, directions);
+        advance_generation(&mut state.evo, offspring);
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Scalarization functions
+// ---------------------------------------------------------------------------
+
+/// Weighted sum scalarization: `sum(w_i * f_i)`.
+fn scalarize_weighted_sum(values: &[f64], weight: &[f64]) -> f64 {
+    values.iter().zip(weight).map(|(&v, &w)| w * v).sum()
+}
+
+/// Tchebycheff scalarization: `max(w_i * |f_i - z_i*|)`.
+fn scalarize_tchebycheff(values: &[f64], weight: &[f64], ideal: &[f64]) -> f64 {
+    values
+        .iter()
+        .zip(weight)
+        .zip(ideal)
+        .map(|((&v, &w), &z)| {
+            let w = if w < 1e-6 { 1e-6 } else { w };
+            w * (v - z).abs()
+        })
+        .fold(f64::NEG_INFINITY, f64::max)
+}
+
+/// PBI scalarization: `d1 + theta * d2`.
+///
+/// d1 = projection onto weight direction, d2 = perpendicular distance.
+fn scalarize_pbi(values: &[f64], weight: &[f64], ideal: &[f64], theta: f64) -> f64 {
+    let n = values.len();
+
+    // Direction from ideal to the point
+    let diff: Vec<f64> = values.iter().zip(ideal).map(|(&v, &z)| v - z).collect();
+
+    // Normalize weight vector
+    let w_norm: f64 = weight.iter().map(|&w| w * w).sum::<f64>().sqrt();
+    if w_norm < 1e-30 {
+        return f64::INFINITY;
+    }
+    let w_unit: Vec<f64> = weight.iter().map(|&w| w / w_norm).collect();
+
+    // d1 = projection of diff onto weight direction
+    let d1: f64 = diff.iter().zip(&w_unit).map(|(&d, &w)| d * w).sum();
+
+    // d2 = perpendicular distance
+    let d2_sq: f64 = (0..n)
+        .map(|i| {
+            let proj = d1 * w_unit[i];
+            (diff[i] - proj).powi(2)
+        })
+        .sum::<f64>();
+
+    d1 + theta * d2_sq.sqrt()
+}
+
+/// Evaluate scalarization for a given decomposition method.
+fn scalarize(values: &[f64], weight: &[f64], ideal: &[f64], decomposition: &Decomposition) -> f64 {
+    match decomposition {
+        Decomposition::WeightedSum => scalarize_weighted_sum(values, weight),
+        Decomposition::Tchebycheff => scalarize_tchebycheff(values, weight, ideal),
+        Decomposition::Pbi { theta } => scalarize_pbi(values, weight, ideal, *theta),
+    }
+}
+
+// ---------------------------------------------------------------------------
+// MOEA/D generation algorithm
+// ---------------------------------------------------------------------------
+
+fn moead_generate_offspring(
+    state: &mut MoeadState,
+    population: &[&MultiObjectiveTrial],
+    directions: &[Direction],
+) -> Vec<Candidate> {
+    let pop_size = state.evo.population_size;
+
+    if population.len() < 2 {
+        return (0..pop_size)
+            .map(|_| {
+                let params = state
+                    .evo
+                    .dimensions
+                    .iter()
+                    .map(|d| sample_random(&mut state.evo.rng, &d.distribution))
+                    .collect();
+                Candidate { params }
+            })
+            .collect();
+    }
+
+    // Extract current population parameters and objective values
+    let current_params: Vec<Vec<ParamValue>> = population
+        .iter()
+        .map(|t| extract_trial_params(t, &state.evo.dimensions, &mut state.evo.rng))
+        .collect();
+
+    let current_values: Vec<Vec<f64>> = population
+        .iter()
+        .map(|t| to_minimize_space(&t.values, directions))
+        .collect();
+
+    // Update ideal point
+    for vals in &current_values {
+        for (i, &v) in vals.iter().enumerate() {
+            if i < state.ideal_point.len() && v < state.ideal_point[i] {
+                state.ideal_point[i] = v;
+            }
+        }
+    }
+
+    // Assign each solution to its best subproblem via scalarization
+    // and select the best solution for each subproblem as its representative
+    let n_weights = state.weight_vectors.len();
+    let mut best_for_subproblem: Vec<usize> = Vec::with_capacity(n_weights);
+
+    for j in 0..n_weights {
+        let mut best_idx = 0;
+        let mut best_val = f64::INFINITY;
+        for (k, vals) in current_values.iter().enumerate() {
+            let s = scalarize(
+                vals,
+                &state.weight_vectors[j],
+                &state.ideal_point,
+                &state.config.decomposition,
+            );
+            if s < best_val {
+                best_val = s;
+                best_idx = k;
+            }
+        }
+        best_for_subproblem.push(best_idx);
+    }
+
+    // Store current population state
+    state.population_values = current_values;
+    state.population_params = current_params;
+
+    // Generate offspring: for each subproblem, mate from neighborhood
+    let mut offspring = Vec::with_capacity(pop_size);
+
+    for i in 0..pop_size.min(state.neighborhoods.len()) {
+        let neighborhood = &state.neighborhoods[i];
+
+        // Pick two parents from the neighborhood using subproblem assignments
+        let n1 = neighborhood[state.evo.rng.usize(0..neighborhood.len())];
+        let n2 = neighborhood[state.evo.rng.usize(0..neighborhood.len())];
+
+        let p1_idx = best_for_subproblem[n1 % best_for_subproblem.len()];
+        let p2_idx = best_for_subproblem[n2 % best_for_subproblem.len()];
+
+        let p1 = &state.population_params[p1_idx];
+        let p2 = &state.population_params[p2_idx];
+
+        let (mut child1, _child2) = crossover(
+            &mut state.evo.rng,
+            p1,
+            p2,
+            &state.evo.dimensions,
+            state.config.crossover_prob,
+            state.config.crossover_eta,
+        );
+
+        mutate(
+            &mut state.evo.rng,
+            &mut child1,
+            &state.evo.dimensions,
+            state.config.mutation_eta,
+        );
+
+        offspring.push(Candidate { params: child1 });
+    }
+
+    // If pop_size > neighborhoods, fill remaining with random neighborhood crossover
+    while offspring.len() < pop_size {
+        let i = state.evo.rng.usize(0..state.neighborhoods.len());
+        let neighborhood = &state.neighborhoods[i];
+        let n1 = neighborhood[state.evo.rng.usize(0..neighborhood.len())];
+        let n2 = neighborhood[state.evo.rng.usize(0..neighborhood.len())];
+
+        let p1_idx = best_for_subproblem[n1 % best_for_subproblem.len()];
+        let p2_idx = best_for_subproblem[n2 % best_for_subproblem.len()];
+
+        let (mut child1, _) = crossover(
+            &mut state.evo.rng,
+            &state.population_params[p1_idx],
+            &state.population_params[p2_idx],
+            &state.evo.dimensions,
+            state.config.crossover_prob,
+            state.config.crossover_eta,
+        );
+
+        mutate(
+            &mut state.evo.rng,
+            &mut child1,
+            &state.evo.dimensions,
+            state.config.mutation_eta,
+        );
+
+        offspring.push(Candidate { params: child1 });
+    }
+
+    offspring
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_scalarize_weighted_sum() {
+        let values = [1.0, 2.0, 3.0];
+        let weight = [0.5, 0.3, 0.2];
+        let result = scalarize_weighted_sum(&values, &weight);
+        assert!((result - (0.5 + 0.6 + 0.6)).abs() < 1e-10);
+    }
+
+    #[test]
+    fn test_scalarize_tchebycheff() {
+        let values = [3.0, 2.0];
+        let weight = [0.5, 0.5];
+        let ideal = [1.0, 1.0];
+        let result = scalarize_tchebycheff(&values, &weight, &ideal);
+        // max(0.5 * |3-1|, 0.5 * |2-1|) = max(1.0, 0.5) = 1.0
+        assert!((result - 1.0).abs() < 1e-10);
+    }
+
+    #[test]
+    fn test_scalarize_pbi() {
+        let values = [2.0, 2.0];
+        let weight = [1.0, 1.0];
+        let ideal = [0.0, 0.0];
+        let result = scalarize_pbi(&values, &weight, &ideal, 5.0);
+        // d1 = projection of (2,2) onto (1/√2, 1/√2) = 2*√2
+        // d2 = 0 (point is on the weight direction)
+        let expected_d1 = 2.0 * (2.0_f64).sqrt();
+        assert!((result - expected_d1).abs() < 1e-10);
+    }
+
+    #[test]
+    fn test_compute_neighborhoods() {
+        let weights = vec![vec![1.0, 0.0], vec![0.5, 0.5], vec![0.0, 1.0]];
+        let neighborhoods = compute_neighborhoods(&weights, 2);
+        assert_eq!(neighborhoods.len(), 3);
+        // Each neighborhood should have 2 entries
+        for n in &neighborhoods {
+            assert_eq!(n.len(), 2);
+        }
+        // First weight [1,0] should be closest to itself and [0.5,0.5]
+        assert_eq!(neighborhoods[0][0], 0); // itself
+        assert_eq!(neighborhoods[0][1], 1); // nearest neighbor
+    }
+}

src/sampler/motpe.rs

@@ -0,0 +1,831 @@
+//! Multi-Objective Tree-Parzen Estimator (MOTPE) sampler.
+//!
+//! MOTPE extends TPE to multi-objective optimization by replacing the gamma-based
+//! split with Pareto non-dominated sorting. This lets the sampler propose
+//! parameters that push the Pareto front forward across all objectives
+//! simultaneously.
+//!
+//! # Algorithm
+//!
+//! In single-objective TPE, trials are sorted by value and split at a gamma
+//! percentile into good/bad groups. MOTPE replaces this with:
+//!
+//! 1. Compute non-dominated sorting on all completed trials.
+//! 2. Use the Pareto front (rank 0) as "good" trials.
+//! 3. Use dominated trials (rank 1+) as "bad" trials.
+//! 4. Build KDE l(x) from good, g(x) from bad.
+//! 5. Sample candidates and score by l(x)/g(x).
+//!
+//! # When to use
+//!
+//! - You have 2+ objectives and want model-guided search (not pure evolutionary).
+//! - Your objectives are relatively smooth and continuous.
+//! - You want a Pareto-aware version of TPE without the overhead of full
+//!   population-based algorithms like NSGA-II or NSGA-III.
+//!
+//! For single-objective problems, use [`TpeSampler`](super::tpe::TpeSampler) instead.
+//! For many-objective (3+) problems with reference-point decomposition, consider
+//! NSGA-III or MOEA/D.
+//!
+//! # Configuration
+//!
+//! - `n_startup_trials` — number of random trials before MOTPE kicks in (default: 11)
+//! - `n_ei_candidates` — candidates evaluated per sample (default: 24)
+//! - `kde_bandwidth` — optional fixed KDE bandwidth; `None` uses Scott's rule
+//! - `seed` — optional seed for reproducibility
+//!
+//! # Examples
+//!
+//! ```
+//! use optimizer::Direction;
+//! use optimizer::multi_objective::MultiObjectiveStudy;
+//! use optimizer::parameter::{FloatParam, Parameter};
+//! use optimizer::sampler::motpe::MotpeSampler;
+//!
+//! let sampler = MotpeSampler::builder().seed(42).build();
+//! let study =
+//!     MultiObjectiveStudy::with_sampler(vec![Direction::Minimize, Direction::Minimize], sampler);
+//!
+//! let x = FloatParam::new(0.0, 1.0);
+//! study
+//!     .optimize(30, |trial: &mut optimizer::Trial| {
+//!         let xv = x.suggest(trial)?;
+//!         Ok::<_, optimizer::Error>(vec![xv, 1.0 - xv])
+//!     })
+//!     .unwrap();
+//!
+//! let front = study.pareto_front();
+//! assert!(!front.is_empty());
+//! ```
+
+use core::sync::atomic::{AtomicU64, Ordering};
+
+use crate::distribution::Distribution;
+use crate::multi_objective::{MultiObjectiveSampler, MultiObjectiveTrial};
+use crate::param::ParamValue;
+use crate::sampler::common;
+use crate::sampler::tpe::common as tpe_common;
+use crate::types::{Direction, TrialState};
+use crate::{pareto, rng_util};
+
+/// Multi-Objective TPE (MOTPE) sampler for multi-objective Bayesian optimization.
+///
+/// Use Pareto non-dominated sorting to split completed trials into "good"
+/// (non-dominated, rank 0) and "bad" (dominated) groups, then fit kernel
+/// density estimators to each group and sample new points that maximize
+/// l(x)/g(x).
+///
+/// During the startup phase (fewer than `n_startup_trials` completed),
+/// MOTPE falls back to random sampling.
+///
+/// # When to use
+///
+/// Use `MotpeSampler` when optimizing 2+ objectives and you want a
+/// model-guided sampler that adapts proposals based on the current
+/// Pareto front. For single-objective problems, use
+/// [`TpeSampler`](super::tpe::TpeSampler) instead.
+///
+/// # Examples
+///
+/// ```
+/// use optimizer::Direction;
+/// use optimizer::multi_objective::MultiObjectiveStudy;
+/// use optimizer::parameter::{FloatParam, Parameter};
+/// use optimizer::sampler::motpe::MotpeSampler;
+///
+/// let sampler = MotpeSampler::builder()
+///     .n_startup_trials(10)
+///     .n_ei_candidates(24)
+///     .seed(42)
+///     .build();
+///
+/// let study =
+///     MultiObjectiveStudy::with_sampler(vec![Direction::Minimize, Direction::Minimize], sampler);
+///
+/// let x = FloatParam::new(0.0, 1.0);
+/// study
+///     .optimize(30, |trial: &mut optimizer::Trial| {
+///         let xv = x.suggest(trial)?;
+///         Ok::<_, optimizer::Error>(vec![xv, 1.0 - xv])
+///     })
+///     .unwrap();
+///
+/// let front = study.pareto_front();
+/// assert!(!front.is_empty());
+/// ```
+pub struct MotpeSampler {
+    /// Number of trials before MOTPE kicks in (uses random sampling before this).
+    n_startup_trials: usize,
+    /// Number of candidate samples to evaluate when selecting the next point.
+    n_ei_candidates: usize,
+    /// Optional fixed bandwidth for KDE. If None, uses Scott's rule.
+    kde_bandwidth: Option<f64>,
+    /// Base seed for deterministic per-call RNG derivation (no mutex needed).
+    seed: u64,
+    /// Monotonic counter to disambiguate calls with identical (`trial_id`, distribution).
+    call_seq: AtomicU64,
+}
+
+impl MotpeSampler {
+    /// Creates a new MOTPE sampler with default settings.
+    ///
+    /// Defaults:
+    /// - `n_startup_trials`: 11
+    /// - `n_ei_candidates`: 24
+    /// - `kde_bandwidth`: None (Scott's rule)
+    #[must_use]
+    pub fn new() -> Self {
+        Self {
+            n_startup_trials: 11,
+            n_ei_candidates: 24,
+            kde_bandwidth: None,
+            seed: fastrand::u64(..),
+            call_seq: AtomicU64::new(0),
+        }
+    }
+
+    /// Creates a new MOTPE sampler with a fixed seed.
+    #[must_use]
+    pub fn with_seed(seed: u64) -> Self {
+        Self {
+            n_startup_trials: 11,
+            n_ei_candidates: 24,
+            kde_bandwidth: None,
+            seed,
+            call_seq: AtomicU64::new(0),
+        }
+    }
+
+    /// Creates a builder for configuring a MOTPE sampler.
+    #[must_use]
+    pub fn builder() -> MotpeSamplerBuilder {
+        MotpeSamplerBuilder::new()
+    }
+
+    /// Splits trials into good (non-dominated) and bad (dominated) groups
+    /// using Pareto non-dominated sorting.
+    fn split_trials<'a>(
+        history: &'a [MultiObjectiveTrial],
+        directions: &[Direction],
+    ) -> (Vec<&'a MultiObjectiveTrial>, Vec<&'a MultiObjectiveTrial>) {
+        let complete: Vec<(usize, &MultiObjectiveTrial)> = history
+            .iter()
+            .enumerate()
+            .filter(|(_, t)| t.state == TrialState::Complete)
+            .collect();
+
+        if complete.is_empty() {
+            return (vec![], vec![]);
+        }
+
+        let values: Vec<Vec<f64>> = complete.iter().map(|(_, t)| t.values.clone()).collect();
+        let constraints: Vec<Vec<f64>> = complete
+            .iter()
+            .map(|(_, t)| t.constraints.clone())
+            .collect();
+        let has_constraints = constraints.iter().any(|c| !c.is_empty());
+
+        let fronts = if has_constraints {
+            pareto::fast_non_dominated_sort_constrained(&values, directions, &constraints)
+        } else {
+            pareto::fast_non_dominated_sort(&values, directions)
+        };
+
+        if fronts.is_empty() {
+            return (vec![], vec![]);
+        }
+
+        // Front 0 = good (non-dominated), everything else = bad
+        let good: Vec<&MultiObjectiveTrial> = fronts[0].iter().map(|&i| complete[i].1).collect();
+
+        let bad: Vec<&MultiObjectiveTrial> = fronts[1..]
+            .iter()
+            .flatten()
+            .map(|&i| complete[i].1)
+            .collect();
+
+        (good, bad)
+    }
+}
+
+impl Default for MotpeSampler {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl MotpeSampler {
+    fn sample_float(
+        &self,
+        d: &crate::distribution::FloatDistribution,
+        good_trials: &[&MultiObjectiveTrial],
+        bad_trials: &[&MultiObjectiveTrial],
+        rng: &mut fastrand::Rng,
+    ) -> ParamValue {
+        let target_dist = Distribution::Float(d.clone());
+        let good_values: Vec<f64> = good_trials
+            .iter()
+            .filter_map(|t| {
+                t.distributions.iter().find_map(|(id, dist)| {
+                    if *dist == target_dist {
+                        t.params.get(id).and_then(|v| match v {
+                            ParamValue::Float(f) => Some(*f),
+                            _ => None,
+                        })
+                    } else {
+                        None
+                    }
+                })
+            })
+            .collect();
+
+        let bad_values: Vec<f64> = bad_trials
+            .iter()
+            .filter_map(|t| {
+                t.distributions.iter().find_map(|(id, dist)| {
+                    if *dist == target_dist {
+                        t.params.get(id).and_then(|v| match v {
+                            ParamValue::Float(f) => Some(*f),
+                            _ => None,
+                        })
+                    } else {
+                        None
+                    }
+                })
+            })
+            .collect();
+
+        if good_values.is_empty() || bad_values.is_empty() {
+            return ParamValue::Float(rng_util::f64_range(rng, d.low, d.high));
+        }
+
+        let value = tpe_common::sample_tpe_float(
+            d,
+            good_values,
+            bad_values,
+            self.n_ei_candidates,
+            self.kde_bandwidth,
+            rng,
+        );
+        ParamValue::Float(value)
+    }
+
+    fn sample_int(
+        &self,
+        d: &crate::distribution::IntDistribution,
+        good_trials: &[&MultiObjectiveTrial],
+        bad_trials: &[&MultiObjectiveTrial],
+        rng: &mut fastrand::Rng,
+    ) -> ParamValue {
+        let target_dist = Distribution::Int(d.clone());
+        let good_values: Vec<i64> = good_trials
+            .iter()
+            .filter_map(|t| {
+                t.distributions.iter().find_map(|(id, dist)| {
+                    if *dist == target_dist {
+                        t.params.get(id).and_then(|v| match v {
+                            ParamValue::Int(i) => Some(*i),
+                            _ => None,
+                        })
+                    } else {
+                        None
+                    }
+                })
+            })
+            .collect();
+
+        let bad_values: Vec<i64> = bad_trials
+            .iter()
+            .filter_map(|t| {
+                t.distributions.iter().find_map(|(id, dist)| {
+                    if *dist == target_dist {
+                        t.params.get(id).and_then(|v| match v {
+                            ParamValue::Int(i) => Some(*i),
+                            _ => None,
+                        })
+                    } else {
+                        None
+                    }
+                })
+            })
+            .collect();
+
+        if good_values.is_empty() || bad_values.is_empty() {
+            return common::sample_random(rng, &Distribution::Int(d.clone()));
+        }
+
+        let value = tpe_common::sample_tpe_int(
+            d,
+            good_values,
+            bad_values,
+            self.n_ei_candidates,
+            self.kde_bandwidth,
+            rng,
+        );
+        ParamValue::Int(value)
+    }
+
+    #[allow(clippy::unused_self)]
+    fn sample_categorical(
+        &self,
+        d: &crate::distribution::CategoricalDistribution,
+        good_trials: &[&MultiObjectiveTrial],
+        bad_trials: &[&MultiObjectiveTrial],
+        rng: &mut fastrand::Rng,
+    ) -> ParamValue {
+        let target_dist = Distribution::Categorical(d.clone());
+        let good_indices: Vec<usize> = good_trials
+            .iter()
+            .filter_map(|t| {
+                t.distributions.iter().find_map(|(id, dist)| {
+                    if *dist == target_dist {
+                        t.params.get(id).and_then(|v| match v {
+                            ParamValue::Categorical(i) => Some(*i),
+                            _ => None,
+                        })
+                    } else {
+                        None
+                    }
+                })
+            })
+            .collect();
+
+        let bad_indices: Vec<usize> = bad_trials
+            .iter()
+            .filter_map(|t| {
+                t.distributions.iter().find_map(|(id, dist)| {
+                    if *dist == target_dist {
+                        t.params.get(id).and_then(|v| match v {
+                            ParamValue::Categorical(i) => Some(*i),
+                            _ => None,
+                        })
+                    } else {
+                        None
+                    }
+                })
+            })
+            .collect();
+
+        if good_indices.is_empty() || bad_indices.is_empty() {
+            return common::sample_random(rng, &Distribution::Categorical(d.clone()));
+        }
+
+        let index =
+            tpe_common::sample_tpe_categorical(d.n_choices, &good_indices, &bad_indices, rng);
+        ParamValue::Categorical(index)
+    }
+}
+
+impl MultiObjectiveSampler for MotpeSampler {
+    fn sample(
+        &self,
+        distribution: &Distribution,
+        trial_id: u64,
+        history: &[MultiObjectiveTrial],
+        directions: &[Direction],
+    ) -> ParamValue {
+        let seq = self.call_seq.fetch_add(1, Ordering::Relaxed);
+        let mut rng = fastrand::Rng::with_seed(rng_util::mix_seed(
+            self.seed,
+            trial_id,
+            rng_util::distribution_fingerprint(distribution).wrapping_add(seq),
+        ));
+
+        // Fall back to random sampling during startup phase
+        let n_complete = history
+            .iter()
+            .filter(|t| t.state == TrialState::Complete)
+            .count();
+        if n_complete < self.n_startup_trials {
+            return common::sample_random(&mut rng, distribution);
+        }
+
+        // Split trials into good (Pareto front) and bad (dominated)
+        let (good_trials, bad_trials) = Self::split_trials(history, directions);
+
+        if good_trials.is_empty() || bad_trials.is_empty() {
+            return common::sample_random(&mut rng, distribution);
+        }
+
+        match distribution {
+            Distribution::Float(d) => self.sample_float(d, &good_trials, &bad_trials, &mut rng),
+            Distribution::Int(d) => self.sample_int(d, &good_trials, &bad_trials, &mut rng),
+            Distribution::Categorical(d) => {
+                self.sample_categorical(d, &good_trials, &bad_trials, &mut rng)
+            }
+        }
+    }
+}
+
+/// Builder for configuring a [`MotpeSampler`].
+///
+/// # Defaults
+///
+/// - `n_startup_trials`: 11
+/// - `n_ei_candidates`: 24
+/// - `kde_bandwidth`: None (Scott's rule)
+/// - `seed`: None (OS entropy)
+///
+/// # Examples
+///
+/// ```
+/// use optimizer::sampler::motpe::MotpeSamplerBuilder;
+///
+/// let sampler = MotpeSamplerBuilder::new()
+///     .n_startup_trials(15)
+///     .n_ei_candidates(32)
+///     .seed(42)
+///     .build();
+/// ```
+#[derive(Debug, Clone)]
+pub struct MotpeSamplerBuilder {
+    n_startup_trials: usize,
+    n_ei_candidates: usize,
+    kde_bandwidth: Option<f64>,
+    seed: Option<u64>,
+}
+
+impl MotpeSamplerBuilder {
+    /// Creates a new builder with default settings.
+    #[must_use]
+    pub fn new() -> Self {
+        Self {
+            n_startup_trials: 11,
+            n_ei_candidates: 24,
+            kde_bandwidth: None,
+            seed: None,
+        }
+    }
+
+    /// Sets the number of startup trials before MOTPE sampling begins.
+    #[must_use]
+    pub fn n_startup_trials(mut self, n: usize) -> Self {
+        self.n_startup_trials = n;
+        self
+    }
+
+    /// Sets the number of EI candidates to evaluate per sample.
+    #[must_use]
+    pub fn n_ei_candidates(mut self, n: usize) -> Self {
+        self.n_ei_candidates = n;
+        self
+    }
+
+    /// Sets a fixed bandwidth for the kernel density estimator.
+    ///
+    /// By default, Scott's rule is used for automatic bandwidth selection.
+    #[must_use]
+    pub fn kde_bandwidth(mut self, bandwidth: f64) -> Self {
+        self.kde_bandwidth = Some(bandwidth);
+        self
+    }
+
+    /// Sets a seed for reproducible sampling.
+    #[must_use]
+    pub fn seed(mut self, seed: u64) -> Self {
+        self.seed = Some(seed);
+        self
+    }
+
+    /// Builds the configured [`MotpeSampler`].
+    #[must_use]
+    pub fn build(self) -> MotpeSampler {
+        MotpeSampler {
+            n_startup_trials: self.n_startup_trials,
+            n_ei_candidates: self.n_ei_candidates,
+            kde_bandwidth: self.kde_bandwidth,
+            seed: self.seed.unwrap_or_else(|| fastrand::u64(..)),
+            call_seq: AtomicU64::new(0),
+        }
+    }
+}
+
+impl Default for MotpeSamplerBuilder {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+#[cfg(test)]
+#[allow(
+    clippy::similar_names,
+    clippy::cast_sign_loss,
+    clippy::cast_precision_loss
+)]
+mod tests {
+    use std::collections::HashMap;
+
+    use super::*;
+    use crate::distribution::{CategoricalDistribution, FloatDistribution, IntDistribution};
+    use crate::parameter::ParamId;
+    use crate::trial::Trial;
+
+    fn create_mo_trial(
+        id: u64,
+        values: Vec<f64>,
+        params: Vec<(ParamId, ParamValue, Distribution)>,
+    ) -> MultiObjectiveTrial {
+        let mut param_map = HashMap::new();
+        let mut dist_map = HashMap::new();
+        let label_map = HashMap::new();
+        for (param_id, pv, dist) in params {
+            param_map.insert(param_id, pv);
+            dist_map.insert(param_id, dist);
+        }
+        MultiObjectiveTrial {
+            id,
+            params: param_map,
+            distributions: dist_map,
+            param_labels: label_map,
+            values,
+            state: TrialState::Complete,
+            user_attrs: HashMap::new(),
+            constraints: Vec::new(),
+        }
+    }
+
+    #[test]
+    fn test_motpe_startup_random_sampling() {
+        let sampler = MotpeSampler::with_seed(42);
+        let dist = Distribution::Float(FloatDistribution {
+            low: 0.0,
+            high: 1.0,
+            log_scale: false,
+            step: None,
+        });
+        let directions = [Direction::Minimize, Direction::Minimize];
+
+        // With no history, should use random sampling
+        let history: Vec<MultiObjectiveTrial> = vec![];
+        for i in 0..50 {
+            let value = sampler.sample(&dist, i, &history, &directions);
+            if let ParamValue::Float(v) = value {
+                assert!((0.0..=1.0).contains(&v));
+            } else {
+                panic!("Expected Float value");
+            }
+        }
+    }
+
+    #[test]
+    fn test_motpe_split_pareto() {
+        let dist = Distribution::Float(FloatDistribution {
+            low: 0.0,
+            high: 1.0,
+            log_scale: false,
+            step: None,
+        });
+        let directions = [Direction::Minimize, Direction::Minimize];
+        let x_id = ParamId::new();
+
+        // Create trials: Pareto front = {(0.1, 0.9), (0.5, 0.5), (0.9, 0.1)}
+        // Dominated = {(0.6, 0.8), (0.8, 0.7)}
+        let history = vec![
+            create_mo_trial(
+                0,
+                vec![0.1, 0.9],
+                vec![(x_id, ParamValue::Float(0.1), dist.clone())],
+            ),
+            create_mo_trial(
+                1,
+                vec![0.5, 0.5],
+                vec![(x_id, ParamValue::Float(0.5), dist.clone())],
+            ),
+            create_mo_trial(
+                2,
+                vec![0.9, 0.1],
+                vec![(x_id, ParamValue::Float(0.9), dist.clone())],
+            ),
+            create_mo_trial(
+                3,
+                vec![0.6, 0.8],
+                vec![(x_id, ParamValue::Float(0.6), dist.clone())],
+            ),
+            create_mo_trial(
+                4,
+                vec![0.8, 0.7],
+                vec![(x_id, ParamValue::Float(0.8), dist.clone())],
+            ),
+        ];
+
+        let (good, bad) = MotpeSampler::split_trials(&history, &directions);
+        assert_eq!(good.len(), 3, "Pareto front should have 3 members");
+        assert_eq!(bad.len(), 2, "2 dominated trials");
+    }
+
+    #[test]
+    fn test_motpe_samples_float() {
+        let sampler = MotpeSampler::builder()
+            .n_startup_trials(5)
+            .n_ei_candidates(24)
+            .seed(42)
+            .build();
+
+        let dist = Distribution::Float(FloatDistribution {
+            low: 0.0,
+            high: 1.0,
+            log_scale: false,
+            step: None,
+        });
+        let directions = [Direction::Minimize, Direction::Minimize];
+        let x_id = ParamId::new();
+
+        // Build history where values near 0.3 are on the Pareto front
+        let mut history = Vec::new();
+        for i in 0..20 {
+            let x = f64::from(i) / 20.0;
+            // Pareto front: f1 = (x - 0.3)^2, f2 = (x - 0.3)^2 + 0.1
+            // Best solutions cluster around x = 0.3
+            let f1 = (x - 0.3).powi(2);
+            let f2 = (x - 0.7).powi(2);
+            history.push(create_mo_trial(
+                i as u64,
+                vec![f1, f2],
+                vec![(x_id, ParamValue::Float(x), dist.clone())],
+            ));
+        }
+
+        // MOTPE should produce values within [0, 1]
+        for i in 0..50 {
+            let value = sampler.sample(&dist, 100 + i, &history, &directions);
+            if let ParamValue::Float(v) = value {
+                assert!((0.0..=1.0).contains(&v), "Value {v} out of range");
+            } else {
+                panic!("Expected Float value");
+            }
+        }
+    }
+
+    #[test]
+    fn test_motpe_int_sampling() {
+        let sampler = MotpeSampler::builder().n_startup_trials(5).seed(42).build();
+
+        let dist = Distribution::Int(IntDistribution {
+            low: 0,
+            high: 100,
+            log_scale: false,
+            step: None,
+        });
+        let directions = [Direction::Minimize, Direction::Minimize];
+        let x_id = ParamId::new();
+
+        let mut history = Vec::new();
+        for i in 0..20 {
+            let x = i * 5;
+            let f1 = ((x as f64) - 30.0).powi(2);
+            let f2 = ((x as f64) - 70.0).powi(2);
+            history.push(create_mo_trial(
+                i as u64,
+                vec![f1, f2],
+                vec![(x_id, ParamValue::Int(x), dist.clone())],
+            ));
+        }
+
+        for i in 0..50 {
+            let value = sampler.sample(&dist, 100 + i, &history, &directions);
+            if let ParamValue::Int(v) = value {
+                assert!((0..=100).contains(&v), "Value {v} out of range");
+            } else {
+                panic!("Expected Int value");
+            }
+        }
+    }
+
+    #[test]
+    fn test_motpe_categorical_sampling() {
+        let sampler = MotpeSampler::builder().n_startup_trials(5).seed(42).build();
+
+        let dist = Distribution::Categorical(CategoricalDistribution { n_choices: 3 });
+        let directions = [Direction::Minimize, Direction::Minimize];
+        let cat_id = ParamId::new();
+
+        // Category 1 is on the Pareto front, others are dominated
+        let mut history = Vec::new();
+        for i in 0..15 {
+            let category = i % 3;
+            let (f1, f2) = match category {
+                0 => (0.8, 0.8), // dominated
+                1 => (0.1, 0.9), // Pareto front
+                2 => (0.9, 0.1), // Pareto front
+                _ => unreachable!(),
+            };
+            history.push(create_mo_trial(
+                i as u64,
+                vec![f1, f2],
+                vec![(
+                    cat_id,
+                    ParamValue::Categorical(category as usize),
+                    dist.clone(),
+                )],
+            ));
+        }
+
+        let mut counts = vec![0usize; 3];
+        for i in 0..200 {
+            let value = sampler.sample(&dist, 100 + i, &history, &directions);
+            if let ParamValue::Categorical(idx) = value {
+                assert!(idx < 3, "Category {idx} out of range");
+                counts[idx] += 1;
+            } else {
+                panic!("Expected Categorical value");
+            }
+        }
+
+        // Categories 1 and 2 (on Pareto front) should dominate category 0
+        assert!(
+            counts[1] + counts[2] > counts[0],
+            "Pareto-front categories should be sampled more: {counts:?}"
+        );
+    }
+
+    #[test]
+    fn test_motpe_reproducibility() {
+        let dist = Distribution::Float(FloatDistribution {
+            low: 0.0,
+            high: 1.0,
+            log_scale: false,
+            step: None,
+        });
+        let directions = [Direction::Minimize, Direction::Minimize];
+        let x_id = ParamId::new();
+
+        let history: Vec<MultiObjectiveTrial> = (0..20)
+            .map(|i| {
+                let x = f64::from(i) / 20.0;
+                create_mo_trial(
+                    i as u64,
+                    vec![x, 1.0 - x],
+                    vec![(x_id, ParamValue::Float(x), dist.clone())],
+                )
+            })
+            .collect();
+
+        let sampler1 = MotpeSampler::builder()
+            .seed(12345)
+            .n_startup_trials(5)
+            .build();
+        let sampler2 = MotpeSampler::builder()
+            .seed(12345)
+            .n_startup_trials(5)
+            .build();
+
+        for i in 0..10 {
+            let v1 = sampler1.sample(&dist, i, &history, &directions);
+            let v2 = sampler2.sample(&dist, i, &history, &directions);
+            assert_eq!(v1, v2, "Samples should be identical with same seed");
+        }
+    }
+
+    #[test]
+    fn test_motpe_with_study() {
+        use crate::multi_objective::MultiObjectiveStudy;
+        use crate::parameter::{FloatParam, Parameter};
+
+        let sampler = MotpeSampler::builder().seed(42).build();
+        let study = MultiObjectiveStudy::with_sampler(
+            vec![Direction::Minimize, Direction::Minimize],
+            sampler,
+        );
+
+        let x = FloatParam::new(0.0, 1.0);
+        study
+            .optimize(30, |trial: &mut Trial| {
+                let xv = x.suggest(trial)?;
+                Ok::<_, crate::Error>(vec![xv, 1.0 - xv])
+            })
+            .unwrap();
+
+        let front = study.pareto_front();
+        assert!(!front.is_empty(), "Should have Pareto-optimal solutions");
+
+        // All front solutions should have values summing to ~1.0
+        for trial in &front {
+            let sum: f64 = trial.values.iter().sum();
+            assert!(
+                (sum - 1.0).abs() < 0.01,
+                "Pareto front values should sum to ~1.0, got {sum}"
+            );
+        }
+    }
+
+    #[test]
+    fn test_motpe_builder_defaults() {
+        let sampler = MotpeSamplerBuilder::new().build();
+        assert_eq!(sampler.n_startup_trials, 11);
+        assert_eq!(sampler.n_ei_candidates, 24);
+        assert!(sampler.kde_bandwidth.is_none());
+    }
+
+    #[test]
+    fn test_motpe_builder_custom() {
+        let sampler = MotpeSamplerBuilder::new()
+            .n_startup_trials(20)
+            .n_ei_candidates(48)
+            .kde_bandwidth(0.5)
+            .seed(99)
+            .build();
+        assert_eq!(sampler.n_startup_trials, 20);
+        assert_eq!(sampler.n_ei_candidates, 48);
+        assert_eq!(sampler.kde_bandwidth, Some(0.5));
+    }
+}

src/sampler/nsga2.rs

@@ -0,0 +1,420 @@
+//! NSGA-II (Non-dominated Sorting Genetic Algorithm II) sampler.
+//!
+//! NSGA-II is one of the most widely used evolutionary multi-objective
+//! optimization algorithms. It ranks the population using **non-dominated
+//! sorting** (fast O(MN²) algorithm) and breaks ties within the same
+//! Pareto front using **crowding distance**, which favors solutions in
+//! less-crowded regions of the objective space.
+//!
+//! # Algorithm
+//!
+//! Each generation proceeds as follows:
+//!
+//! 1. **Non-dominated sorting** — partition the combined parent+offspring
+//!    population into Pareto fronts F₁, F₂, …
+//! 2. **Crowding distance** — for each front, compute per-solution crowding
+//!    distance (sum of normalized neighbor gaps in each objective).
+//! 3. **Selection** — fill the next population front-by-front. When a front
+//!    only partially fits, prefer solutions with higher crowding distance.
+//! 4. **Binary tournament** — select parents using (rank, crowding distance)
+//!    comparisons.
+//! 5. **SBX crossover + polynomial mutation** — generate offspring.
+//!
+//! # When to use
+//!
+//! - Two-objective problems where you want a well-spread Pareto front.
+//! - General-purpose multi-objective optimization with moderate population
+//!   sizes.
+//! - Problems that benefit from diversity preservation via crowding distance.
+//!
+//! For problems with **three or more objectives**, consider
+//! [`Nsga3Sampler`](super::nsga3::Nsga3Sampler) (reference-point niching)
+//! or [`MoeadSampler`](super::moead::MoeadSampler) (decomposition).
+//!
+//! # Configuration
+//!
+//! | Parameter | Builder method | Default |
+//! |-----------|---------------|---------|
+//! | Population size | [`population_size`](Nsga2SamplerBuilder::population_size) | `4 + floor(3 * ln(n_params))`, min 4 |
+//! | Crossover probability | [`crossover_prob`](Nsga2SamplerBuilder::crossover_prob) | 0.9 |
+//! | SBX distribution index | [`crossover_eta`](Nsga2SamplerBuilder::crossover_eta) | 20.0 |
+//! | Mutation distribution index | [`mutation_eta`](Nsga2SamplerBuilder::mutation_eta) | 20.0 |
+//! | Random seed | [`seed`](Nsga2SamplerBuilder::seed) | random |
+//!
+//! # Examples
+//!
+//! ```
+//! use optimizer::Direction;
+//! use optimizer::multi_objective::MultiObjectiveStudy;
+//! use optimizer::parameter::{FloatParam, Parameter};
+//! use optimizer::sampler::nsga2::Nsga2Sampler;
+//!
+//! let sampler = Nsga2Sampler::with_seed(42);
+//! let study =
+//!     MultiObjectiveStudy::with_sampler(vec![Direction::Minimize, Direction::Minimize], sampler);
+//!
+//! let x = FloatParam::new(0.0, 1.0);
+//! study
+//!     .optimize(50, |trial: &mut optimizer::Trial| {
+//!         let xv = x.suggest(trial)?;
+//!         Ok::<_, optimizer::Error>(vec![xv * xv, (xv - 1.0).powi(2)])
+//!     })
+//!     .unwrap();
+//! ```
+
+use parking_lot::Mutex;
+
+use super::genetic::{
+    self, Candidate, EvolutionaryState, Phase, advance_generation, collect_evaluated_generation,
+    crossover, extract_trial_params, finalize_discovery, generate_random_candidates, mutate,
+    sample_from_candidate, sample_random,
+};
+use crate::distribution::Distribution;
+use crate::multi_objective::MultiObjectiveTrial;
+use crate::param::ParamValue;
+use crate::pareto;
+use crate::types::Direction;
+
+/// NSGA-II sampler for multi-objective optimization.
+///
+/// Use non-dominated sorting with crowding-distance tie-breaking to
+/// evolve a well-spread Pareto front. Best suited for bi-objective
+/// problems; for 3+ objectives prefer [`Nsga3Sampler`](super::nsga3::Nsga3Sampler).
+///
+/// Create with [`Nsga2Sampler::new`], [`Nsga2Sampler::with_seed`], or
+/// [`Nsga2Sampler::builder`] for full configuration.
+pub struct Nsga2Sampler {
+    state: Mutex<Nsga2State>,
+}
+
+impl Nsga2Sampler {
+    /// Creates a new NSGA-II sampler with a random seed.
+    #[must_use]
+    pub fn new() -> Self {
+        Self {
+            state: Mutex::new(Nsga2State::new(Nsga2Config::default(), None)),
+        }
+    }
+
+    /// Creates a new NSGA-II sampler with a fixed seed.
+    #[must_use]
+    pub fn with_seed(seed: u64) -> Self {
+        Self {
+            state: Mutex::new(Nsga2State::new(Nsga2Config::default(), Some(seed))),
+        }
+    }
+
+    /// Creates a builder for configuring an `Nsga2Sampler`.
+    #[must_use]
+    pub fn builder() -> Nsga2SamplerBuilder {
+        Nsga2SamplerBuilder::default()
+    }
+}
+
+impl Default for Nsga2Sampler {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+/// Builder for [`Nsga2Sampler`].
+#[derive(Debug, Clone, Default)]
+pub struct Nsga2SamplerBuilder {
+    population_size: Option<usize>,
+    crossover_prob: Option<f64>,
+    crossover_eta: Option<f64>,
+    mutation_eta: Option<f64>,
+    seed: Option<u64>,
+}
+
+impl Nsga2SamplerBuilder {
+    /// Sets the population size. Default: `4 + floor(3 * ln(n_params))`, minimum 4.
+    #[must_use]
+    pub fn population_size(mut self, size: usize) -> Self {
+        self.population_size = Some(size);
+        self
+    }
+
+    /// Sets the crossover probability. Default: 0.9.
+    #[must_use]
+    pub fn crossover_prob(mut self, prob: f64) -> Self {
+        self.crossover_prob = Some(prob);
+        self
+    }
+
+    /// Sets the SBX distribution index. Default: 20.0.
+    #[must_use]
+    pub fn crossover_eta(mut self, eta: f64) -> Self {
+        self.crossover_eta = Some(eta);
+        self
+    }
+
+    /// Sets the polynomial mutation distribution index. Default: 20.0.
+    #[must_use]
+    pub fn mutation_eta(mut self, eta: f64) -> Self {
+        self.mutation_eta = Some(eta);
+        self
+    }
+
+    /// Sets the random seed for reproducibility.
+    #[must_use]
+    pub fn seed(mut self, seed: u64) -> Self {
+        self.seed = Some(seed);
+        self
+    }
+
+    /// Builds the configured [`Nsga2Sampler`].
+    #[must_use]
+    pub fn build(self) -> Nsga2Sampler {
+        let config = Nsga2Config {
+            user_population_size: self.population_size,
+            crossover_prob: self.crossover_prob.unwrap_or(0.9),
+            crossover_eta: self.crossover_eta.unwrap_or(20.0),
+            mutation_eta: self.mutation_eta.unwrap_or(20.0),
+        };
+        Nsga2Sampler {
+            state: Mutex::new(Nsga2State::new(config, self.seed)),
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Internal types
+// ---------------------------------------------------------------------------
+
+#[derive(Clone, Debug)]
+struct Nsga2Config {
+    user_population_size: Option<usize>,
+    crossover_prob: f64,
+    crossover_eta: f64,
+    mutation_eta: f64,
+}
+
+impl Default for Nsga2Config {
+    fn default() -> Self {
+        Self {
+            user_population_size: None,
+            crossover_prob: 0.9,
+            crossover_eta: 20.0,
+            mutation_eta: 20.0,
+        }
+    }
+}
+
+struct Nsga2State {
+    evo: EvolutionaryState,
+    config: Nsga2Config,
+}
+
+impl Nsga2State {
+    fn new(config: Nsga2Config, seed: Option<u64>) -> Self {
+        Self {
+            evo: EvolutionaryState::new(seed),
+            config,
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// MultiObjectiveSampler implementation
+// ---------------------------------------------------------------------------
+
+impl crate::multi_objective::MultiObjectiveSampler for Nsga2Sampler {
+    fn sample(
+        &self,
+        distribution: &Distribution,
+        trial_id: u64,
+        history: &[MultiObjectiveTrial],
+        directions: &[Direction],
+    ) -> ParamValue {
+        let mut state = self.state.lock();
+
+        match &state.evo.phase {
+            Phase::Discovery => {
+                if let Some(value) =
+                    genetic::sample_discovery(&mut state.evo, distribution, trial_id)
+                {
+                    return value;
+                }
+                // Transitioned to active phase
+                let user_pop = state.config.user_population_size;
+                finalize_discovery(&mut state.evo, user_pop);
+                generate_random_candidates(&mut state.evo);
+                sample_from_candidate(&mut state.evo, trial_id)
+            }
+            Phase::Active => {
+                maybe_generate_new_generation(&mut state, history, directions);
+                sample_from_candidate(&mut state.evo, trial_id)
+            }
+        }
+    }
+}
+
+/// Check if all candidates in the current generation have been evaluated;
+/// if so, run NSGA-II selection and generate offspring.
+fn maybe_generate_new_generation(
+    state: &mut Nsga2State,
+    history: &[MultiObjectiveTrial],
+    directions: &[Direction],
+) {
+    if state.evo.candidates.is_empty() {
+        generate_random_candidates(&mut state.evo);
+        return;
+    }
+
+    if let Some(evaluated) = collect_evaluated_generation(&state.evo, history) {
+        let offspring = nsga2_generate_offspring(state, &evaluated, directions);
+        advance_generation(&mut state.evo, offspring);
+    }
+}
+
+// ---------------------------------------------------------------------------
+// NSGA-II generation algorithm
+// ---------------------------------------------------------------------------
+
+/// Performs NSGA-II selection: non-dominated sort + crowding distance,
+/// then selects `pop_size` parents from the population.
+fn nsga2_select(
+    state: &mut Nsga2State,
+    population: &[&MultiObjectiveTrial],
+    directions: &[Direction],
+) -> (Vec<Vec<ParamValue>>, Vec<usize>, Vec<f64>) {
+    let pop_size = state.evo.population_size;
+
+    let values: Vec<Vec<f64>> = population.iter().map(|t| t.values.clone()).collect();
+    let constraints: Vec<Vec<f64>> = population.iter().map(|t| t.constraints.clone()).collect();
+    let has_constraints = constraints.iter().any(|c| !c.is_empty());
+
+    let fronts = if has_constraints {
+        pareto::fast_non_dominated_sort_constrained(&values, directions, &constraints)
+    } else {
+        pareto::fast_non_dominated_sort(&values, directions)
+    };
+
+    let n = population.len();
+    let mut rank = vec![0_usize; n];
+    let mut crowding = vec![0.0_f64; n];
+
+    for (front_rank, front) in fronts.iter().enumerate() {
+        let cd = pareto::crowding_distance_indexed(front, &values);
+        for (i, &idx) in front.iter().enumerate() {
+            rank[idx] = front_rank;
+            crowding[idx] = cd[i];
+        }
+    }
+
+    let mut selected: Vec<usize> = Vec::with_capacity(pop_size);
+    for front in &fronts {
+        if selected.len() + front.len() <= pop_size {
+            selected.extend_from_slice(front);
+        } else {
+            let remaining = pop_size - selected.len();
+            let mut front_sorted: Vec<usize> = front.clone();
+            front_sorted.sort_by(|&a, &b| {
+                crowding[b]
+                    .partial_cmp(&crowding[a])
+                    .unwrap_or(core::cmp::Ordering::Equal)
+            });
+            selected.extend_from_slice(&front_sorted[..remaining]);
+            break;
+        }
+    }
+
+    while selected.len() < pop_size {
+        selected.push(state.evo.rng.usize(0..n));
+    }
+
+    let parents: Vec<Vec<ParamValue>> = selected
+        .iter()
+        .map(|&idx| {
+            extract_trial_params(population[idx], &state.evo.dimensions, &mut state.evo.rng)
+        })
+        .collect();
+
+    let sel_rank: Vec<usize> = selected.iter().map(|&i| rank[i]).collect();
+    let sel_crowding: Vec<f64> = selected.iter().map(|&i| crowding[i]).collect();
+
+    (parents, sel_rank, sel_crowding)
+}
+
+/// Runs NSGA-II selection and generates offspring candidates.
+fn nsga2_generate_offspring(
+    state: &mut Nsga2State,
+    population: &[&MultiObjectiveTrial],
+    directions: &[Direction],
+) -> Vec<Candidate> {
+    let pop_size = state.evo.population_size;
+
+    if population.len() < 2 {
+        return (0..pop_size)
+            .map(|_| {
+                let params = state
+                    .evo
+                    .dimensions
+                    .iter()
+                    .map(|d| sample_random(&mut state.evo.rng, &d.distribution))
+                    .collect();
+                Candidate { params }
+            })
+            .collect();
+    }
+
+    let (parents, sel_rank, sel_crowding) = nsga2_select(state, population, directions);
+
+    let mut offspring = Vec::with_capacity(pop_size);
+    while offspring.len() < pop_size {
+        let p1 = tournament_select(&mut state.evo.rng, &sel_rank, &sel_crowding, parents.len());
+        let p2 = tournament_select(&mut state.evo.rng, &sel_rank, &sel_crowding, parents.len());
+
+        let (mut child1, mut child2) = crossover(
+            &mut state.evo.rng,
+            &parents[p1],
+            &parents[p2],
+            &state.evo.dimensions,
+            state.config.crossover_prob,
+            state.config.crossover_eta,
+        );
+
+        mutate(
+            &mut state.evo.rng,
+            &mut child1,
+            &state.evo.dimensions,
+            state.config.mutation_eta,
+        );
+        mutate(
+            &mut state.evo.rng,
+            &mut child2,
+            &state.evo.dimensions,
+            state.config.mutation_eta,
+        );
+
+        offspring.push(Candidate { params: child1 });
+        if offspring.len() < pop_size {
+            offspring.push(Candidate { params: child2 });
+        }
+    }
+
+    offspring
+}
+
+/// Tournament selection: pick 2 random individuals, return index of winner.
+/// Winner has lower rank; ties broken by higher crowding distance.
+fn tournament_select(
+    rng: &mut fastrand::Rng,
+    ranks: &[usize],
+    crowding: &[f64],
+    n: usize,
+) -> usize {
+    let a = rng.usize(0..n);
+    let b = rng.usize(0..n);
+
+    if ranks[a] < ranks[b] {
+        a
+    } else if ranks[b] < ranks[a] {
+        b
+    } else if crowding[a] >= crowding[b] {
+        a
+    } else {
+        b
+    }
+}

src/sampler/nsga3.rs

@@ -0,0 +1,773 @@
+//! NSGA-III (Non-dominated Sorting Genetic Algorithm III) sampler.
+//!
+//! NSGA-III extends NSGA-II to handle **many-objective** (3+) problems
+//! where crowding distance loses effectiveness. Instead of crowding
+//! distance, it uses **reference-point-based niching** with structured
+//! Das-Dennis reference points distributed on the unit simplex to guide
+//! the population toward a well-diversified Pareto front.
+//!
+//! # Algorithm
+//!
+//! Each generation proceeds as follows:
+//!
+//! 1. **Non-dominated sorting** — same as NSGA-II, partition the
+//!    combined population into Pareto fronts F₁, F₂, …
+//! 2. **Normalize objectives** — translate by ideal point and scale by
+//!    intercepts so all objectives lie in roughly \[0, 1\].
+//! 3. **Associate with reference points** — assign each solution to the
+//!    closest Das-Dennis reference direction by perpendicular distance.
+//! 4. **Niching selection** — when the last front only partially fits,
+//!    prefer solutions associated with under-represented reference points
+//!    (lowest niche count first, closest distance second).
+//! 5. **SBX crossover + polynomial mutation** — generate offspring via
+//!    rank-based tournament selection.
+//!
+//! # When to use
+//!
+//! - **Three or more objectives** — NSGA-III maintains diversity far
+//!   better than NSGA-II as the number of objectives grows.
+//! - Problems where you want a **uniformly distributed** Pareto front
+//!   guided by structured reference points.
+//! - Scales well up to ~10 objectives with appropriate division settings.
+//!
+//! For bi-objective problems, [`Nsga2Sampler`](super::nsga2::Nsga2Sampler)
+//! is simpler and equally effective. For decomposition-based optimization,
+//! see [`MoeadSampler`](super::moead::MoeadSampler).
+//!
+//! # Configuration
+//!
+//! | Parameter | Builder method | Default |
+//! |-----------|---------------|---------|
+//! | Population size | [`population_size`](Nsga3SamplerBuilder::population_size) | Number of Das-Dennis reference points |
+//! | Das-Dennis divisions (H) | [`n_divisions`](Nsga3SamplerBuilder::n_divisions) | Auto-chosen from population size and objectives |
+//! | Crossover probability | [`crossover_prob`](Nsga3SamplerBuilder::crossover_prob) | 1.0 |
+//! | SBX distribution index | [`crossover_eta`](Nsga3SamplerBuilder::crossover_eta) | 30.0 |
+//! | Mutation distribution index | [`mutation_eta`](Nsga3SamplerBuilder::mutation_eta) | 20.0 |
+//! | Random seed | [`seed`](Nsga3SamplerBuilder::seed) | random |
+//!
+//! # Examples
+//!
+//! ```
+//! use optimizer::Direction;
+//! use optimizer::multi_objective::MultiObjectiveStudy;
+//! use optimizer::parameter::{FloatParam, Parameter};
+//! use optimizer::sampler::nsga3::Nsga3Sampler;
+//!
+//! let sampler = Nsga3Sampler::with_seed(42);
+//! let study = MultiObjectiveStudy::with_sampler(
+//!     vec![
+//!         Direction::Minimize,
+//!         Direction::Minimize,
+//!         Direction::Minimize,
+//!     ],
+//!     sampler,
+//! );
+//!
+//! let x = FloatParam::new(0.0, 1.0);
+//! let y = FloatParam::new(0.0, 1.0);
+//! study
+//!     .optimize(100, |trial: &mut optimizer::Trial| {
+//!         let xv = x.suggest(trial)?;
+//!         let yv = y.suggest(trial)?;
+//!         Ok::<_, optimizer::Error>(vec![xv, yv, (1.0 - xv - yv).abs()])
+//!     })
+//!     .unwrap();
+//! ```
+
+use parking_lot::Mutex;
+
+use super::genetic::{
+    self, Candidate, EvolutionaryState, Phase, advance_generation, auto_divisions,
+    collect_evaluated_generation, crossover, das_dennis, extract_trial_params,
+    generate_random_candidates, mutate, sample_from_candidate, sample_random,
+};
+use crate::distribution::Distribution;
+use crate::multi_objective::MultiObjectiveTrial;
+use crate::param::ParamValue;
+use crate::pareto;
+use crate::types::Direction;
+
+/// NSGA-III sampler for multi-objective optimization.
+///
+/// Use reference-point niching with Das-Dennis structured points to
+/// maintain diversity in many-objective (3+) problems. For bi-objective
+/// problems, [`Nsga2Sampler`](super::nsga2::Nsga2Sampler) is simpler.
+///
+/// Create with [`Nsga3Sampler::new`], [`Nsga3Sampler::with_seed`], or
+/// [`Nsga3Sampler::builder`] for full configuration.
+pub struct Nsga3Sampler {
+    state: Mutex<Nsga3State>,
+}
+
+impl Nsga3Sampler {
+    /// Creates a new NSGA-III sampler with a random seed.
+    #[must_use]
+    pub fn new() -> Self {
+        Self {
+            state: Mutex::new(Nsga3State::new(Nsga3Config::default(), None)),
+        }
+    }
+
+    /// Creates a new NSGA-III sampler with a fixed seed.
+    #[must_use]
+    pub fn with_seed(seed: u64) -> Self {
+        Self {
+            state: Mutex::new(Nsga3State::new(Nsga3Config::default(), Some(seed))),
+        }
+    }
+
+    /// Creates a builder for configuring an `Nsga3Sampler`.
+    #[must_use]
+    pub fn builder() -> Nsga3SamplerBuilder {
+        Nsga3SamplerBuilder::default()
+    }
+}
+
+impl Default for Nsga3Sampler {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+/// Builder for [`Nsga3Sampler`].
+#[derive(Debug, Clone, Default)]
+pub struct Nsga3SamplerBuilder {
+    population_size: Option<usize>,
+    n_divisions: Option<usize>,
+    crossover_prob: Option<f64>,
+    crossover_eta: Option<f64>,
+    mutation_eta: Option<f64>,
+    seed: Option<u64>,
+}
+
+impl Nsga3SamplerBuilder {
+    /// Sets the population size. If unset, equals the number of
+    /// Das-Dennis reference points.
+    #[must_use]
+    pub fn population_size(mut self, size: usize) -> Self {
+        self.population_size = Some(size);
+        self
+    }
+
+    /// Sets the number of divisions (H) for Das-Dennis reference points.
+    /// If unset, automatically chosen based on population size and number
+    /// of objectives.
+    #[must_use]
+    pub fn n_divisions(mut self, h: usize) -> Self {
+        self.n_divisions = Some(h);
+        self
+    }
+
+    /// Sets the crossover probability. Default: 1.0.
+    #[must_use]
+    pub fn crossover_prob(mut self, prob: f64) -> Self {
+        self.crossover_prob = Some(prob);
+        self
+    }
+
+    /// Sets the SBX distribution index. Default: 30.0.
+    #[must_use]
+    pub fn crossover_eta(mut self, eta: f64) -> Self {
+        self.crossover_eta = Some(eta);
+        self
+    }
+
+    /// Sets the polynomial mutation distribution index. Default: 20.0.
+    #[must_use]
+    pub fn mutation_eta(mut self, eta: f64) -> Self {
+        self.mutation_eta = Some(eta);
+        self
+    }
+
+    /// Sets the random seed for reproducibility.
+    #[must_use]
+    pub fn seed(mut self, seed: u64) -> Self {
+        self.seed = Some(seed);
+        self
+    }
+
+    /// Builds the configured [`Nsga3Sampler`].
+    #[must_use]
+    pub fn build(self) -> Nsga3Sampler {
+        let config = Nsga3Config {
+            user_population_size: self.population_size,
+            n_divisions: self.n_divisions,
+            crossover_prob: self.crossover_prob.unwrap_or(1.0),
+            crossover_eta: self.crossover_eta.unwrap_or(30.0),
+            mutation_eta: self.mutation_eta.unwrap_or(20.0),
+        };
+        Nsga3Sampler {
+            state: Mutex::new(Nsga3State::new(config, self.seed)),
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Internal types
+// ---------------------------------------------------------------------------
+
+#[derive(Clone, Debug)]
+struct Nsga3Config {
+    user_population_size: Option<usize>,
+    n_divisions: Option<usize>,
+    crossover_prob: f64,
+    crossover_eta: f64,
+    mutation_eta: f64,
+}
+
+impl Default for Nsga3Config {
+    fn default() -> Self {
+        Self {
+            user_population_size: None,
+            n_divisions: None,
+            crossover_prob: 1.0,
+            crossover_eta: 30.0,
+            mutation_eta: 20.0,
+        }
+    }
+}
+
+struct Nsga3State {
+    evo: EvolutionaryState,
+    config: Nsga3Config,
+    /// Das-Dennis reference points (lazily generated once objectives are known).
+    reference_points: Vec<Vec<f64>>,
+    /// Best value seen per objective (minimize-space).
+    ideal_point: Vec<f64>,
+    /// Whether reference points have been initialized.
+    initialized: bool,
+}
+
+impl Nsga3State {
+    fn new(config: Nsga3Config, seed: Option<u64>) -> Self {
+        Self {
+            evo: EvolutionaryState::new(seed),
+            config,
+            reference_points: Vec::new(),
+            ideal_point: Vec::new(),
+            initialized: false,
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// MultiObjectiveSampler implementation
+// ---------------------------------------------------------------------------
+
+impl crate::multi_objective::MultiObjectiveSampler for Nsga3Sampler {
+    fn sample(
+        &self,
+        distribution: &Distribution,
+        trial_id: u64,
+        history: &[MultiObjectiveTrial],
+        directions: &[Direction],
+    ) -> ParamValue {
+        let mut state = self.state.lock();
+
+        match &state.evo.phase {
+            Phase::Discovery => {
+                if let Some(value) =
+                    genetic::sample_discovery(&mut state.evo, distribution, trial_id)
+                {
+                    return value;
+                }
+                // Transitioned to active phase
+                initialize_nsga3(&mut state, directions);
+                generate_random_candidates(&mut state.evo);
+                sample_from_candidate(&mut state.evo, trial_id)
+            }
+            Phase::Active => {
+                maybe_generate_new_generation(&mut state, history, directions);
+                sample_from_candidate(&mut state.evo, trial_id)
+            }
+        }
+    }
+}
+
+/// Initialize NSGA-III: generate reference points and set population size.
+fn initialize_nsga3(state: &mut Nsga3State, directions: &[Direction]) {
+    let n_obj = directions.len();
+
+    // Determine divisions
+    let divisions = state
+        .config
+        .n_divisions
+        .unwrap_or_else(|| auto_divisions(n_obj, state.config.user_population_size.unwrap_or(100)));
+
+    state.reference_points = das_dennis(n_obj, divisions);
+    let n_ref = state.reference_points.len();
+
+    // Population size = number of reference points (or user override, at least n_ref)
+    let pop_size = state.config.user_population_size.unwrap_or(n_ref).max(4);
+    state.evo.population_size = pop_size;
+    state.evo.phase = Phase::Active;
+    state.ideal_point = vec![f64::INFINITY; n_obj];
+    state.initialized = true;
+}
+
+fn maybe_generate_new_generation(
+    state: &mut Nsga3State,
+    history: &[MultiObjectiveTrial],
+    directions: &[Direction],
+) {
+    if state.evo.candidates.is_empty() {
+        generate_random_candidates(&mut state.evo);
+        return;
+    }
+
+    if let Some(evaluated) = collect_evaluated_generation(&state.evo, history) {
+        let offspring = nsga3_generate_offspring(state, &evaluated, directions);
+        advance_generation(&mut state.evo, offspring);
+    }
+}
+
+// ---------------------------------------------------------------------------
+// NSGA-III selection algorithm
+// ---------------------------------------------------------------------------
+
+/// Normalize objectives to minimize-space.
+fn to_minimize_space(values: &[f64], directions: &[Direction]) -> Vec<f64> {
+    values
+        .iter()
+        .zip(directions)
+        .map(|(&v, d)| match d {
+            Direction::Minimize => v,
+            Direction::Maximize => -v,
+        })
+        .collect()
+}
+
+/// Update ideal point with new observations.
+fn update_ideal_point(ideal: &mut [f64], normalized_values: &[Vec<f64>]) {
+    for vals in normalized_values {
+        for (i, &v) in vals.iter().enumerate() {
+            if v < ideal[i] {
+                ideal[i] = v;
+            }
+        }
+    }
+}
+
+/// Compute Achievement Scalarizing Function (ASF) for extreme point finding.
+fn asf(point: &[f64], weight: &[f64], ideal: &[f64]) -> f64 {
+    point
+        .iter()
+        .zip(weight)
+        .zip(ideal)
+        .map(|((&p, &w), &z)| {
+            let w = if w < 1e-6 { 1e-6 } else { w };
+            (p - z) / w
+        })
+        .fold(f64::NEG_INFINITY, f64::max)
+}
+
+/// Find intercepts for normalization via extreme points.
+///
+/// For each objective, find the point with best ASF (using a weight vector
+/// that emphasizes that objective). The intercepts are where the hyperplane
+/// through the extreme points crosses each axis.
+fn find_intercepts(normalized_values: &[Vec<f64>], ideal: &[f64]) -> Vec<f64> {
+    let n_obj = ideal.len();
+    let n = normalized_values.len();
+
+    if n == 0 || n_obj == 0 {
+        return vec![1.0; n_obj];
+    }
+
+    // Find extreme points (one per objective)
+    let mut extreme_indices = Vec::with_capacity(n_obj);
+    for obj in 0..n_obj {
+        let mut weight = vec![1e-6; n_obj];
+        weight[obj] = 1.0;
+
+        let mut best_idx = 0;
+        let mut best_asf = f64::INFINITY;
+        for (i, vals) in normalized_values.iter().enumerate() {
+            let a = asf(vals, &weight, ideal);
+            if a < best_asf {
+                best_asf = a;
+                best_idx = i;
+            }
+        }
+        extreme_indices.push(best_idx);
+    }
+
+    // Try to compute hyperplane intercepts
+    // For stability, if the extreme points are degenerate, fall back to
+    // max - ideal per objective
+    let mut intercepts = Vec::with_capacity(n_obj);
+    for obj in 0..n_obj {
+        let max_val = normalized_values
+            .iter()
+            .map(|v| v[obj])
+            .fold(f64::NEG_INFINITY, f64::max);
+        let intercept = max_val - ideal[obj];
+        intercepts.push(if intercept > 1e-10 { intercept } else { 1.0 });
+    }
+
+    intercepts
+}
+
+/// Normalize objective values: subtract ideal, divide by intercepts.
+fn normalize_objectives(values: &[Vec<f64>], ideal: &[f64], intercepts: &[f64]) -> Vec<Vec<f64>> {
+    values
+        .iter()
+        .map(|v| {
+            v.iter()
+                .zip(ideal)
+                .zip(intercepts)
+                .map(|((&val, &z), &a)| {
+                    let norm = if a > 1e-10 { a } else { 1.0 };
+                    (val - z) / norm
+                })
+                .collect()
+        })
+        .collect()
+}
+
+/// Perpendicular distance from a point to a reference line (direction vector).
+fn perpendicular_distance(point: &[f64], reference: &[f64]) -> f64 {
+    let dot: f64 = point.iter().zip(reference).map(|(&p, &r)| p * r).sum();
+    let ref_norm_sq: f64 = reference.iter().map(|&r| r * r).sum();
+
+    if ref_norm_sq < 1e-30 {
+        return f64::INFINITY;
+    }
+
+    let proj_scalar = dot / ref_norm_sq;
+    let dist_sq: f64 = point
+        .iter()
+        .zip(reference)
+        .map(|(&p, &r)| {
+            let proj = proj_scalar * r;
+            (p - proj).powi(2)
+        })
+        .sum();
+
+    dist_sq.sqrt()
+}
+
+/// Associate each solution with its nearest reference point.
+/// Returns (`closest_ref_idx`, distance) for each solution.
+fn associate_to_reference_points(
+    normalized: &[Vec<f64>],
+    reference_points: &[Vec<f64>],
+) -> Vec<(usize, f64)> {
+    normalized
+        .iter()
+        .map(|point| {
+            let mut best_ref = 0;
+            let mut best_dist = f64::INFINITY;
+            for (j, rp) in reference_points.iter().enumerate() {
+                let d = perpendicular_distance(point, rp);
+                if d < best_dist {
+                    best_dist = d;
+                    best_ref = j;
+                }
+            }
+            (best_ref, best_dist)
+        })
+        .collect()
+}
+
+/// NSGA-III niching-based selection from the last front.
+///
+/// `already_selected` are indices into the combined population that are
+/// already accepted (from fronts 0..L-1). `last_front` contains indices
+/// from front L. We need to pick `remaining` more from `last_front`.
+fn niching_select(
+    rng: &mut fastrand::Rng,
+    associations: &[(usize, f64)],
+    already_selected: &[usize],
+    last_front: &[usize],
+    n_reference_points: usize,
+    remaining: usize,
+) -> Vec<usize> {
+    // Count niche per reference point for already selected
+    let mut niche_count = vec![0_usize; n_reference_points];
+    for &idx in already_selected {
+        niche_count[associations[idx].0] += 1;
+    }
+
+    // Build per-reference-point candidate lists from the last front
+    let mut ref_candidates: Vec<Vec<(usize, f64)>> = vec![Vec::new(); n_reference_points];
+    for &idx in last_front {
+        let (ref_idx, dist) = associations[idx];
+        ref_candidates[ref_idx].push((idx, dist));
+    }
+
+    let mut selected = Vec::with_capacity(remaining);
+    let mut excluded = vec![false; associations.len()];
+
+    for _ in 0..remaining {
+        // Find minimum niche count among reference points that still have candidates
+        let min_count = (0..n_reference_points)
+            .filter(|&j| ref_candidates[j].iter().any(|&(idx, _)| !excluded[idx]))
+            .map(|j| niche_count[j])
+            .min();
+
+        let Some(min_count) = min_count else {
+            break;
+        };
+
+        // Collect reference points with this minimum count that have candidates
+        let min_refs: Vec<usize> = (0..n_reference_points)
+            .filter(|&j| {
+                niche_count[j] == min_count
+                    && ref_candidates[j].iter().any(|&(idx, _)| !excluded[idx])
+            })
+            .collect();
+
+        if min_refs.is_empty() {
+            break;
+        }
+
+        // Pick a random reference point from the minimum set
+        let chosen_ref = min_refs[rng.usize(0..min_refs.len())];
+
+        // Available candidates for this reference point
+        let available: Vec<(usize, f64)> = ref_candidates[chosen_ref]
+            .iter()
+            .filter(|&&(idx, _)| !excluded[idx])
+            .copied()
+            .collect();
+
+        if available.is_empty() {
+            continue;
+        }
+
+        let chosen_idx = if min_count == 0 {
+            // Pick closest to reference line
+            available
+                .iter()
+                .min_by(|a, b| a.1.partial_cmp(&b.1).unwrap_or(core::cmp::Ordering::Equal))
+                .unwrap()
+                .0
+        } else {
+            // Pick random
+            available[rng.usize(0..available.len())].0
+        };
+
+        selected.push(chosen_idx);
+        excluded[chosen_idx] = true;
+        niche_count[chosen_ref] += 1;
+    }
+
+    selected
+}
+
+/// Perform NSGA-III selection: non-dominated sort + reference-point niching.
+fn nsga3_select(
+    state: &mut Nsga3State,
+    population: &[&MultiObjectiveTrial],
+    directions: &[Direction],
+) -> (Vec<Vec<ParamValue>>, Vec<usize>) {
+    let pop_size = state.evo.population_size;
+    let n_obj = directions.len();
+
+    // Convert to minimize-space
+    let min_values: Vec<Vec<f64>> = population
+        .iter()
+        .map(|t| to_minimize_space(&t.values, directions))
+        .collect();
+
+    // Non-dominated sort
+    let constraints: Vec<Vec<f64>> = population.iter().map(|t| t.constraints.clone()).collect();
+    let has_constraints = constraints.iter().any(|c| !c.is_empty());
+    let fronts = if has_constraints {
+        pareto::fast_non_dominated_sort_constrained(
+            &min_values,
+            &vec![Direction::Minimize; n_obj],
+            &constraints,
+        )
+    } else {
+        pareto::fast_non_dominated_sort(&min_values, &vec![Direction::Minimize; n_obj])
+    };
+
+    // Fill front-by-front
+    let mut selected: Vec<usize> = Vec::with_capacity(pop_size);
+    let mut last_front_idx = None;
+
+    for (fi, front) in fronts.iter().enumerate() {
+        if selected.len() + front.len() <= pop_size {
+            selected.extend_from_slice(front);
+        } else {
+            last_front_idx = Some(fi);
+            break;
+        }
+    }
+
+    // If we filled exactly or all fronts fit, done
+    if selected.len() < pop_size
+        && let Some(lf_idx) = last_front_idx
+    {
+        // Need niching from the last partial front
+        let remaining = pop_size - selected.len();
+
+        // Update ideal point
+        update_ideal_point(&mut state.ideal_point, &min_values);
+
+        // Find intercepts and normalize
+        let intercepts = find_intercepts(&min_values, &state.ideal_point);
+        let normalized = normalize_objectives(&min_values, &state.ideal_point, &intercepts);
+
+        // Associate all solutions with reference points
+        let associations = associate_to_reference_points(&normalized, &state.reference_points);
+
+        // Select from last front using niching
+        let last_front = &fronts[lf_idx];
+        let additional = niching_select(
+            &mut state.evo.rng,
+            &associations,
+            &selected,
+            last_front,
+            state.reference_points.len(),
+            remaining,
+        );
+        selected.extend(additional);
+    }
+
+    // Pad if needed
+    let n = population.len();
+    while selected.len() < pop_size {
+        selected.push(state.evo.rng.usize(0..n));
+    }
+
+    let params = selected
+        .iter()
+        .map(|&idx| {
+            extract_trial_params(population[idx], &state.evo.dimensions, &mut state.evo.rng)
+        })
+        .collect();
+    (params, selected)
+}
+
+/// Tournament selection based on rank only (no crowding distance in NSGA-III).
+fn tournament_select_rank(rng: &mut fastrand::Rng, ranks: &[usize], n: usize) -> usize {
+    let a = rng.usize(0..n);
+    let b = rng.usize(0..n);
+
+    if ranks[a] <= ranks[b] { a } else { b }
+}
+
+fn nsga3_generate_offspring(
+    state: &mut Nsga3State,
+    population: &[&MultiObjectiveTrial],
+    directions: &[Direction],
+) -> Vec<Candidate> {
+    let pop_size = state.evo.population_size;
+
+    if population.len() < 2 {
+        return (0..pop_size)
+            .map(|_| {
+                let params = state
+                    .evo
+                    .dimensions
+                    .iter()
+                    .map(|d| sample_random(&mut state.evo.rng, &d.distribution))
+                    .collect();
+                Candidate { params }
+            })
+            .collect();
+    }
+
+    // Initialize reference points and ideal on first generation
+    if !state.initialized {
+        initialize_nsga3(state, directions);
+    }
+
+    let (parents, selected_indices) = nsga3_select(state, population, directions);
+
+    // Assign Pareto front ranks for tournament selection
+    let n_obj = directions.len();
+    let min_values: Vec<Vec<f64>> = population
+        .iter()
+        .map(|t| to_minimize_space(&t.values, directions))
+        .collect();
+    let fronts = pareto::fast_non_dominated_sort(&min_values, &vec![Direction::Minimize; n_obj]);
+    // Build rank lookup for population indices
+    let mut pop_rank = vec![0_usize; population.len()];
+    for (front_rank, front) in fronts.iter().enumerate() {
+        for &idx in front {
+            if idx < pop_rank.len() {
+                pop_rank[idx] = front_rank;
+            }
+        }
+    }
+    // Map population ranks to selected parent indices
+    let parent_ranks: Vec<usize> = selected_indices
+        .iter()
+        .map(|&idx| {
+            if idx < pop_rank.len() {
+                pop_rank[idx]
+            } else {
+                0
+            }
+        })
+        .collect();
+
+    let mut offspring = Vec::with_capacity(pop_size);
+    while offspring.len() < pop_size {
+        let p1 = tournament_select_rank(&mut state.evo.rng, &parent_ranks, parents.len());
+        let p2 = tournament_select_rank(&mut state.evo.rng, &parent_ranks, parents.len());
+
+        let (mut child1, mut child2) = crossover(
+            &mut state.evo.rng,
+            &parents[p1],
+            &parents[p2],
+            &state.evo.dimensions,
+            state.config.crossover_prob,
+            state.config.crossover_eta,
+        );
+
+        mutate(
+            &mut state.evo.rng,
+            &mut child1,
+            &state.evo.dimensions,
+            state.config.mutation_eta,
+        );
+        mutate(
+            &mut state.evo.rng,
+            &mut child2,
+            &state.evo.dimensions,
+            state.config.mutation_eta,
+        );
+
+        offspring.push(Candidate { params: child1 });
+        if offspring.len() < pop_size {
+            offspring.push(Candidate { params: child2 });
+        }
+    }
+
+    offspring
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_perpendicular_distance() {
+        // Point (1, 0) to reference line (1, 1) (45-degree line)
+        let d = perpendicular_distance(&[1.0, 0.0], &[1.0, 1.0]);
+        // Projection is (0.5, 0.5), distance = sqrt(0.25 + 0.25) = sqrt(0.5)
+        assert!((d - (0.5_f64).sqrt()).abs() < 1e-10);
+    }
+
+    #[test]
+    fn test_perpendicular_distance_on_line() {
+        // Point on the reference line
+        let d = perpendicular_distance(&[2.0, 2.0], &[1.0, 1.0]);
+        assert!(d < 1e-10);
+    }
+
+    #[test]
+    fn test_normalize_objectives() {
+        let values = vec![vec![2.0, 4.0], vec![4.0, 2.0]];
+        let ideal = vec![1.0, 1.0];
+        let intercepts = vec![3.0, 3.0];
+        let normalized = normalize_objectives(&values, &ideal, &intercepts);
+        assert!((normalized[0][0] - 1.0 / 3.0).abs() < 1e-10);
+        assert!((normalized[0][1] - 1.0).abs() < 1e-10);
+    }
+}

src/sampler/random.rs

@@ -1,18 +1,50 @@
-//! Random sampler implementation.
+//! Random sampler — uniform independent sampling.
+//!
+//! [`RandomSampler`] draws each parameter value independently and uniformly
+//! at random, ignoring trial history entirely. It respects log-scale and
+//! step-size constraints defined by the parameter distribution.
+//!
+//! # When to use
+//!
+//! - **Baseline comparison** — run Random alongside smarter samplers to
+//!   quantify their benefit.
+//! - **Startup phase** — many model-based samplers (TPE, GP, CMA-ES) use
+//!   random sampling for their first *n* trials before fitting a surrogate.
+//! - **Very high dimensions** — when the search space is too large for
+//!   structured exploration, random search with enough budget can be
+//!   surprisingly competitive.
+//!
+//! For better uniform coverage without model fitting, consider
+//! `SobolSampler` (requires the `sobol`
+//! feature flag).
+//!
+//! # Example
+//!
+//! ```
+//! use optimizer::prelude::*;
+//! use optimizer::sampler::random::RandomSampler;
+//!
+//! let study: Study<f64> = Study::with_sampler(Direction::Minimize, RandomSampler::with_seed(42));
+//! ```
 
-use parking_lot::Mutex;
-use rand::rngs::StdRng;
-use rand::{RngExt, SeedableRng};
+use core::sync::atomic::{AtomicU64, Ordering};
 
 use crate::distribution::Distribution;
+use crate::multi_objective::{MultiObjectiveSampler, MultiObjectiveTrial};
 use crate::param::ParamValue;
+use crate::rng_util;
 use crate::sampler::{CompletedTrial, Sampler};
+use crate::types::Direction;
 
-/// A simple random sampler that samples uniformly from distributions.
+/// Uniform independent random sampler.
 ///
-/// This sampler ignores the trial history and samples uniformly at random,
-/// respecting log scale and step size constraints. It serves as a baseline
-/// sampler and is used during the startup phase of more sophisticated samplers.
+/// Sample each parameter value uniformly at random, respecting log-scale and
+/// step-size constraints. Trial history is ignored — every sample is drawn
+/// independently.
+///
+/// This is the default sampler used by [`Study::new`](crate::Study::new)
+/// and during the startup phase of model-based samplers such as
+/// [`TpeSampler`](super::tpe::TpeSampler).
 ///
 /// # Examples
 ///
@@ -26,7 +58,9 @@ use crate::sampler::{CompletedTrial, Sampler};
 /// let sampler = RandomSampler::with_seed(42);
 /// ```
 pub struct RandomSampler {
-    rng: Mutex<StdRng>,
+    seed: u64,
+    /// Monotonic counter to disambiguate calls with identical (`trial_id`, distribution).
+    call_seq: AtomicU64,
 }
 
 impl RandomSampler {
@@ -34,7 +68,8 @@ impl RandomSampler {
     #[must_use]
     pub fn new() -> Self {
         Self {
-            rng: Mutex::new(rand::make_rng()),
+            seed: fastrand::u64(..),
+            call_seq: AtomicU64::new(0),
         }
     }
 
@@ -44,11 +79,33 @@ impl RandomSampler {
     #[must_use]
     pub fn with_seed(seed: u64) -> Self {
         Self {
-            rng: Mutex::new(StdRng::seed_from_u64(seed)),
+            seed,
+            call_seq: AtomicU64::new(0),
         }
     }
 }
 
+/// Default multi-objective sampler that delegates to [`RandomSampler`].
+pub(crate) struct RandomMultiObjectiveSampler(RandomSampler);
+
+impl RandomMultiObjectiveSampler {
+    pub(crate) fn new() -> Self {
+        Self(RandomSampler::new())
+    }
+}
+
+impl MultiObjectiveSampler for RandomMultiObjectiveSampler {
+    fn sample(
+        &self,
+        distribution: &Distribution,
+        trial_id: u64,
+        _history: &[MultiObjectiveTrial],
+        _directions: &[Direction],
+    ) -> ParamValue {
+        self.0.sample(distribution, trial_id, &[])
+    }
+}
+
 impl Default for RandomSampler {
     fn default() -> Self {
         Self::new()
@@ -60,55 +117,17 @@ impl Sampler for RandomSampler {
     fn sample(
         &self,
         distribution: &Distribution,
-        _trial_id: u64,
+        trial_id: u64,
         _history: &[CompletedTrial],
     ) -> ParamValue {
-        let mut rng = self.rng.lock();
+        let seq = self.call_seq.fetch_add(1, Ordering::Relaxed);
+        let mut rng = fastrand::Rng::with_seed(rng_util::mix_seed(
+            self.seed,
+            trial_id,
+            rng_util::distribution_fingerprint(distribution).wrapping_add(seq),
+        ));
 
-        match distribution {
-            Distribution::Float(d) => {
-                let value = if d.log_scale {
-                    // Sample uniformly in log space
-                    let log_low = d.low.ln();
-                    let log_high = d.high.ln();
-                    let log_value = rng.random_range(log_low..=log_high);
-                    log_value.exp()
-                } else if let Some(step) = d.step {
-                    // Sample from step grid
-                    let n_steps = ((d.high - d.low) / step).floor() as i64;
-                    let k = rng.random_range(0..=n_steps);
-                    d.low + (k as f64) * step
-                } else {
-                    // Uniform sampling
-                    rng.random_range(d.low..=d.high)
-                };
-                ParamValue::Float(value)
-            }
-            Distribution::Int(d) => {
-                let value = if d.log_scale {
-                    // Sample uniformly in log space, then round
-                    let log_low = (d.low as f64).ln();
-                    let log_high = (d.high as f64).ln();
-                    let log_value = rng.random_range(log_low..=log_high);
-                    let raw = log_value.exp().round() as i64;
-                    // Clamp to bounds since rounding might push outside
-                    raw.clamp(d.low, d.high)
-                } else if let Some(step) = d.step {
-                    // Sample from step grid
-                    let n_steps = (d.high - d.low) / step;
-                    let k = rng.random_range(0..=n_steps);
-                    d.low + k * step
-                } else {
-                    // Uniform sampling
-                    rng.random_range(d.low..=d.high)
-                };
-                ParamValue::Int(value)
-            }
-            Distribution::Categorical(d) => {
-                let index = rng.random_range(0..d.n_choices);
-                ParamValue::Categorical(index)
-            }
-        }
+        super::common::sample_random(&mut rng, distribution)
     }
 }
 
@@ -128,8 +147,8 @@ mod tests {
             step: None,
         });
 
-        for _ in 0..100 {
-            let value = sampler.sample(&dist, 0, &[]);
+        for i in 0..100 {
+            let value = sampler.sample(&dist, i, &[]);
             if let ParamValue::Float(v) = value {
                 assert!((0.0..=1.0).contains(&v));
             } else {
@@ -148,8 +167,8 @@ mod tests {
             step: None,
         });
 
-        for _ in 0..100 {
-            let value = sampler.sample(&dist, 0, &[]);
+        for i in 0..100 {
+            let value = sampler.sample(&dist, i, &[]);
             if let ParamValue::Float(v) = value {
                 assert!((1e-5..=1.0).contains(&v));
             } else {
@@ -168,8 +187,8 @@ mod tests {
             step: Some(0.25),
         });
 
-        for _ in 0..100 {
-            let value = sampler.sample(&dist, 0, &[]);
+        for i in 0..100 {
+            let value = sampler.sample(&dist, i, &[]);
             if let ParamValue::Float(v) = value {
                 assert!((0.0..=1.0).contains(&v));
                 // Check it's on the step grid
@@ -192,8 +211,8 @@ mod tests {
             step: None,
         });
 
-        for _ in 0..100 {
-            let value = sampler.sample(&dist, 0, &[]);
+        for i in 0..100 {
+            let value = sampler.sample(&dist, i, &[]);
             if let ParamValue::Int(v) = value {
                 assert!((0..=10).contains(&v));
             } else {
@@ -212,8 +231,8 @@ mod tests {
             step: None,
         });
 
-        for _ in 0..100 {
-            let value = sampler.sample(&dist, 0, &[]);
+        for i in 0..100 {
+            let value = sampler.sample(&dist, i, &[]);
             if let ParamValue::Int(v) = value {
                 assert!((1..=1000).contains(&v));
             } else {
@@ -232,8 +251,8 @@ mod tests {
             step: Some(2),
         });
 
-        for _ in 0..100 {
-            let value = sampler.sample(&dist, 0, &[]);
+        for i in 0..100 {
+            let value = sampler.sample(&dist, i, &[]);
             if let ParamValue::Int(v) = value {
                 assert!((0..=10).contains(&v));
                 // Check it's on the step grid: 0, 2, 4, 6, 8, 10
@@ -249,8 +268,8 @@ mod tests {
         let sampler = RandomSampler::with_seed(42);
         let dist = Distribution::Categorical(CategoricalDistribution { n_choices: 5 });
 
-        for _ in 0..100 {
-            let value = sampler.sample(&dist, 0, &[]);
+        for i in 0..100 {
+            let value = sampler.sample(&dist, i, &[]);
             if let ParamValue::Categorical(idx) = value {
                 assert!(idx < 5);
             } else {
@@ -270,9 +289,9 @@ mod tests {
             step: None,
         });
 
-        for _ in 0..10 {
-            let v1 = sampler1.sample(&dist, 0, &[]);
-            let v2 = sampler2.sample(&dist, 0, &[]);
+        for i in 0..10 {
+            let v1 = sampler1.sample(&dist, i, &[]);
+            let v2 = sampler2.sample(&dist, i, &[]);
             assert_eq!(v1, v2);
         }
     }

src/sampler/sobol.rs

@@ -1,4 +1,50 @@
 //! Quasi-random sampler using Sobol low-discrepancy sequences.
+//!
+//! [`SobolSampler`] generates points from a Sobol sequence (scrambled via the
+//! Burley 2020 algorithm) to fill the parameter space more uniformly than
+//! pure random sampling. Where [`RandomSampler`](super::random::RandomSampler)
+//! may cluster points in some regions by chance, Sobol sequences are
+//! constructed to spread points evenly across all dimensions.
+//!
+//! # When to use
+//!
+//! - **Better-than-random baseline** — when you want uniform coverage
+//!   without the cost of model fitting (TPE, GP, etc.).
+//! - **Startup phase replacement** — use Sobol instead of random for the
+//!   initial exploration phase of adaptive samplers.
+//! - **Moderate dimensionality** — Sobol uniformity is strongest up to
+//!   ~20 dimensions; beyond that the advantage over random sampling
+//!   diminishes.
+//! - **Deterministic exploration** — Sobol sequences are fully deterministic
+//!   for a given seed, making experiments reproducible.
+//!
+//! # How it works
+//!
+//! Each trial maps to a Sobol sequence index, and each parameter within a
+//! trial maps to a separate Sobol dimension. The resulting quasi-random
+//! point in \[0, 1) is then scaled to the parameter's distribution (linear,
+//! log-scale, or step grid).
+//!
+//! **Important:** parameters must be suggested in the same order across
+//! trials for consistent dimension assignment.
+//!
+//! Requires the **`sobol`** feature flag:
+//!
+//! ```toml
+//! [dependencies]
+//! optimizer = { version = "...", features = ["sobol"] }
+//! ```
+//!
+//! # Example
+//!
+//! ```
+//! use optimizer::prelude::*;
+//! use optimizer::sampler::sobol::SobolSampler;
+//!
+//! let study: Study<f64> = Study::with_sampler(Direction::Minimize, SobolSampler::with_seed(42));
+//! ```
+
+use std::collections::HashMap;
 
 use parking_lot::Mutex;
 use sobol_burley::sample;
@@ -7,23 +53,19 @@ use crate::distribution::Distribution;
 use crate::param::ParamValue;
 use crate::sampler::{CompletedTrial, Sampler};
 
-/// Internal state for tracking the dimension counter within a trial.
+/// Internal state for tracking per-trial dimension counters.
 struct SobolState {
-    /// The `trial_id` of the current trial (used to reset dimension counter).
-    current_trial: u64,
-    /// Next Sobol dimension to use for the current trial.
-    next_dimension: u32,
+    /// Next Sobol dimension for each in-flight trial.
+    dimensions: HashMap<u64, u32>,
 }
 
 /// Quasi-random sampler using Sobol low-discrepancy sequences.
 ///
-/// Provides better uniform coverage of the parameter space than
-/// [`RandomSampler`](super::random::RandomSampler). Useful as a baseline or
-/// for the startup phase of model-based samplers.
-///
-/// Unlike random sampling, Sobol sequences are deterministic and fill the
-/// space more evenly, reducing the number of trials needed to adequately
-/// cover the search space.
+/// Produce better uniform coverage of the parameter space than
+/// [`RandomSampler`](super::random::RandomSampler) by using a
+/// scrambled Sobol sequence (Burley 2020). Useful as a standalone
+/// baseline or as a drop-in replacement for the random startup
+/// phase of model-based samplers.
 ///
 /// Each trial uses a different Sobol sequence index, and each parameter
 /// within a trial maps to a different Sobol dimension. Parameters must be
@@ -33,7 +75,7 @@ struct SobolState {
 /// Sobol sequences are most effective in moderate dimensions (up to ~20).
 /// For very high dimensions, the uniformity advantage diminishes.
 ///
-/// Requires the `sobol` feature flag.
+/// Requires the **`sobol`** feature flag.
 ///
 /// # Examples
 ///
@@ -65,8 +107,7 @@ impl SobolSampler {
         Self {
             seed: seed as u32,
             state: Mutex::new(SobolState {
-                current_trial: u64::MAX,
-                next_dimension: 0,
+                dimensions: HashMap::new(),
             }),
         }
     }
@@ -88,20 +129,15 @@ impl Sampler for SobolSampler {
     ) -> ParamValue {
         let mut state = self.state.lock();
 
-        // Reset dimension counter when a new trial starts.
-        if state.current_trial != trial_id {
-            state.current_trial = trial_id;
-            state.next_dimension = 0;
-        }
-
-        let dimension = state.next_dimension;
-        state.next_dimension = dimension + 1;
+        let dimension = state.dimensions.entry(trial_id).or_insert(0);
+        let dim = *dimension;
+        *dimension = dim + 1;
 
         // Use trial_id as the Sobol sequence index.
         let index = trial_id as u32;
 
         // Generate a quasi-random point in [0, 1).
-        let point = f64::from(sample(index, dimension, self.seed));
+        let point = f64::from(sample(index, dim, self.seed));
 
         map_point_to_distribution(point, distribution)
     }

src/sampler/tpe/common.rs

@@ -0,0 +1,216 @@
+//! Shared TPE sampling functions used by both `TpeSampler` and `MotpeSampler`.
+
+use crate::distribution::{FloatDistribution, IntDistribution};
+use crate::kde::KernelDensityEstimator;
+use crate::rng_util;
+
+/// Samples using TPE for float distributions.
+pub(crate) fn sample_tpe_float(
+    dist: &FloatDistribution,
+    good_values: Vec<f64>,
+    bad_values: Vec<f64>,
+    n_ei_candidates: usize,
+    kde_bandwidth: Option<f64>,
+    rng: &mut fastrand::Rng,
+) -> f64 {
+    let (low, high, log_scale, step) = (dist.low, dist.high, dist.log_scale, dist.step);
+
+    // Transform to internal space (log space if needed)
+    let (internal_low, internal_high, good_internal, bad_internal) = if log_scale {
+        let i_low = low.ln();
+        let i_high = high.ln();
+        let g = {
+            let mut v = good_values;
+            for x in &mut v {
+                *x = x.ln();
+            }
+            v
+        };
+        let b = {
+            let mut v = bad_values;
+            for x in &mut v {
+                *x = x.ln();
+            }
+            v
+        };
+        (i_low, i_high, g, b)
+    } else {
+        (low, high, good_values, bad_values)
+    };
+
+    // Fit KDEs to good and bad groups
+    let l_kde = match kde_bandwidth {
+        Some(bw) => KernelDensityEstimator::with_bandwidth(good_internal, bw),
+        None => KernelDensityEstimator::new(good_internal),
+    };
+    let g_kde = match kde_bandwidth {
+        Some(bw) => KernelDensityEstimator::with_bandwidth(bad_internal, bw),
+        None => KernelDensityEstimator::new(bad_internal),
+    };
+
+    // If KDE construction fails, fall back to uniform sampling
+    let (Ok(l_kde), Ok(g_kde)) = (l_kde, g_kde) else {
+        return rng_util::f64_range(rng, low, high);
+    };
+
+    // Generate candidates from l(x) and select the one with best l(x)/g(x) ratio
+    let mut best_candidate = internal_low;
+    let mut best_ratio = f64::NEG_INFINITY;
+
+    for _ in 0..n_ei_candidates {
+        let candidate = l_kde.sample(rng).clamp(internal_low, internal_high);
+
+        let l_density = l_kde.pdf(candidate);
+        let g_density = g_kde.pdf(candidate);
+
+        // Compute l(x)/g(x) ratio, handling zero density
+        let ratio = if g_density < f64::EPSILON {
+            if l_density > f64::EPSILON {
+                f64::INFINITY
+            } else {
+                0.0
+            }
+        } else {
+            l_density / g_density
+        };
+
+        if ratio > best_ratio {
+            best_ratio = ratio;
+            best_candidate = candidate;
+        }
+    }
+
+    // Transform back from internal space
+    let mut value = if log_scale {
+        best_candidate.exp()
+    } else {
+        best_candidate
+    };
+
+    // Apply step constraint if present
+    if let Some(step) = step {
+        let k = ((value - low) / step).round();
+        value = low + k * step;
+    }
+
+    // Ensure value is within bounds
+    value.clamp(low, high)
+}
+
+/// Samples using TPE for integer distributions.
+#[allow(clippy::cast_precision_loss, clippy::cast_possible_truncation)]
+pub(crate) fn sample_tpe_int(
+    dist: &IntDistribution,
+    good_values: Vec<i64>,
+    bad_values: Vec<i64>,
+    n_ei_candidates: usize,
+    kde_bandwidth: Option<f64>,
+    rng: &mut fastrand::Rng,
+) -> i64 {
+    let (low, high, log_scale, step) = (dist.low, dist.high, dist.log_scale, dist.step);
+
+    // Convert to floats for KDE
+    let good_floats: Vec<f64> = good_values.into_iter().map(|v| v as f64).collect();
+    let bad_floats: Vec<f64> = bad_values.into_iter().map(|v| v as f64).collect();
+
+    let float_dist = FloatDistribution {
+        low: low as f64,
+        high: high as f64,
+        log_scale,
+        step: step.map(|s| s as f64),
+    };
+
+    // Use float TPE sampling
+    let float_value = sample_tpe_float(
+        &float_dist,
+        good_floats,
+        bad_floats,
+        n_ei_candidates,
+        kde_bandwidth,
+        rng,
+    );
+
+    // Round to nearest integer
+    let int_value = float_value.round() as i64;
+
+    // Apply step constraint if present
+    let int_value = if let Some(step) = step {
+        let k = ((int_value - low) as f64 / step as f64).round() as i64;
+        low + k * step
+    } else {
+        int_value
+    };
+
+    // Ensure value is within bounds
+    int_value.clamp(low, high)
+}
+
+/// Samples using TPE for categorical distributions.
+#[allow(clippy::cast_precision_loss)]
+pub(crate) fn sample_tpe_categorical(
+    n_choices: usize,
+    good_indices: &[usize],
+    bad_indices: &[usize],
+    rng: &mut fastrand::Rng,
+) -> usize {
+    // Stack-allocate for the common case (<=32 choices), heap for rare large cases
+    let mut good_buf = [0usize; 32];
+    let mut bad_buf = [0usize; 32];
+    let mut weight_buf = [0.0f64; 32];
+
+    let mut good_vec;
+    let mut bad_vec;
+    let mut weight_vec;
+
+    let (good_counts, bad_counts, weights): (&mut [usize], &mut [usize], &mut [f64]) =
+        if n_choices <= 32 {
+            (
+                &mut good_buf[..n_choices],
+                &mut bad_buf[..n_choices],
+                &mut weight_buf[..n_choices],
+            )
+        } else {
+            good_vec = vec![0usize; n_choices];
+            bad_vec = vec![0usize; n_choices];
+            weight_vec = vec![0.0f64; n_choices];
+            (&mut good_vec, &mut bad_vec, &mut weight_vec)
+        };
+
+    // Count occurrences in good and bad groups
+    for &idx in good_indices {
+        if idx < n_choices {
+            good_counts[idx] += 1;
+        }
+    }
+    for &idx in bad_indices {
+        if idx < n_choices {
+            bad_counts[idx] += 1;
+        }
+    }
+
+    // Add smoothing (Laplace smoothing) to avoid zero probabilities
+    let good_total = good_indices.len() as f64 + n_choices as f64;
+    let bad_total = bad_indices.len() as f64 + n_choices as f64;
+
+    // Calculate l(x)/g(x) ratio for each category
+    for i in 0..n_choices {
+        let l_prob = (good_counts[i] as f64 + 1.0) / good_total;
+        let g_prob = (bad_counts[i] as f64 + 1.0) / bad_total;
+        weights[i] = l_prob / g_prob;
+    }
+
+    // Sample proportionally to weights
+    let total_weight: f64 = weights.iter().sum();
+    let threshold = rng.f64() * total_weight;
+
+    let mut cumulative = 0.0;
+    for (i, &w) in weights.iter().enumerate() {
+        cumulative += w;
+        if cumulative >= threshold {
+            return i;
+        }
+    }
+
+    // Fallback to last index (shouldn't happen)
+    n_choices - 1
+}

src/sampler/tpe/mod.rs

@@ -1,8 +1,65 @@
-//! Tree-Parzen Estimator (TPE) sampler implementation and utilities.
+//! Tree-Parzen Estimator (TPE) sampler family for Bayesian optimization.
 //!
-//! This module provides TPE-based sampling for Bayesian optimization,
-//! including support for intersection search space calculation.
+//! TPE is a sequential model-based optimization algorithm that models P(x|y) instead
+//! of P(y|x). It splits completed trials into "good" (below the gamma quantile) and
+//! "bad" groups, fits a kernel density estimator (KDE) to each, and proposes new
+//! points by maximizing the l(x)/g(x) ratio — an approximation of Expected Improvement.
+//!
+//! # Samplers
+//!
+//! | Sampler | Models parameters | Best for |
+//! |---------|-------------------|----------|
+//! | [`TpeSampler`] | Independently | General-purpose single-objective optimization |
+//! | [`MultivariateTpeSampler`] | Jointly | Problems with correlated parameters |
+//!
+//! # Gamma strategies
+//!
+//! The gamma quantile controls how many trials are considered "good". This module
+//! provides four built-in strategies via the [`GammaStrategy`] trait:
+//!
+//! | Strategy | Formula | Default |
+//! |----------|---------|---------|
+//! | [`FixedGamma`] | Constant value | gamma = 0.25 |
+//! | [`LinearGamma`] | Linear ramp from min to max | 0.10 → 0.25 over 100 trials |
+//! | [`SqrtGamma`] | 1/√n decay (Optuna-style) | factor = 1.0, max = 0.25 |
+//! | [`HyperoptGamma`] | (base+1)/n (Hyperopt-style) | base = 24, max = 0.25 |
+//!
+//! You can also implement [`GammaStrategy`] for a custom splitting rule.
+//!
+//! # Search-space utilities
+//!
+//! The [`search_space`] submodule provides [`IntersectionSearchSpace`] for computing
+//! the common parameter set across trials, and [`GroupDecomposedSearchSpace`] for
+//! splitting parameters into independent groups based on co-occurrence.
+//!
+//! # Examples
+//!
+//! Basic TPE with default settings:
+//!
+//! ```
+//! use optimizer::sampler::tpe::TpeSampler;
+//! use optimizer::{Direction, Study};
+//!
+//! let study: Study<f64> = Study::with_sampler(Direction::Minimize, TpeSampler::new());
+//! ```
+//!
+//! Multivariate TPE for correlated parameters:
+//!
+//! ```
+//! use optimizer::sampler::tpe::MultivariateTpeSampler;
+//! use optimizer::{Direction, Study};
+//!
+//! let sampler = MultivariateTpeSampler::builder()
+//!     .gamma(0.15)
+//!     .n_startup_trials(20)
+//!     .group(true)
+//!     .seed(42)
+//!     .build()
+//!     .unwrap();
+//! let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+//! ```
 
+pub(crate) mod common;
 mod gamma;
 mod multivariate;
 mod sampler;

src/sampler/tpe/multivariate/engine.rs

@@ -0,0 +1,623 @@
+//! Core multivariate TPE sampling logic.
+//!
+//! Contains the main sampling engine: group decomposition, single-group multivariate
+//! TPE, candidate selection, independent fallbacks, and value conversion.
+
+use std::collections::HashMap;
+
+use crate::distribution::Distribution;
+use crate::param::ParamValue;
+use crate::parameter::ParamId;
+use crate::sampler::CompletedTrial;
+
+use super::MultivariateTpeSampler;
+
+impl MultivariateTpeSampler {
+    /// Samples parameters by decomposing the search space into independent groups.
+    ///
+    /// When `group=true`, this method analyzes the trial history to identify groups of
+    /// parameters that always appear together, then samples each group independently
+    /// using multivariate TPE. This is more efficient when parameters naturally partition
+    /// into independent subsets (e.g., due to conditional search spaces).
+    ///
+    /// # Arguments
+    ///
+    /// * `search_space` - The full search space containing all parameters to sample.
+    /// * `history` - Completed trials from the optimization history.
+    ///
+    /// # Returns
+    ///
+    /// A `HashMap` mapping parameter names to their sampled values.
+    pub(crate) fn sample_with_groups(
+        &self,
+        search_space: &HashMap<ParamId, Distribution>,
+        history: &[CompletedTrial],
+    ) -> HashMap<ParamId, ParamValue> {
+        use std::collections::HashSet;
+
+        use crate::sampler::tpe::GroupDecomposedSearchSpace;
+
+        // Decompose the search space into independent parameter groups
+        let groups = GroupDecomposedSearchSpace::calculate(history);
+
+        let mut result: HashMap<ParamId, ParamValue> = HashMap::new();
+
+        // Sample each group independently
+        for group in &groups {
+            // Build a sub-search space for this group
+            let group_search_space: HashMap<ParamId, Distribution> = search_space
+                .iter()
+                .filter(|(id, _)| group.contains(id))
+                .map(|(id, dist)| (*id, dist.clone()))
+                .collect();
+
+            if group_search_space.is_empty() {
+                continue;
+            }
+
+            // Filter history to trials that have at least one parameter in this group
+            let group_history: Vec<&CompletedTrial> = history
+                .iter()
+                .filter(|trial| {
+                    trial
+                        .distributions
+                        .keys()
+                        .any(|param_id| group.contains(param_id))
+                })
+                .collect();
+
+            // Build completed trials from references for the group
+            // We need to create a temporary slice for sample_group_internal
+            let group_history_owned: Vec<CompletedTrial> =
+                group_history.iter().map(|t| (*t).clone()).collect();
+
+            // Sample this group using multivariate TPE
+            let mut rng = self.rng.lock();
+            let group_result =
+                self.sample_single_group(&group_search_space, &group_history_owned, &mut rng);
+            drop(rng);
+
+            // Merge group results into the main result
+            for (id, value) in group_result {
+                result.insert(id, value);
+            }
+        }
+
+        // Handle parameters not in any group (sample independently)
+        let grouped_params: HashSet<ParamId> = groups.iter().flatten().copied().collect();
+        let ungrouped_params: HashMap<ParamId, Distribution> = search_space
+            .iter()
+            .filter(|(id, _)| !grouped_params.contains(id) && !result.contains_key(id))
+            .map(|(id, dist)| (*id, dist.clone()))
+            .collect();
+
+        if !ungrouped_params.is_empty() {
+            // Sample ungrouped parameters uniformly (no history for them)
+            let mut rng = self.rng.lock();
+            for (id, dist) in &ungrouped_params {
+                let value = crate::sampler::common::sample_random(&mut rng, dist);
+                result.insert(*id, value);
+            }
+        }
+
+        result
+    }
+
+    /// Validates observations and fits multivariate KDEs for good and bad groups.
+    ///
+    /// Returns `None` if observations are invalid or KDE construction fails.
+    fn try_fit_kdes(
+        good_obs: Vec<Vec<f64>>,
+        bad_obs: Vec<Vec<f64>>,
+        expected_dims: usize,
+    ) -> Option<(crate::kde::MultivariateKDE, crate::kde::MultivariateKDE)> {
+        use crate::kde::MultivariateKDE;
+
+        let valid = !good_obs.is_empty()
+            && !bad_obs.is_empty()
+            && good_obs.iter().all(|obs| obs.len() == expected_dims)
+            && bad_obs.iter().all(|obs| obs.len() == expected_dims);
+
+        if !valid {
+            return None;
+        }
+
+        let good_kde = MultivariateKDE::new(good_obs).ok()?;
+        let bad_kde = MultivariateKDE::new(bad_obs).ok()?;
+        Some((good_kde, bad_kde))
+    }
+
+    /// Samples parameters as a single group using multivariate TPE.
+    ///
+    /// This is the core multivariate TPE sampling logic, used both in non-grouped mode
+    /// and for sampling individual groups in grouped mode.
+    ///
+    /// # Arguments
+    ///
+    /// * `search_space` - The search space for this group of parameters.
+    /// * `history` - Completed trials to use for model fitting.
+    /// * `rng` - Random number generator (caller must hold lock).
+    ///
+    /// # Returns
+    ///
+    /// A `HashMap` mapping parameter names to their sampled values.
+    pub(crate) fn sample_single_group(
+        &self,
+        search_space: &HashMap<ParamId, Distribution>,
+        history: &[CompletedTrial],
+        rng: &mut fastrand::Rng,
+    ) -> HashMap<ParamId, ParamValue> {
+        use crate::sampler::tpe::IntersectionSearchSpace;
+        use crate::sampler::tpe::common;
+
+        // Early returns for cases requiring random sampling
+        if history.len() < self.n_startup_trials {
+            return self.sample_all_uniform(search_space, rng);
+        }
+
+        let intersection = IntersectionSearchSpace::calculate(history);
+        if intersection.is_empty() {
+            return self.sample_all_independent_with_rng(search_space, history, rng);
+        }
+
+        let filtered = self.filter_trials(history, &intersection);
+        if filtered.len() < 2 {
+            return self.sample_all_independent_with_rng(search_space, history, rng);
+        }
+
+        let (good, bad) = self.split_trials(&filtered);
+
+        // Sample categorical parameters using TPE with l(x)/g(x) ratio
+        let mut result: HashMap<ParamId, ParamValue> = HashMap::new();
+        for (param_id, dist) in &intersection {
+            if let Distribution::Categorical(d) = dist {
+                let good_indices = Self::extract_categorical_indices(&good, *param_id);
+                let bad_indices = Self::extract_categorical_indices(&bad, *param_id);
+                let idx =
+                    common::sample_tpe_categorical(d.n_choices, &good_indices, &bad_indices, rng);
+                result.insert(*param_id, ParamValue::Categorical(idx));
+            }
+        }
+
+        // Collect continuous parameters
+        let mut param_order: Vec<ParamId> = intersection
+            .iter()
+            .filter(|(_, dist)| !matches!(dist, Distribution::Categorical(_)))
+            .map(|(id, _)| *id)
+            .collect();
+
+        if param_order.is_empty() {
+            self.fill_remaining_independent_with_rng(
+                search_space,
+                &intersection,
+                history,
+                &mut result,
+                rng,
+            );
+            return result;
+        }
+
+        param_order.sort();
+
+        // Extract observations, validate, and fit KDEs
+        let good_obs = self.extract_observations(&good, &param_order);
+        let bad_obs = self.extract_observations(&bad, &param_order);
+
+        let Some((good_kde, bad_kde)) = Self::try_fit_kdes(good_obs, bad_obs, param_order.len())
+        else {
+            self.fill_remaining_independent_with_rng(
+                search_space,
+                &intersection,
+                history,
+                &mut result,
+                rng,
+            );
+            return result;
+        };
+
+        // Compute parameter bounds for each dimension so candidates are clamped
+        let bounds: Vec<(f64, f64)> = param_order
+            .iter()
+            .filter_map(|id| {
+                intersection.get(id).and_then(|dist| match dist {
+                    Distribution::Float(d) => Some((d.low, d.high)),
+                    #[allow(clippy::cast_precision_loss)]
+                    Distribution::Int(d) => Some((d.low as f64, d.high as f64)),
+                    Distribution::Categorical(_) => None,
+                })
+            })
+            .collect();
+
+        let selected = self.select_candidate_with_rng(&good_kde, &bad_kde, &bounds, rng);
+
+        // Map selected values to parameter ids
+        for (idx, param_id) in param_order.iter().enumerate() {
+            if let Some(dist) = intersection.get(param_id) {
+                let value = selected[idx];
+                let param_value = self.convert_to_param_value(value, dist);
+                if let Some(pv) = param_value {
+                    result.insert(*param_id, pv);
+                }
+            }
+        }
+
+        // Fill remaining parameters using independent TPE sampling
+        self.fill_remaining_independent_with_rng(
+            search_space,
+            &intersection,
+            history,
+            &mut result,
+            rng,
+        );
+        result
+    }
+
+    /// Converts a raw f64 value to a `ParamValue` based on the distribution.
+    #[allow(clippy::unused_self)]
+    pub(crate) fn convert_to_param_value(
+        &self,
+        value: f64,
+        dist: &Distribution,
+    ) -> Option<ParamValue> {
+        match dist {
+            Distribution::Float(d) => {
+                let clamped = value.clamp(d.low, d.high);
+                let stepped = if let Some(step) = d.step {
+                    let steps = ((clamped - d.low) / step).round();
+                    (d.low + steps * step).clamp(d.low, d.high)
+                } else {
+                    clamped
+                };
+                Some(ParamValue::Float(stepped))
+            }
+            Distribution::Int(d) => {
+                #[allow(clippy::cast_possible_truncation)]
+                let int_value = value.round() as i64;
+                let clamped = int_value.clamp(d.low, d.high);
+                let stepped = if let Some(step) = d.step {
+                    let steps = (clamped - d.low) / step;
+                    (d.low + steps * step).clamp(d.low, d.high)
+                } else {
+                    clamped
+                };
+                Some(ParamValue::Int(stepped))
+            }
+            Distribution::Categorical(_) => None,
+        }
+    }
+
+    /// Clamps each dimension of a candidate to the corresponding parameter bounds.
+    fn clamp_candidate(candidate: &mut [f64], bounds: &[(f64, f64)]) {
+        for (val, &(lo, hi)) in candidate.iter_mut().zip(bounds.iter()) {
+            *val = val.clamp(lo, hi);
+        }
+    }
+
+    /// Selects the best candidate from a set of samples using the joint acquisition function.
+    ///
+    /// This method implements the core TPE selection criterion: it generates candidates
+    /// from the "good" KDE (l(x)) and selects the one that maximizes the ratio l(x)/g(x),
+    /// which is equivalent to maximizing `log(l(x)) - log(g(x))`.
+    ///
+    /// Candidates are clamped to parameter bounds before evaluation so the acquisition
+    /// function scores the values that will actually be used.
+    #[must_use]
+    #[cfg(test)]
+    pub(crate) fn select_candidate(
+        &self,
+        good_kde: &crate::kde::MultivariateKDE,
+        bad_kde: &crate::kde::MultivariateKDE,
+        bounds: &[(f64, f64)],
+    ) -> Vec<f64> {
+        let mut rng = self.rng.lock();
+
+        // Generate candidates from the good distribution, clamped to bounds
+        let candidates: Vec<Vec<f64>> = (0..self.n_ei_candidates)
+            .map(|_| {
+                let mut c = good_kde.sample(&mut rng);
+                Self::clamp_candidate(&mut c, bounds);
+                c
+            })
+            .collect();
+
+        // Compute log(l(x)) - log(g(x)) for each candidate
+        // This is equivalent to log(l(x)/g(x)) which we want to maximize
+        let log_ratios: Vec<f64> = candidates
+            .iter()
+            .map(|candidate| {
+                let log_l = good_kde.log_pdf(candidate);
+                let log_g = bad_kde.log_pdf(candidate);
+                log_l - log_g
+            })
+            .collect();
+
+        // Find the candidate with the maximum log ratio
+        let mut best_idx = 0;
+        let mut best_ratio = f64::NEG_INFINITY;
+
+        for (idx, &ratio) in log_ratios.iter().enumerate() {
+            // Handle NaN by treating it as worse than any finite value
+            if ratio > best_ratio || (best_ratio.is_nan() && !ratio.is_nan()) {
+                best_ratio = ratio;
+                best_idx = idx;
+            }
+        }
+
+        candidates.into_iter().nth(best_idx).unwrap_or_default()
+    }
+
+    /// Selects the best candidate using an external RNG.
+    ///
+    /// This variant accepts an external RNG, used when the caller already holds the lock.
+    /// Candidates are clamped to parameter bounds before evaluation.
+    pub(crate) fn select_candidate_with_rng(
+        &self,
+        good_kde: &crate::kde::MultivariateKDE,
+        bad_kde: &crate::kde::MultivariateKDE,
+        bounds: &[(f64, f64)],
+        rng: &mut fastrand::Rng,
+    ) -> Vec<f64> {
+        // Generate candidates from the good distribution, clamped to bounds
+        let candidates: Vec<Vec<f64>> = (0..self.n_ei_candidates)
+            .map(|_| {
+                let mut c = good_kde.sample(rng);
+                Self::clamp_candidate(&mut c, bounds);
+                c
+            })
+            .collect();
+
+        // Compute log(l(x)) - log(g(x)) for each candidate
+        let log_ratios: Vec<f64> = candidates
+            .iter()
+            .map(|candidate| {
+                let log_l = good_kde.log_pdf(candidate);
+                let log_g = bad_kde.log_pdf(candidate);
+                log_l - log_g
+            })
+            .collect();
+
+        // Find the candidate with the maximum log ratio
+        let mut best_idx = 0;
+        let mut best_ratio = f64::NEG_INFINITY;
+
+        for (idx, &ratio) in log_ratios.iter().enumerate() {
+            if ratio > best_ratio || (best_ratio.is_nan() && !ratio.is_nan()) {
+                best_ratio = ratio;
+                best_idx = idx;
+            }
+        }
+
+        candidates.into_iter().nth(best_idx).unwrap_or_default()
+    }
+
+    /// Fills remaining parameters using independent TPE sampling with an external RNG.
+    ///
+    /// This variant accepts an external RNG, used when the caller already holds the lock.
+    pub(crate) fn fill_remaining_independent_with_rng(
+        &self,
+        search_space: &HashMap<ParamId, Distribution>,
+        _intersection: &HashMap<ParamId, Distribution>,
+        history: &[CompletedTrial],
+        result: &mut HashMap<ParamId, ParamValue>,
+        rng: &mut fastrand::Rng,
+    ) {
+        // Identify parameters not in result, sorted for deterministic RNG consumption
+        let mut missing_params: Vec<(&ParamId, &Distribution)> = search_space
+            .iter()
+            .filter(|(id, _)| !result.contains_key(id))
+            .collect();
+
+        if missing_params.is_empty() {
+            return;
+        }
+
+        missing_params.sort_by_key(|(id, _)| *id);
+
+        // Split trials for independent sampling
+        let (good_trials, bad_trials) = self.split_trials(&history.iter().collect::<Vec<_>>());
+
+        for (param_id, dist) in missing_params {
+            let value =
+                self.sample_independent_tpe(*param_id, dist, &good_trials, &bad_trials, rng);
+            result.insert(*param_id, value);
+        }
+    }
+
+    /// Samples all parameters using independent TPE sampling.
+    ///
+    /// This is used as a complete fallback when no intersection search space exists.
+    /// Parameters are sorted by `ParamId` for deterministic RNG consumption order.
+    #[cfg(test)]
+    pub(crate) fn sample_all_independent(
+        &self,
+        search_space: &HashMap<ParamId, Distribution>,
+        history: &[CompletedTrial],
+    ) -> HashMap<ParamId, ParamValue> {
+        // Split trials for independent sampling
+        let (good_trials, bad_trials) = self.split_trials(&history.iter().collect::<Vec<_>>());
+
+        let mut rng = self.rng.lock();
+        let mut result = HashMap::new();
+
+        let mut sorted: Vec<_> = search_space.iter().collect();
+        sorted.sort_by_key(|(id, _)| *id);
+        for (param_id, dist) in sorted {
+            let value =
+                self.sample_independent_tpe(*param_id, dist, &good_trials, &bad_trials, &mut rng);
+            result.insert(*param_id, value);
+        }
+
+        result
+    }
+
+    /// Samples all parameters using independent TPE sampling with an external RNG.
+    ///
+    /// This variant accepts an external RNG, used when the caller already holds the lock.
+    /// Parameters are sorted by `ParamId` for deterministic RNG consumption order.
+    pub(crate) fn sample_all_independent_with_rng(
+        &self,
+        search_space: &HashMap<ParamId, Distribution>,
+        history: &[CompletedTrial],
+        rng: &mut fastrand::Rng,
+    ) -> HashMap<ParamId, ParamValue> {
+        // Split trials for independent sampling
+        let (good_trials, bad_trials) = self.split_trials(&history.iter().collect::<Vec<_>>());
+
+        let mut result = HashMap::new();
+
+        let mut sorted: Vec<_> = search_space.iter().collect();
+        sorted.sort_by_key(|(id, _)| *id);
+        for (param_id, dist) in sorted {
+            let value =
+                self.sample_independent_tpe(*param_id, dist, &good_trials, &bad_trials, rng);
+            result.insert(*param_id, value);
+        }
+
+        result
+    }
+
+    /// Samples a single parameter using independent TPE.
+    ///
+    /// This method extracts values for the given parameter from good and bad trials,
+    /// fits univariate KDEs, and samples using the TPE acquisition function.
+    pub(crate) fn sample_independent_tpe(
+        &self,
+        param_id: ParamId,
+        distribution: &Distribution,
+        good_trials: &[&CompletedTrial],
+        bad_trials: &[&CompletedTrial],
+        rng: &mut fastrand::Rng,
+    ) -> ParamValue {
+        match distribution {
+            Distribution::Float(d) => {
+                self.sample_independent_float(param_id, d, good_trials, bad_trials, rng)
+            }
+            Distribution::Int(d) => {
+                self.sample_independent_int(param_id, d, good_trials, bad_trials, rng)
+            }
+            Distribution::Categorical(d) => {
+                self.sample_independent_categorical(param_id, d, good_trials, bad_trials, rng)
+            }
+        }
+    }
+
+    fn sample_independent_float(
+        &self,
+        param_id: ParamId,
+        d: &crate::distribution::FloatDistribution,
+        good_trials: &[&CompletedTrial],
+        bad_trials: &[&CompletedTrial],
+        rng: &mut fastrand::Rng,
+    ) -> ParamValue {
+        use crate::sampler::tpe::common;
+
+        let good_values: Vec<f64> = good_trials
+            .iter()
+            .filter_map(|t| t.params.get(&param_id))
+            .filter_map(|v| match v {
+                ParamValue::Float(f) => Some(*f),
+                _ => None,
+            })
+            .filter(|&v| v >= d.low && v <= d.high)
+            .collect();
+
+        let bad_values: Vec<f64> = bad_trials
+            .iter()
+            .filter_map(|t| t.params.get(&param_id))
+            .filter_map(|v| match v {
+                ParamValue::Float(f) => Some(*f),
+                _ => None,
+            })
+            .filter(|&v| v >= d.low && v <= d.high)
+            .collect();
+
+        if good_values.is_empty() || bad_values.is_empty() {
+            return crate::sampler::common::sample_random(rng, &Distribution::Float(d.clone()));
+        }
+
+        let value =
+            common::sample_tpe_float(d, good_values, bad_values, self.n_ei_candidates, None, rng);
+        ParamValue::Float(value)
+    }
+
+    fn sample_independent_int(
+        &self,
+        param_id: ParamId,
+        d: &crate::distribution::IntDistribution,
+        good_trials: &[&CompletedTrial],
+        bad_trials: &[&CompletedTrial],
+        rng: &mut fastrand::Rng,
+    ) -> ParamValue {
+        use crate::sampler::tpe::common;
+
+        let good_values: Vec<i64> = good_trials
+            .iter()
+            .filter_map(|t| t.params.get(&param_id))
+            .filter_map(|v| match v {
+                ParamValue::Int(i) => Some(*i),
+                _ => None,
+            })
+            .filter(|&v| v >= d.low && v <= d.high)
+            .collect();
+
+        let bad_values: Vec<i64> = bad_trials
+            .iter()
+            .filter_map(|t| t.params.get(&param_id))
+            .filter_map(|v| match v {
+                ParamValue::Int(i) => Some(*i),
+                _ => None,
+            })
+            .filter(|&v| v >= d.low && v <= d.high)
+            .collect();
+
+        if good_values.is_empty() || bad_values.is_empty() {
+            return crate::sampler::common::sample_random(rng, &Distribution::Int(d.clone()));
+        }
+
+        let value =
+            common::sample_tpe_int(d, good_values, bad_values, self.n_ei_candidates, None, rng);
+        ParamValue::Int(value)
+    }
+
+    #[allow(clippy::unused_self)]
+    fn sample_independent_categorical(
+        &self,
+        param_id: ParamId,
+        d: &crate::distribution::CategoricalDistribution,
+        good_trials: &[&CompletedTrial],
+        bad_trials: &[&CompletedTrial],
+        rng: &mut fastrand::Rng,
+    ) -> ParamValue {
+        use crate::sampler::tpe::common;
+
+        let good_indices: Vec<usize> = good_trials
+            .iter()
+            .filter_map(|t| t.params.get(&param_id))
+            .filter_map(|v| match v {
+                ParamValue::Categorical(i) => Some(*i),
+                _ => None,
+            })
+            .filter(|&i| i < d.n_choices)
+            .collect();
+
+        let bad_indices: Vec<usize> = bad_trials
+            .iter()
+            .filter_map(|t| t.params.get(&param_id))
+            .filter_map(|v| match v {
+                ParamValue::Categorical(i) => Some(*i),
+                _ => None,
+            })
+            .filter(|&i| i < d.n_choices)
+            .collect();
+
+        if good_indices.is_empty() || bad_indices.is_empty() {
+            return crate::sampler::common::sample_random(
+                rng,
+                &Distribution::Categorical(d.clone()),
+            );
+        }
+
+        let idx = common::sample_tpe_categorical(d.n_choices, &good_indices, &bad_indices, rng);
+        ParamValue::Categorical(idx)
+    }
+}

src/sampler/tpe/multivariate/trials.rs

@@ -0,0 +1,220 @@
+//! Trial processing for the multivariate TPE sampler.
+//!
+//! Contains constant-liar imputation, trial filtering, good/bad splitting,
+//! observation extraction, and categorical index extraction.
+
+use std::collections::HashMap;
+
+use crate::distribution::Distribution;
+use crate::param::ParamValue;
+use crate::parameter::ParamId;
+use crate::sampler::{CompletedTrial, PendingTrial};
+
+use super::{ConstantLiarStrategy, MultivariateTpeSampler};
+
+impl MultivariateTpeSampler {
+    /// Imputes objective values for pending trials based on the constant liar strategy.
+    ///
+    /// In parallel optimization, multiple trials may be running simultaneously. This method
+    /// assigns "lie" values to pending trials so they can be included in the model fitting,
+    /// which helps avoid redundant exploration of the same region.
+    ///
+    /// # Arguments
+    ///
+    /// * `pending_trials` - Trials that are currently running and have no objective value yet.
+    /// * `completed_trials` - Trials that have completed and have objective values.
+    ///
+    /// # Returns
+    ///
+    /// A vector of `CompletedTrial` objects containing both the original completed trials
+    /// and the pending trials with imputed values. If the strategy is `None`, returns
+    /// only the completed trials (pending trials are ignored).
+    #[must_use]
+    pub fn impute_pending_trials(
+        &self,
+        pending_trials: &[PendingTrial],
+        completed_trials: &[CompletedTrial],
+    ) -> Vec<CompletedTrial> {
+        // Start with a copy of completed trials
+        let mut result: Vec<CompletedTrial> = completed_trials.to_vec();
+
+        // If strategy is None or no pending trials, just return completed trials
+        if matches!(self.constant_liar, ConstantLiarStrategy::None) || pending_trials.is_empty() {
+            return result;
+        }
+
+        // Compute the imputation value based on strategy
+        let imputed_value = self.compute_imputation_value(completed_trials);
+
+        // Convert pending trials to completed trials with imputed values
+        for pending in pending_trials {
+            result.push(CompletedTrial::new(
+                pending.id,
+                pending.params.clone(),
+                pending.distributions.clone(),
+                HashMap::new(),
+                imputed_value,
+            ));
+        }
+
+        result
+    }
+
+    /// Computes the imputation value based on the constant liar strategy.
+    ///
+    /// This is a helper method used by [`impute_pending_trials`](Self::impute_pending_trials).
+    #[allow(clippy::cast_precision_loss)]
+    pub(crate) fn compute_imputation_value(&self, completed_trials: &[CompletedTrial]) -> f64 {
+        match self.constant_liar {
+            ConstantLiarStrategy::None => 0.0, // This case is handled before calling this method
+            ConstantLiarStrategy::Mean => {
+                if completed_trials.is_empty() {
+                    0.0
+                } else {
+                    let sum: f64 = completed_trials.iter().map(|t| t.value).sum();
+                    sum / completed_trials.len() as f64
+                }
+            }
+            ConstantLiarStrategy::Best => {
+                // Best means minimum for minimization problems
+                completed_trials
+                    .iter()
+                    .map(|t| t.value)
+                    .fold(f64::INFINITY, f64::min)
+            }
+            ConstantLiarStrategy::Worst => {
+                // Worst means maximum for minimization problems
+                completed_trials
+                    .iter()
+                    .map(|t| t.value)
+                    .fold(f64::NEG_INFINITY, f64::max)
+            }
+            ConstantLiarStrategy::Custom(v) => v,
+        }
+    }
+
+    /// Filters trials to those containing all parameters in the search space.
+    ///
+    /// Only trials that contain ALL parameters in the search space are included,
+    /// ensuring we can model the joint distribution over all parameters.
+    #[must_use]
+    pub fn filter_trials<'a>(
+        &self,
+        history: &'a [CompletedTrial],
+        search_space: &HashMap<ParamId, Distribution>,
+    ) -> Vec<&'a CompletedTrial> {
+        history
+            .iter()
+            .filter(|trial| {
+                // Include trial only if it has ALL parameters in the search space
+                search_space
+                    .keys()
+                    .all(|param_id| trial.params.contains_key(param_id))
+            })
+            .collect()
+    }
+
+    /// Splits filtered trials into good and bad groups based on the gamma quantile.
+    ///
+    /// The gamma value is computed dynamically using the configured [`super::GammaStrategy`].
+    /// Trials are sorted by objective value (ascending for minimization), and the
+    /// gamma quantile determines the split point.
+    #[allow(
+        clippy::cast_precision_loss,
+        clippy::cast_possible_truncation,
+        clippy::cast_sign_loss
+    )]
+    #[must_use]
+    pub fn split_trials<'a>(
+        &self,
+        trials: &[&'a CompletedTrial],
+    ) -> (Vec<&'a CompletedTrial>, Vec<&'a CompletedTrial>) {
+        if trials.is_empty() {
+            return (vec![], vec![]);
+        }
+
+        // Sort trials by objective value (ascending for minimization)
+        let mut sorted_indices: Vec<usize> = (0..trials.len()).collect();
+        sorted_indices.sort_by(|&a, &b| {
+            trials[a]
+                .value
+                .partial_cmp(&trials[b].value)
+                .unwrap_or(core::cmp::Ordering::Equal)
+        });
+
+        // Compute gamma using the strategy and clamp to valid range
+        let gamma = self
+            .gamma_strategy
+            .gamma(trials.len())
+            .clamp(f64::EPSILON, 1.0 - f64::EPSILON);
+
+        // Calculate the split point (gamma quantile)
+        // Ensure at least 1 trial in each group if possible
+        let n_good = ((trials.len() as f64 * gamma).ceil() as usize)
+            .max(1)
+            .min(trials.len().saturating_sub(1));
+
+        // Handle edge case: if we have only 1 trial, put it in good
+        if trials.len() == 1 {
+            return (vec![trials[0]], vec![]);
+        }
+
+        let good: Vec<_> = sorted_indices[..n_good]
+            .iter()
+            .map(|&i| trials[i])
+            .collect();
+        let bad: Vec<_> = sorted_indices[n_good..]
+            .iter()
+            .map(|&i| trials[i])
+            .collect();
+
+        (good, bad)
+    }
+
+    /// Extracts parameter values from trials as a numeric observation matrix.
+    ///
+    /// Each row in the output represents one trial's parameter values in the specified order.
+    /// Categorical parameters are skipped.
+    #[must_use]
+    #[allow(clippy::cast_precision_loss)]
+    pub fn extract_observations(
+        &self,
+        trials: &[&CompletedTrial],
+        param_order: &[ParamId],
+    ) -> Vec<Vec<f64>> {
+        trials
+            .iter()
+            .map(|trial| {
+                param_order
+                    .iter()
+                    .filter_map(|param_id| {
+                        trial.params.get(param_id).and_then(|value| match value {
+                            crate::param::ParamValue::Float(f) => Some(*f),
+                            crate::param::ParamValue::Int(i) => Some(*i as f64),
+                            crate::param::ParamValue::Categorical(_) => None, // Skip categorical
+                        })
+                    })
+                    .collect()
+            })
+            .collect()
+    }
+
+    /// Extracts categorical indices from trials for a specific parameter.
+    pub(crate) fn extract_categorical_indices(
+        trials: &[&CompletedTrial],
+        param_id: ParamId,
+    ) -> Vec<usize> {
+        trials
+            .iter()
+            .filter_map(|trial| {
+                trial.params.get(&param_id).and_then(|value| {
+                    if let ParamValue::Categorical(idx) = value {
+                        Some(*idx)
+                    } else {
+                        None
+                    }
+                })
+            })
+            .collect()
+    }
+}

src/sampler/tpe/sampler.rs

@@ -56,19 +56,19 @@
 //! ```
 
 use core::fmt::Debug;
+use core::sync::atomic::{AtomicU64, Ordering};
 use std::sync::Arc;
 
-use parking_lot::Mutex;
-use rand::rngs::StdRng;
-use rand::{RngExt, SeedableRng};
-
 use crate::distribution::Distribution;
 use crate::error::{Error, Result};
-use crate::kde::KernelDensityEstimator;
 use crate::param::ParamValue;
+use crate::rng_util;
+use crate::sampler::common;
 use crate::sampler::tpe::gamma::{FixedGamma, GammaStrategy};
 use crate::sampler::{CompletedTrial, Sampler};
 
+use super::common as tpe_common;
+
 // ============================================================================
 // Gamma Strategy Trait and Implementations
 // ============================================================================
@@ -103,7 +103,7 @@ use crate::sampler::{CompletedTrial, Sampler};
 ///
 /// // Create with custom settings using the builder
 /// let sampler = TpeSampler::builder()
-///     .gamma(0.15)  // Shorthand for Fixednew(0.15)
+///     .gamma(0.15)  // Shorthand for FixedGamma::new(0.15)
 ///     .n_startup_trials(20)
 ///     .n_ei_candidates(32)
 ///     .seed(42)
@@ -130,8 +130,10 @@ pub struct TpeSampler {
     n_ei_candidates: usize,
     /// Optional fixed bandwidth for KDE. If None, uses Scott's rule.
     kde_bandwidth: Option<f64>,
-    /// Thread-safe RNG for sampling.
-    rng: Mutex<StdRng>,
+    /// Base seed for deterministic per-call RNG derivation (no mutex needed).
+    seed: u64,
+    /// Monotonic counter to disambiguate calls with identical (`trial_id`, distribution).
+    call_seq: AtomicU64,
 }
 
 impl TpeSampler {
@@ -149,7 +151,8 @@ impl TpeSampler {
             n_startup_trials: 10,
             n_ei_candidates: 24,
             kde_bandwidth: None,
-            rng: Mutex::new(rand::make_rng()),
+            seed: fastrand::u64(..),
+            call_seq: AtomicU64::new(0),
         }
     }
 
@@ -249,17 +252,13 @@ impl TpeSampler {
             return Err(Error::InvalidBandwidth(bw));
         }
 
-        let rng = match seed {
-            Some(s) => StdRng::seed_from_u64(s),
-            None => rand::make_rng(),
-        };
-
         Ok(Self {
             gamma_strategy: Arc::new(gamma_strategy),
             n_startup_trials,
             n_ei_candidates,
             kde_bandwidth,
-            rng: Mutex::new(rng),
+            seed: seed.unwrap_or_else(|| fastrand::u64(..)),
+            call_seq: AtomicU64::new(0),
         })
     }
 
@@ -289,15 +288,6 @@ impl TpeSampler {
             return (vec![], vec![]);
         }
 
-        // Sort trials by value (ascending for minimization)
-        let mut sorted_indices: Vec<usize> = (0..history.len()).collect();
-        sorted_indices.sort_by(|&a, &b| {
-            history[a]
-                .value
-                .partial_cmp(&history[b].value)
-                .unwrap_or(core::cmp::Ordering::Equal)
-        });
-
         // Compute gamma using the strategy and clamp to valid range
         let gamma = self
             .gamma_strategy
@@ -310,243 +300,23 @@ impl TpeSampler {
             .max(1)
             .min(history.len() - 1);
 
-        let good: Vec<_> = sorted_indices[..n_good]
-            .iter()
-            .map(|&i| &history[i])
-            .collect();
-        let bad: Vec<_> = sorted_indices[n_good..]
-            .iter()
-            .map(|&i| &history[i])
-            .collect();
+        // Use quickselect (O(n)) to partition indices instead of full sort (O(n log n)).
+        // We only need to know which trials are in the top gamma-quantile, not their order.
+        let mut indices: Vec<usize> = (0..history.len()).collect();
+        if n_good > 0 {
+            indices.select_nth_unstable_by(n_good - 1, |&a, &b| {
+                history[a]
+                    .value
+                    .partial_cmp(&history[b].value)
+                    .unwrap_or(core::cmp::Ordering::Equal)
+            });
+        }
+
+        let good: Vec<_> = indices[..n_good].iter().map(|&i| &history[i]).collect();
+        let bad: Vec<_> = indices[n_good..].iter().map(|&i| &history[i]).collect();
 
         (good, bad)
     }
-
-    /// Samples uniformly from a distribution (used during startup phase).
-    #[allow(
-        clippy::cast_possible_truncation,
-        clippy::cast_precision_loss,
-        clippy::unused_self
-    )]
-    fn sample_uniform(&self, distribution: &Distribution, rng: &mut StdRng) -> ParamValue {
-        match distribution {
-            Distribution::Float(d) => {
-                let value = if d.log_scale {
-                    let log_low = d.low.ln();
-                    let log_high = d.high.ln();
-                    rng.random_range(log_low..=log_high).exp()
-                } else if let Some(step) = d.step {
-                    let n_steps = ((d.high - d.low) / step).floor() as i64;
-                    let k = rng.random_range(0..=n_steps);
-                    d.low + (k as f64) * step
-                } else {
-                    rng.random_range(d.low..=d.high)
-                };
-                ParamValue::Float(value)
-            }
-            Distribution::Int(d) => {
-                let value = if d.log_scale {
-                    let log_low = (d.low as f64).ln();
-                    let log_high = (d.high as f64).ln();
-                    let raw = rng.random_range(log_low..=log_high).exp().round() as i64;
-                    raw.clamp(d.low, d.high)
-                } else if let Some(step) = d.step {
-                    let n_steps = (d.high - d.low) / step;
-                    let k = rng.random_range(0..=n_steps);
-                    d.low + k * step
-                } else {
-                    rng.random_range(d.low..=d.high)
-                };
-                ParamValue::Int(value)
-            }
-            Distribution::Categorical(d) => {
-                ParamValue::Categorical(rng.random_range(0..d.n_choices))
-            }
-        }
-    }
-
-    /// Samples using TPE for float distributions.
-    #[allow(clippy::too_many_arguments)]
-    fn sample_tpe_float(
-        &self,
-        low: f64,
-        high: f64,
-        log_scale: bool,
-        step: Option<f64>,
-        good_values: Vec<f64>,
-        bad_values: Vec<f64>,
-        rng: &mut StdRng,
-    ) -> f64 {
-        // Transform to internal space (log space if needed)
-        let (internal_low, internal_high, good_internal, bad_internal) = if log_scale {
-            let i_low = low.ln();
-            let i_high = high.ln();
-            let g: Vec<f64> = good_values.iter().map(|&v| v.ln()).collect();
-            let b: Vec<f64> = bad_values.iter().map(|&v| v.ln()).collect();
-            (i_low, i_high, g, b)
-        } else {
-            (low, high, good_values, bad_values)
-        };
-
-        // Fit KDEs to good and bad groups
-        let l_kde = match self.kde_bandwidth {
-            Some(bw) => KernelDensityEstimator::with_bandwidth(good_internal, bw),
-            None => KernelDensityEstimator::new(good_internal),
-        };
-        let g_kde = match self.kde_bandwidth {
-            Some(bw) => KernelDensityEstimator::with_bandwidth(bad_internal, bw),
-            None => KernelDensityEstimator::new(bad_internal),
-        };
-
-        // If KDE construction fails, fall back to uniform sampling
-        let (Ok(l_kde), Ok(g_kde)) = (l_kde, g_kde) else {
-            return rng.random_range(low..=high);
-        };
-
-        // Generate candidates from l(x) and select the one with best l(x)/g(x) ratio
-        let mut best_candidate = internal_low;
-        let mut best_ratio = f64::NEG_INFINITY;
-
-        for _ in 0..self.n_ei_candidates {
-            let candidate = l_kde.sample(rng);
-
-            // Clamp to bounds
-            let candidate = candidate.clamp(internal_low, internal_high);
-
-            let l_density = l_kde.pdf(candidate);
-            let g_density = g_kde.pdf(candidate);
-
-            // Compute l(x)/g(x) ratio, handling zero density
-            let ratio = if g_density < f64::EPSILON {
-                if l_density > f64::EPSILON {
-                    f64::INFINITY
-                } else {
-                    0.0
-                }
-            } else {
-                l_density / g_density
-            };
-
-            if ratio > best_ratio {
-                best_ratio = ratio;
-                best_candidate = candidate;
-            }
-        }
-
-        // Transform back from internal space
-        let mut value = if log_scale {
-            best_candidate.exp()
-        } else {
-            best_candidate
-        };
-
-        // Apply step constraint if present
-        if let Some(step) = step {
-            let k = ((value - low) / step).round();
-            value = low + k * step;
-        }
-
-        // Ensure value is within bounds
-        value.clamp(low, high)
-    }
-
-    /// Samples using TPE for integer distributions.
-    #[allow(
-        clippy::too_many_arguments,
-        clippy::cast_precision_loss,
-        clippy::cast_possible_truncation
-    )]
-    fn sample_tpe_int(
-        &self,
-        low: i64,
-        high: i64,
-        log_scale: bool,
-        step: Option<i64>,
-        good_values: &[i64],
-        bad_values: &[i64],
-        rng: &mut StdRng,
-    ) -> i64 {
-        // Convert to floats for KDE
-        let good_floats: Vec<f64> = good_values.iter().map(|&v| v as f64).collect();
-        let bad_floats: Vec<f64> = bad_values.iter().map(|&v| v as f64).collect();
-
-        // Use float TPE sampling
-        let float_value = self.sample_tpe_float(
-            low as f64,
-            high as f64,
-            log_scale,
-            step.map(|s| s as f64),
-            good_floats,
-            bad_floats,
-            rng,
-        );
-
-        // Round to nearest integer
-        let int_value = float_value.round() as i64;
-
-        // Apply step constraint if present
-        let int_value = if let Some(step) = step {
-            let k = ((int_value - low) as f64 / step as f64).round() as i64;
-            low + k * step
-        } else {
-            int_value
-        };
-
-        // Ensure value is within bounds
-        int_value.clamp(low, high)
-    }
-
-    /// Samples using TPE for categorical distributions.
-    #[allow(clippy::cast_precision_loss, clippy::unused_self)]
-    fn sample_tpe_categorical(
-        &self,
-        n_choices: usize,
-        good_indices: &[usize],
-        bad_indices: &[usize],
-        rng: &mut StdRng,
-    ) -> usize {
-        // Count occurrences in good and bad groups
-        let mut good_counts = vec![0usize; n_choices];
-        let mut bad_counts = vec![0usize; n_choices];
-
-        for &idx in good_indices {
-            if idx < n_choices {
-                good_counts[idx] += 1;
-            }
-        }
-        for &idx in bad_indices {
-            if idx < n_choices {
-                bad_counts[idx] += 1;
-            }
-        }
-
-        // Add smoothing (Laplace smoothing) to avoid zero probabilities
-        let good_total = good_indices.len() as f64 + n_choices as f64;
-        let bad_total = bad_indices.len() as f64 + n_choices as f64;
-
-        // Calculate l(x)/g(x) ratio for each category
-        let mut weights = vec![0.0f64; n_choices];
-        for i in 0..n_choices {
-            let l_prob = (good_counts[i] as f64 + 1.0) / good_total;
-            let g_prob = (bad_counts[i] as f64 + 1.0) / bad_total;
-            weights[i] = l_prob / g_prob;
-        }
-
-        // Sample proportionally to weights
-        let total_weight: f64 = weights.iter().sum();
-        let threshold = rng.random::<f64>() * total_weight;
-
-        let mut cumulative = 0.0;
-        for (i, &w) in weights.iter().enumerate() {
-            cumulative += w;
-            if cumulative >= threshold {
-                return i;
-            }
-        }
-
-        // Fallback to last index (shouldn't happen)
-        n_choices - 1
-    }
 }
 
 impl Default for TpeSampler {
@@ -855,17 +625,13 @@ impl TpeSamplerBuilder {
             return Err(Error::InvalidBandwidth(bw));
         }
 
-        let rng = match self.seed {
-            Some(s) => StdRng::seed_from_u64(s),
-            None => rand::make_rng(),
-        };
-
         Ok(TpeSampler {
             gamma_strategy,
             n_startup_trials: self.n_startup_trials,
             n_ei_candidates: self.n_ei_candidates,
             kde_bandwidth: self.kde_bandwidth,
-            rng: Mutex::new(rng),
+            seed: self.seed.unwrap_or_else(|| fastrand::u64(..)),
+            call_seq: AtomicU64::new(0),
         })
     }
 }
@@ -876,19 +642,185 @@ impl Default for TpeSamplerBuilder {
     }
 }
 
+impl TpeSampler {
+    fn sample_float(
+        &self,
+        d: &crate::distribution::FloatDistribution,
+        good_trials: &[&CompletedTrial],
+        bad_trials: &[&CompletedTrial],
+        rng: &mut fastrand::Rng,
+    ) -> ParamValue {
+        let target_dist = Distribution::Float(d.clone());
+        let good_values: Vec<f64> = good_trials
+            .iter()
+            .filter_map(|t| {
+                t.distributions.iter().find_map(|(id, dist)| {
+                    if *dist == target_dist {
+                        t.params.get(id).and_then(|v| match v {
+                            ParamValue::Float(f) => Some(*f),
+                            _ => None,
+                        })
+                    } else {
+                        None
+                    }
+                })
+            })
+            .collect();
+
+        let bad_values: Vec<f64> = bad_trials
+            .iter()
+            .filter_map(|t| {
+                t.distributions.iter().find_map(|(id, dist)| {
+                    if *dist == target_dist {
+                        t.params.get(id).and_then(|v| match v {
+                            ParamValue::Float(f) => Some(*f),
+                            _ => None,
+                        })
+                    } else {
+                        None
+                    }
+                })
+            })
+            .collect();
+
+        if good_values.is_empty() || bad_values.is_empty() {
+            return ParamValue::Float(rng_util::f64_range(rng, d.low, d.high));
+        }
+
+        let value = tpe_common::sample_tpe_float(
+            d,
+            good_values,
+            bad_values,
+            self.n_ei_candidates,
+            self.kde_bandwidth,
+            rng,
+        );
+        ParamValue::Float(value)
+    }
+
+    fn sample_int(
+        &self,
+        d: &crate::distribution::IntDistribution,
+        good_trials: &[&CompletedTrial],
+        bad_trials: &[&CompletedTrial],
+        rng: &mut fastrand::Rng,
+    ) -> ParamValue {
+        let target_dist = Distribution::Int(d.clone());
+        let good_values: Vec<i64> = good_trials
+            .iter()
+            .filter_map(|t| {
+                t.distributions.iter().find_map(|(id, dist)| {
+                    if *dist == target_dist {
+                        t.params.get(id).and_then(|v| match v {
+                            ParamValue::Int(i) => Some(*i),
+                            _ => None,
+                        })
+                    } else {
+                        None
+                    }
+                })
+            })
+            .collect();
+
+        let bad_values: Vec<i64> = bad_trials
+            .iter()
+            .filter_map(|t| {
+                t.distributions.iter().find_map(|(id, dist)| {
+                    if *dist == target_dist {
+                        t.params.get(id).and_then(|v| match v {
+                            ParamValue::Int(i) => Some(*i),
+                            _ => None,
+                        })
+                    } else {
+                        None
+                    }
+                })
+            })
+            .collect();
+
+        if good_values.is_empty() || bad_values.is_empty() {
+            return common::sample_random(rng, &Distribution::Int(d.clone()));
+        }
+
+        let value = tpe_common::sample_tpe_int(
+            d,
+            good_values,
+            bad_values,
+            self.n_ei_candidates,
+            self.kde_bandwidth,
+            rng,
+        );
+        ParamValue::Int(value)
+    }
+
+    #[allow(clippy::unused_self)]
+    fn sample_categorical(
+        &self,
+        d: &crate::distribution::CategoricalDistribution,
+        good_trials: &[&CompletedTrial],
+        bad_trials: &[&CompletedTrial],
+        rng: &mut fastrand::Rng,
+    ) -> ParamValue {
+        let target_dist = Distribution::Categorical(d.clone());
+        let good_indices: Vec<usize> = good_trials
+            .iter()
+            .filter_map(|t| {
+                t.distributions.iter().find_map(|(id, dist)| {
+                    if *dist == target_dist {
+                        t.params.get(id).and_then(|v| match v {
+                            ParamValue::Categorical(i) => Some(*i),
+                            _ => None,
+                        })
+                    } else {
+                        None
+                    }
+                })
+            })
+            .collect();
+
+        let bad_indices: Vec<usize> = bad_trials
+            .iter()
+            .filter_map(|t| {
+                t.distributions.iter().find_map(|(id, dist)| {
+                    if *dist == target_dist {
+                        t.params.get(id).and_then(|v| match v {
+                            ParamValue::Categorical(i) => Some(*i),
+                            _ => None,
+                        })
+                    } else {
+                        None
+                    }
+                })
+            })
+            .collect();
+
+        if good_indices.is_empty() || bad_indices.is_empty() {
+            return common::sample_random(rng, &Distribution::Categorical(d.clone()));
+        }
+
+        let index =
+            tpe_common::sample_tpe_categorical(d.n_choices, &good_indices, &bad_indices, rng);
+        ParamValue::Categorical(index)
+    }
+}
+
 impl Sampler for TpeSampler {
-    #[allow(clippy::too_many_lines)]
     fn sample(
         &self,
         distribution: &Distribution,
-        _trial_id: u64,
+        trial_id: u64,
         history: &[CompletedTrial],
     ) -> ParamValue {
-        let mut rng = self.rng.lock();
+        let seq = self.call_seq.fetch_add(1, Ordering::Relaxed);
+        let mut rng = fastrand::Rng::with_seed(rng_util::mix_seed(
+            self.seed,
+            trial_id,
+            rng_util::distribution_fingerprint(distribution).wrapping_add(seq),
+        ));
 
         // Fall back to random sampling during startup phase
         if history.len() < self.n_startup_trials {
-            return self.sample_uniform(distribution, &mut rng);
+            return common::sample_random(&mut rng, distribution);
         }
 
         // Split trials into good and bad groups
@@ -896,118 +828,14 @@ impl Sampler for TpeSampler {
 
         // Need at least 1 trial in each group for TPE
         if good_trials.is_empty() || bad_trials.is_empty() {
-            return self.sample_uniform(distribution, &mut rng);
+            return common::sample_random(&mut rng, distribution);
         }
 
-        // Extract parameter values for this distribution
-        // Since we don't have the parameter name here, we need to look at all
-        // trials and find matching distributions
-        // Note: This is a simplification - in practice, we'd need the param name
-        // For now, we'll collect values from trials that have this exact distribution type
-
         match distribution {
-            Distribution::Float(d) => {
-                // Collect float values from trials
-                let good_values: Vec<f64> = good_trials
-                    .iter()
-                    .flat_map(|t| t.params.values())
-                    .filter_map(|v| match v {
-                        ParamValue::Float(f) => Some(*f),
-                        _ => None,
-                    })
-                    .filter(|&v| v >= d.low && v <= d.high)
-                    .collect();
-
-                let bad_values: Vec<f64> = bad_trials
-                    .iter()
-                    .flat_map(|t| t.params.values())
-                    .filter_map(|v| match v {
-                        ParamValue::Float(f) => Some(*f),
-                        _ => None,
-                    })
-                    .filter(|&v| v >= d.low && v <= d.high)
-                    .collect();
-
-                // Need values in both groups for TPE
-                if good_values.is_empty() || bad_values.is_empty() {
-                    return self.sample_uniform(distribution, &mut rng);
-                }
-
-                let value = self.sample_tpe_float(
-                    d.low,
-                    d.high,
-                    d.log_scale,
-                    d.step,
-                    good_values,
-                    bad_values,
-                    &mut rng,
-                );
-                ParamValue::Float(value)
-            }
-            Distribution::Int(d) => {
-                let good_values: Vec<i64> = good_trials
-                    .iter()
-                    .flat_map(|t| t.params.values())
-                    .filter_map(|v| match v {
-                        ParamValue::Int(i) => Some(*i),
-                        _ => None,
-                    })
-                    .filter(|&v| v >= d.low && v <= d.high)
-                    .collect();
-
-                let bad_values: Vec<i64> = bad_trials
-                    .iter()
-                    .flat_map(|t| t.params.values())
-                    .filter_map(|v| match v {
-                        ParamValue::Int(i) => Some(*i),
-                        _ => None,
-                    })
-                    .filter(|&v| v >= d.low && v <= d.high)
-                    .collect();
-
-                if good_values.is_empty() || bad_values.is_empty() {
-                    return self.sample_uniform(distribution, &mut rng);
-                }
-
-                let value = self.sample_tpe_int(
-                    d.low,
-                    d.high,
-                    d.log_scale,
-                    d.step,
-                    &good_values,
-                    &bad_values,
-                    &mut rng,
-                );
-                ParamValue::Int(value)
-            }
+            Distribution::Float(d) => self.sample_float(d, &good_trials, &bad_trials, &mut rng),
+            Distribution::Int(d) => self.sample_int(d, &good_trials, &bad_trials, &mut rng),
             Distribution::Categorical(d) => {
-                let good_indices: Vec<usize> = good_trials
-                    .iter()
-                    .flat_map(|t| t.params.values())
-                    .filter_map(|v| match v {
-                        ParamValue::Categorical(i) => Some(*i),
-                        _ => None,
-                    })
-                    .filter(|&i| i < d.n_choices)
-                    .collect();
-
-                let bad_indices: Vec<usize> = bad_trials
-                    .iter()
-                    .flat_map(|t| t.params.values())
-                    .filter_map(|v| match v {
-                        ParamValue::Categorical(i) => Some(*i),
-                        _ => None,
-                    })
-                    .filter(|&i| i < d.n_choices)
-                    .collect();
-
-                if good_indices.is_empty() || bad_indices.is_empty() {
-                    return self.sample_uniform(distribution, &mut rng);
-                }
-
-                let index =
-                    self.sample_tpe_categorical(d.n_choices, &good_indices, &bad_indices, &mut rng);
-                ParamValue::Categorical(index)
+                self.sample_categorical(d, &good_trials, &bad_trials, &mut rng)
             }
         }
     }
@@ -1083,8 +911,8 @@ mod tests {
         // With fewer than n_startup_trials, should use random sampling
         let history: Vec<CompletedTrial> = vec![];
 
-        for _ in 0..100 {
-            let value = sampler.sample(&dist, 0, &history);
+        for i in 0..100 {
+            let value = sampler.sample(&dist, i, &history);
             if let ParamValue::Float(v) = value {
                 assert!((0.0..=1.0).contains(&v));
             } else {

src/storage/journal.rs

@@ -0,0 +1,362 @@
+//! JSONL-based journal storage backend.
+//!
+//! [`JournalStorage`] persists completed trials as one JSON object per
+//! line ([JSONL / JSON Lines](https://jsonlines.org/)) while keeping a
+//! full copy in memory for fast read access.
+//!
+//! # File format
+//!
+//! Each line is a self-contained JSON serialization of a
+//! [`CompletedTrial<V>`](crate::sampler::CompletedTrial).  The file
+//! is append-only — no existing lines are ever modified or deleted.
+//!
+//! ```text
+//! {"id":0,"params":{...},"value":1.23,"state":"Completed",...}
+//! {"id":1,"params":{...},"value":0.87,"state":"Completed",...}
+//! ```
+//!
+//! # File locking
+//!
+//! Concurrent access is coordinated with `fs2` file locks:
+//!
+//! - **Writes** acquire an *exclusive* lock so only one process
+//!   appends at a time.
+//! - **Reads** ([`refresh`](super::Storage::refresh)) acquire a
+//!   *shared* lock so readers never see a partially written line.
+//!
+//! This makes it safe for multiple processes to share the same JSONL
+//! file — for example, distributed workers each running their own
+//! [`Study`](crate::Study) with a `JournalStorage` pointing to a
+//! shared path.
+//!
+//! # Resuming a study
+//!
+//! Use [`JournalStorage::open`] to reload previously persisted trials
+//! and continue optimization from where you left off:
+//!
+//! ```no_run
+//! use optimizer::prelude::*;
+//! use optimizer::storage::JournalStorage;
+//!
+//! // First run — creates the file.
+//! let storage = JournalStorage::<f64>::new("trials.jsonl");
+//! let mut study = Study::builder().minimize().storage(storage).build();
+//! study
+//!     .optimize(50, |trial: &mut optimizer::Trial| {
+//!         let x = FloatParam::new(-5.0, 5.0).suggest(trial)?;
+//!         Ok::<_, optimizer::Error>(x * x)
+//!     })
+//!     .unwrap();
+//!
+//! // Later run — reloads previous 50 trials, then adds 50 more.
+//! let storage = JournalStorage::<f64>::open("trials.jsonl").unwrap();
+//! let mut study = Study::builder().minimize().storage(storage).build();
+//! study
+//!     .optimize(50, |trial: &mut optimizer::Trial| {
+//!         let x = FloatParam::new(-5.0, 5.0).suggest(trial)?;
+//!         Ok::<_, optimizer::Error>(x * x)
+//!     })
+//!     .unwrap();
+//! ```
+//!
+//! # When to use
+//!
+//! - **Persistence** — survive process crashes or intentional restarts.
+//! - **Multi-process** — several workers collaborating on a single study.
+//! - **Inspection** — `cat trials.jsonl | jq .` for quick debugging.
+//!
+//! For pure in-memory usage without disk I/O, use
+//! [`MemoryStorage`](super::MemoryStorage) instead (the default).
+//!
+//! # Security considerations
+//!
+//! Both [`JournalStorage::open`] and [`refresh`](super::Storage::refresh)
+//! read the entire JSONL file into memory.  A very large file will
+//! consume memory proportional to its size, which could lead to
+//! out-of-memory conditions.
+//!
+//! If your application accepts externally-provided file paths, consider:
+//!
+//! - Checking the file size before calling `open`.
+//! - Validating or sanitizing untrusted JSONL content.
+//! - Imposing an upper bound on the number of trials or file size you
+//!   are willing to load.
+
+use core::marker::PhantomData;
+use core::sync::atomic::{AtomicU64, Ordering};
+use std::fs::{File, OpenOptions};
+use std::io::{BufRead, BufReader, Read as _, Seek, SeekFrom, Write};
+use std::path::{Path, PathBuf};
+use std::sync::Arc;
+
+use fs2::FileExt;
+use parking_lot::{Mutex, RwLock};
+use serde::Serialize;
+use serde::de::DeserializeOwned;
+
+use super::{MemoryStorage, Storage};
+use crate::sampler::CompletedTrial;
+
+/// Append-only JSONL storage backend with file locking.
+///
+/// Trials are kept in memory (via an inner [`MemoryStorage`]) for fast
+/// read access and simultaneously appended to a JSONL file on disk.
+/// Multiple processes can safely share the same file thanks to
+/// `fs2` file locks — writes use an exclusive lock, reads use a
+/// shared lock.
+///
+/// The type parameter `V` is the objective value type (typically
+/// `f64`).  It must implement [`Serialize`](serde::Serialize) and
+/// [`DeserializeOwned`](serde::de::DeserializeOwned) so trials can be
+/// written to and read from disk.
+///
+/// See the [`storage`](super) module docs for file format details
+/// and a resumption example.
+///
+/// # Example
+///
+/// ```no_run
+/// use optimizer::prelude::*;
+/// use optimizer::storage::JournalStorage;
+///
+/// let storage = JournalStorage::<f64>::new("trials.jsonl");
+/// let mut study = Study::builder().minimize().storage(storage).build();
+/// ```
+///
+/// # Security considerations
+///
+/// File contents are loaded into memory in full; see the
+/// module-level docs for details and mitigations.
+pub struct JournalStorage<V = f64> {
+    memory: MemoryStorage<V>,
+    path: PathBuf,
+    /// Serialise in-process writes and refreshes so they don't race.
+    io_lock: Mutex<()>,
+    /// Byte offset of last-read position for incremental refresh.
+    file_offset: AtomicU64,
+    _marker: PhantomData<V>,
+}
+
+impl<V: Serialize + DeserializeOwned + Send + Sync> JournalStorage<V> {
+    /// Create a new journal storage that writes to the given path.
+    ///
+    /// The file does not need to exist yet — it will be created on the
+    /// first write.  Existing trials in the file are **not** loaded
+    /// until [`refresh`](Storage::refresh) is called (which happens
+    /// automatically at the start of each trial via the [`Study`](crate::Study)).
+    ///
+    /// To pre-load existing trials at construction time, use
+    /// [`JournalStorage::open`] instead.
+    #[must_use]
+    pub fn new(path: impl AsRef<Path>) -> Self {
+        let path = path
+            .as_ref()
+            .canonicalize()
+            .unwrap_or_else(|_| path.as_ref().to_path_buf());
+        Self {
+            memory: MemoryStorage::new(),
+            path,
+            io_lock: Mutex::new(()),
+            file_offset: AtomicU64::new(0),
+            _marker: PhantomData,
+        }
+    }
+
+    /// Open an existing journal file and load all stored trials.
+    ///
+    /// If the file does not exist, return an empty storage (no error).
+    /// This is the primary way to **resume** a study after a restart.
+    ///
+    /// # Errors
+    ///
+    /// Return a [`Storage`](crate::Error::Storage) error if the file
+    /// exists but cannot be read or parsed.
+    pub fn open(path: impl AsRef<Path>) -> crate::Result<Self> {
+        let path = path
+            .as_ref()
+            .canonicalize()
+            .unwrap_or_else(|_| path.as_ref().to_path_buf());
+        let (trials, offset) = load_trials_from_file(&path)?;
+        Ok(Self {
+            memory: MemoryStorage::with_trials(trials),
+            path,
+            io_lock: Mutex::new(()),
+            file_offset: AtomicU64::new(offset),
+            _marker: PhantomData,
+        })
+    }
+
+    /// Append a single trial to the JSONL file (best-effort).
+    ///
+    /// Does **not** advance `file_offset` — that is left to `refresh`
+    /// so that externally-written data between the old offset and our
+    /// write is never skipped.
+    fn write_to_file(&self, trial: &CompletedTrial<V>) -> crate::Result<()> {
+        let _guard = self.io_lock.lock();
+
+        let mut file = OpenOptions::new()
+            .create(true)
+            .truncate(false)
+            .read(true)
+            .write(true)
+            .open(&self.path)
+            .map_err(|e| crate::Error::Storage(e.to_string()))?;
+
+        file.lock_exclusive()
+            .map_err(|e| crate::Error::Storage(e.to_string()))?;
+
+        file.seek(SeekFrom::End(0))
+            .map_err(|e| crate::Error::Storage(e.to_string()))?;
+
+        let line =
+            serde_json::to_string(trial).map_err(|e| crate::Error::Storage(e.to_string()))?;
+
+        writeln!(file, "{line}").map_err(|e| crate::Error::Storage(e.to_string()))?;
+        file.sync_data()
+            .map_err(|e| crate::Error::Storage(e.to_string()))?;
+
+        file.unlock()
+            .map_err(|e| crate::Error::Storage(e.to_string()))?;
+
+        Ok(())
+    }
+}
+
+impl<V: Serialize + DeserializeOwned + Send + Sync> Storage<V> for JournalStorage<V> {
+    fn push(&self, trial: CompletedTrial<V>) {
+        // Best-effort persist; the trial stays in memory regardless.
+        let _ = self.write_to_file(&trial);
+        self.memory.push(trial);
+    }
+
+    fn trials_arc(&self) -> &Arc<RwLock<Vec<CompletedTrial<V>>>> {
+        self.memory.trials_arc()
+    }
+
+    fn next_trial_id(&self) -> u64 {
+        self.memory.next_trial_id()
+    }
+
+    fn peek_next_trial_id(&self) -> u64 {
+        self.memory.peek_next_trial_id()
+    }
+
+    fn refresh(&self) -> bool {
+        let _guard = self.io_lock.lock();
+
+        let Ok(file) = File::open(&self.path) else {
+            return false;
+        };
+
+        if file.lock_shared().is_err() {
+            return false;
+        }
+
+        let offset = self.file_offset.load(Ordering::SeqCst);
+
+        let file_size = if let Ok(m) = file.metadata() {
+            m.len()
+        } else {
+            let _ = file.unlock();
+            return false;
+        };
+
+        if file_size <= offset {
+            let _ = file.unlock();
+            return false;
+        }
+
+        let mut buf = String::new();
+        let mut handle = &file;
+        if handle.seek(SeekFrom::Start(offset)).is_err() {
+            let _ = file.unlock();
+            return false;
+        }
+        if handle.read_to_string(&mut buf).is_err() {
+            let _ = file.unlock();
+            return false;
+        }
+
+        let _ = file.unlock();
+
+        let bytes_read = buf.len() as u64;
+        let new_offset = offset + bytes_read;
+        let mut new_trials = Vec::new();
+
+        for line in buf.lines() {
+            let line = line.trim();
+            if line.is_empty() {
+                continue;
+            }
+            let trial: CompletedTrial<V> = match serde_json::from_str(line) {
+                Ok(t) => t,
+                Err(_) => return false,
+            };
+            if trial.validate().is_err() {
+                return false;
+            }
+            new_trials.push(trial);
+        }
+
+        if new_trials.is_empty() {
+            self.file_offset.fetch_max(new_offset, Ordering::SeqCst);
+            return false;
+        }
+
+        let mut mem_guard = self.memory.trials_arc().write();
+
+        // Deduplicate: only add trials whose IDs are not already in memory.
+        let existing_ids: std::collections::HashSet<u64> = mem_guard.iter().map(|t| t.id).collect();
+        new_trials.retain(|t| !existing_ids.contains(&t.id));
+
+        if let Some(max_id) = new_trials.iter().map(|t| t.id).max() {
+            self.memory.bump_next_id(max_id + 1);
+        }
+        let added = !new_trials.is_empty();
+        mem_guard.extend(new_trials);
+        self.file_offset.fetch_max(new_offset, Ordering::SeqCst);
+        added
+    }
+}
+
+/// Read all trials from a JSONL file.  Returns an empty vec (and
+/// offset 0) if the file does not exist.  The returned `u64` is the
+/// file size at the time of reading, suitable for initialising the
+/// incremental-refresh offset.
+fn load_trials_from_file<V: DeserializeOwned>(
+    path: &Path,
+) -> crate::Result<(Vec<CompletedTrial<V>>, u64)> {
+    let file = match File::open(path) {
+        Ok(f) => f,
+        Err(e) if e.kind() == std::io::ErrorKind::NotFound => return Ok((Vec::new(), 0)),
+        Err(e) => return Err(crate::Error::Storage(e.to_string())),
+    };
+
+    file.lock_shared()
+        .map_err(|e| crate::Error::Storage(e.to_string()))?;
+
+    let file_size = file
+        .metadata()
+        .map_err(|e| crate::Error::Storage(e.to_string()))?
+        .len();
+
+    let reader = BufReader::new(&file);
+    let mut trials = Vec::new();
+
+    for line in reader.lines() {
+        let line = line.map_err(|e| crate::Error::Storage(e.to_string()))?;
+        let line = line.trim();
+        if line.is_empty() {
+            continue;
+        }
+        let trial: CompletedTrial<V> =
+            serde_json::from_str(line).map_err(|e| crate::Error::Storage(e.to_string()))?;
+        trial.validate().map_err(crate::Error::Storage)?;
+        trials.push(trial);
+    }
+
+    file.unlock()
+        .map_err(|e| crate::Error::Storage(e.to_string()))?;
+
+    Ok((trials, file_size))
+}

src/storage/memory.rs

@@ -0,0 +1,111 @@
+//! In-memory storage backend.
+//!
+//! [`MemoryStorage`] is the default backend used by every
+//! [`Study`](crate::Study).  Trials are stored in a
+//! `Vec<CompletedTrial<V>>` behind a [`parking_lot::RwLock`] for
+//! thread-safe access.
+//!
+//! # When to use
+//!
+//! - **Single-process** studies where persistence is not needed.
+//! - **Testing** or **prototyping** — zero configuration required.
+//! - When you want the **fastest** possible read/write performance
+//!   (no disk I/O).
+//!
+//! For persistent storage that survives process restarts, see
+//! `JournalStorage` (requires the `journal` feature).
+//!
+//! # Example
+//!
+//! ```
+//! use optimizer::prelude::*;
+//! use optimizer::storage::MemoryStorage;
+//!
+//! // Explicit memory storage (equivalent to the default)
+//! let storage = MemoryStorage::<f64>::new();
+//! let study = Study::builder().minimize().storage(storage).build();
+//! ```
+
+use core::sync::atomic::{AtomicU64, Ordering};
+use std::sync::Arc;
+
+use parking_lot::RwLock;
+
+use super::Storage;
+use crate::sampler::CompletedTrial;
+
+/// In-memory trial storage (the default).
+///
+/// Wrap a `Vec<CompletedTrial<V>>` behind a read-write lock so that
+/// trials can be appended from any thread.  This is the backend that
+/// [`Study`](crate::Study) uses when no explicit storage is provided.
+///
+/// Use [`with_trials`](Self::with_trials) to seed a study with
+/// previously collected data.
+///
+/// # Example
+///
+/// ```
+/// use optimizer::storage::{MemoryStorage, Storage};
+///
+/// let storage = MemoryStorage::<f64>::new();
+/// assert_eq!(storage.trials_arc().read().len(), 0);
+/// ```
+pub struct MemoryStorage<V> {
+    trials: Arc<RwLock<Vec<CompletedTrial<V>>>>,
+    next_id: AtomicU64,
+}
+
+impl<V> MemoryStorage<V> {
+    /// Create a new, empty in-memory store.
+    #[must_use]
+    pub fn new() -> Self {
+        Self {
+            trials: Arc::new(RwLock::new(Vec::new())),
+            next_id: AtomicU64::new(0),
+        }
+    }
+
+    /// Create an in-memory store pre-populated with `trials`.
+    ///
+    /// The internal ID counter is set to one past the highest trial ID
+    /// so that subsequent trials receive unique IDs.
+    #[must_use]
+    pub fn with_trials(trials: Vec<CompletedTrial<V>>) -> Self {
+        let next_id = trials.iter().map(|t| t.id).max().map_or(0, |id| id + 1);
+        Self {
+            trials: Arc::new(RwLock::new(trials)),
+            next_id: AtomicU64::new(next_id),
+        }
+    }
+
+    /// Ensure the ID counter is at least `min_value`.
+    #[cfg(feature = "journal")]
+    pub(crate) fn bump_next_id(&self, min_value: u64) {
+        self.next_id.fetch_max(min_value, Ordering::SeqCst);
+    }
+}
+
+impl<V> Default for MemoryStorage<V> {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl<V: Send + Sync> Storage<V> for MemoryStorage<V> {
+    fn push(&self, trial: CompletedTrial<V>) {
+        self.trials.write().push(trial);
+    }
+
+    fn trials_arc(&self) -> &Arc<RwLock<Vec<CompletedTrial<V>>>> {
+        &self.trials
+    }
+
+    fn next_trial_id(&self) -> u64 {
+        self.next_id.fetch_add(1, Ordering::SeqCst)
+    }
+
+    fn peek_next_trial_id(&self) -> u64 {
+        self.next_id.load(Ordering::SeqCst)
+    }
+}

src/storage/mod.rs

@@ -0,0 +1,94 @@
+//! Trial storage backends.
+//!
+//! The [`Storage`] trait defines how completed trials are persisted and
+//! retrieved.  Every [`Study`](crate::Study) owns an `Arc<dyn Storage<V>>`
+//! so storage is transparently shared across threads.
+//!
+//! # Available backends
+//!
+//! | Backend | Description | Feature flag |
+//! |---------|-------------|-------------|
+//! | [`MemoryStorage`] | In-memory `Vec` behind a read-write lock (the default) | — |
+//! | `JournalStorage` | JSONL file with `fs2` file locking for multi-process sharing | `journal` |
+//!
+//! # When to swap backends
+//!
+//! The default [`MemoryStorage`] is sufficient for single-process studies
+//! where persistence is not needed.  Switch to `JournalStorage` when you
+//! want to:
+//!
+//! - **Resume** a study after a process restart.
+//! - **Share state** across multiple processes writing to the same file.
+//! - **Inspect** trial history in a human-readable JSONL file.
+//!
+//! # Implementing a custom backend
+//!
+//! Implement the [`Storage`] trait to plug in your own backend (e.g. a
+//! database).  The trait requires four methods: [`push`](Storage::push),
+//! [`trials_arc`](Storage::trials_arc), [`next_trial_id`](Storage::next_trial_id),
+//! and optionally [`refresh`](Storage::refresh) for external data sources.
+//!
+//! Inject your storage into a study via the builder:
+//!
+//! ```
+//! use optimizer::prelude::*;
+//! use optimizer::storage::MemoryStorage;
+//!
+//! let storage = MemoryStorage::<f64>::new();
+//! let study = Study::builder().minimize().storage(storage).build();
+//! ```
+
+#[cfg(feature = "journal")]
+mod journal;
+
+use std::sync::Arc;
+
+#[cfg(feature = "journal")]
+pub use journal::JournalStorage;
+use parking_lot::RwLock;
+
+mod memory;
+pub use memory::MemoryStorage;
+
+use crate::sampler::CompletedTrial;
+
+/// Trait for storing and retrieving completed trials.
+///
+/// Every [`Study`](crate::Study) owns an `Arc<dyn Storage<V>>`.  The
+/// default implementation is [`MemoryStorage`], which keeps trials in
+/// a plain `Vec` behind a read-write lock.
+///
+/// Implementations must be `Send + Sync` because a study may be shared
+/// across threads (e.g. via `Study::optimize_parallel`).
+pub trait Storage<V>: Send + Sync {
+    /// Append a completed trial to the store.
+    fn push(&self, trial: CompletedTrial<V>);
+
+    /// Return a reference to the in-memory trial buffer.
+    ///
+    /// All implementations must maintain an `Arc<RwLock<Vec<…>>>` that
+    /// reflects the current set of trials.  Callers may acquire a read
+    /// lock for efficient, allocation-free access.
+    fn trials_arc(&self) -> &Arc<RwLock<Vec<CompletedTrial<V>>>>;
+
+    /// Atomically return the next unique trial ID.
+    ///
+    /// Each call increments an internal counter so that consecutive
+    /// calls always produce distinct IDs.
+    fn next_trial_id(&self) -> u64;
+
+    /// Return the current value of the next-trial-ID counter without incrementing.
+    ///
+    /// This is used for persistence (e.g. `Study::save`) to capture the
+    /// counter's exact position, including IDs assigned to failed trials
+    /// that are not stored.
+    fn peek_next_trial_id(&self) -> u64;
+
+    /// Reload from an external source (e.g. a file written by another
+    /// process).  Return `true` if the in-memory buffer was updated.
+    ///
+    /// The default implementation is a no-op that returns `false`.
+    fn refresh(&self) -> bool {
+        false
+    }
+}

src/study/analysis.rs

@@ -0,0 +1,379 @@
+use crate::sampler::CompletedTrial;
+use crate::types::TrialState;
+
+use super::Study;
+
+impl<V> Study<V>
+where
+    V: PartialOrd,
+{
+    /// Return the trial with the best objective value.
+    ///
+    /// The "best" trial depends on the optimization direction:
+    /// - `Direction::Minimize`: Returns the trial with the lowest objective value.
+    /// - `Direction::Maximize`: Returns the trial with the highest objective value.
+    ///
+    /// When constraints are present, feasible trials always rank above infeasible
+    /// trials. Among infeasible trials, those with lower total constraint violation
+    /// are preferred.
+    ///
+    /// # Errors
+    ///
+    /// Returns `Error::NoCompletedTrials` if no trials have been completed.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::parameter::{FloatParam, Parameter};
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let study: Study<f64> = Study::new(Direction::Minimize);
+    ///
+    /// // Error when no trials completed
+    /// assert!(study.best_trial().is_err());
+    ///
+    /// let x_param = FloatParam::new(0.0, 1.0);
+    ///
+    /// let mut trial1 = study.create_trial();
+    /// let _ = x_param.suggest(&mut trial1);
+    /// study.complete_trial(trial1, 0.8);
+    ///
+    /// let mut trial2 = study.create_trial();
+    /// let _ = x_param.suggest(&mut trial2);
+    /// study.complete_trial(trial2, 0.3);
+    ///
+    /// let best = study.best_trial().unwrap();
+    /// assert_eq!(best.value, 0.3); // Minimize: lower is better
+    /// ```
+    pub fn best_trial(&self) -> crate::Result<CompletedTrial<V>>
+    where
+        V: Clone,
+    {
+        let trials = self.storage.trials_arc().read();
+        let direction = self.direction;
+
+        let best = trials
+            .iter()
+            .filter(|t| t.state == TrialState::Complete)
+            .max_by(|a, b| Self::compare_trials(a, b, direction))
+            .ok_or(crate::Error::NoCompletedTrials)?;
+
+        Ok(best.clone())
+    }
+
+    /// Return the best objective value found so far.
+    ///
+    /// The "best" value depends on the optimization direction:
+    /// - `Direction::Minimize`: Returns the lowest objective value.
+    /// - `Direction::Maximize`: Returns the highest objective value.
+    ///
+    /// # Errors
+    ///
+    /// Returns `Error::NoCompletedTrials` if no trials have been completed.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::parameter::{FloatParam, Parameter};
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let study: Study<f64> = Study::new(Direction::Maximize);
+    ///
+    /// // Error when no trials completed
+    /// assert!(study.best_value().is_err());
+    ///
+    /// let x_param = FloatParam::new(0.0, 1.0);
+    ///
+    /// let mut trial1 = study.create_trial();
+    /// let _ = x_param.suggest(&mut trial1);
+    /// study.complete_trial(trial1, 0.3);
+    ///
+    /// let mut trial2 = study.create_trial();
+    /// let _ = x_param.suggest(&mut trial2);
+    /// study.complete_trial(trial2, 0.8);
+    ///
+    /// let best = study.best_value().unwrap();
+    /// assert_eq!(best, 0.8); // Maximize: higher is better
+    /// ```
+    pub fn best_value(&self) -> crate::Result<V>
+    where
+        V: Clone,
+    {
+        self.best_trial().map(|trial| trial.value)
+    }
+
+    /// Return the top `n` trials sorted by objective value.
+    ///
+    /// For `Direction::Minimize`, returns trials with the lowest values.
+    /// For `Direction::Maximize`, returns trials with the highest values.
+    /// Only includes completed trials (not failed or pruned).
+    ///
+    /// If fewer than `n` completed trials exist, returns all of them.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::parameter::{FloatParam, Parameter};
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let study: Study<f64> = Study::new(Direction::Minimize);
+    /// let x = FloatParam::new(0.0, 10.0);
+    ///
+    /// for val in [5.0, 1.0, 3.0] {
+    ///     let mut t = study.create_trial();
+    ///     let _ = x.suggest(&mut t);
+    ///     study.complete_trial(t, val);
+    /// }
+    ///
+    /// let top2 = study.top_trials(2);
+    /// assert_eq!(top2.len(), 2);
+    /// assert!(top2[0].value <= top2[1].value);
+    /// ```
+    #[must_use]
+    pub fn top_trials(&self, n: usize) -> Vec<CompletedTrial<V>>
+    where
+        V: Clone,
+    {
+        let trials = self.storage.trials_arc().read();
+        let direction = self.direction;
+        // Sort indices instead of cloning all trials, then clone only the top N.
+        let mut indices: Vec<usize> = trials
+            .iter()
+            .enumerate()
+            .filter(|(_, t)| t.state == TrialState::Complete)
+            .map(|(i, _)| i)
+            .collect();
+        // Sort best-first: reverse the compare_trials ordering (which is designed for max_by)
+        indices.sort_by(|&a, &b| Self::compare_trials(&trials[b], &trials[a], direction));
+        indices.truncate(n);
+        indices.iter().map(|&i| trials[i].clone()).collect()
+    }
+}
+
+impl<V> Study<V>
+where
+    V: PartialOrd + Clone + Into<f64>,
+{
+    /// Compute parameter importance scores using Spearman rank correlation.
+    ///
+    /// For each parameter, the absolute Spearman correlation between its values
+    /// and the objective values is computed across all completed trials. Scores
+    /// are normalized so they sum to 1.0 and sorted in descending order.
+    ///
+    /// Parameters that appear in fewer than 2 trials are omitted.
+    /// Returns an empty `Vec` if the study has fewer than 2 completed trials.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::parameter::{FloatParam, Parameter};
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let study: Study<f64> = Study::new(Direction::Minimize);
+    /// let x = FloatParam::new(0.0, 10.0).name("x");
+    ///
+    /// study
+    ///     .optimize(20, |trial: &mut optimizer::Trial| {
+    ///         let xv = x.suggest(trial)?;
+    ///         Ok::<_, optimizer::Error>(xv * xv)
+    ///     })
+    ///     .unwrap();
+    ///
+    /// let importance = study.param_importance();
+    /// assert_eq!(importance.len(), 1);
+    /// assert_eq!(importance[0].0, "x");
+    /// ```
+    #[must_use]
+    #[allow(clippy::cast_precision_loss)]
+    pub fn param_importance(&self) -> Vec<(String, f64)> {
+        use std::collections::BTreeSet;
+
+        use crate::importance::spearman;
+        use crate::param::ParamValue;
+        use crate::types::TrialState;
+
+        let trials = self.storage.trials_arc().read();
+        let complete: Vec<_> = trials
+            .iter()
+            .filter(|t| t.state == TrialState::Complete)
+            .collect();
+
+        if complete.len() < 2 {
+            return Vec::new();
+        }
+
+        // Collect all parameter IDs across trials.
+        let all_param_ids: BTreeSet<_> = complete.iter().flat_map(|t| t.params.keys()).collect();
+
+        let mut scores: Vec<(String, f64)> = Vec::with_capacity(all_param_ids.len());
+
+        for &param_id in &all_param_ids {
+            // Collect (param_value_f64, objective_f64) for trials that have this param.
+            let mut param_vals = Vec::with_capacity(complete.len());
+            let mut obj_vals = Vec::with_capacity(complete.len());
+
+            for trial in &complete {
+                if let Some(pv) = trial.params.get(param_id) {
+                    let f = match *pv {
+                        ParamValue::Float(v) => v,
+                        ParamValue::Int(v) => v as f64,
+                        ParamValue::Categorical(v) => v as f64,
+                    };
+                    param_vals.push(f);
+                    obj_vals.push(trial.value.clone().into());
+                }
+            }
+
+            if param_vals.len() < 2 {
+                continue;
+            }
+
+            let corr = spearman(&param_vals, &obj_vals).abs();
+
+            // Determine label: use param_labels if available, else "param_{id}".
+            let label = complete
+                .iter()
+                .find_map(|t| t.param_labels.get(param_id))
+                .map_or_else(|| param_id.to_string(), Clone::clone);
+
+            scores.push((label, corr));
+        }
+
+        // Normalize so scores sum to 1.0.
+        let sum: f64 = scores.iter().map(|(_, s)| *s).sum();
+        if sum > 0.0 {
+            for entry in &mut scores {
+                entry.1 /= sum;
+            }
+        }
+
+        // Sort descending by score.
+        scores.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap_or(core::cmp::Ordering::Equal));
+
+        scores
+    }
+
+    /// Compute parameter importance using fANOVA (functional ANOVA) with
+    /// default configuration.
+    ///
+    /// Fits a random forest to the trial data and decomposes variance into
+    /// per-parameter main effects and pairwise interaction effects. This is
+    /// more accurate than correlation-based importance ([`Self::param_importance`])
+    /// and can detect non-linear relationships and parameter interactions.
+    ///
+    /// # Errors
+    ///
+    /// Returns [`crate::Error::NoCompletedTrials`] if fewer than 2 trials have completed.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::parameter::{FloatParam, Parameter};
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let study: Study<f64> = Study::new(Direction::Minimize);
+    /// let x = FloatParam::new(0.0, 10.0).name("x");
+    /// let y = FloatParam::new(0.0, 10.0).name("y");
+    ///
+    /// study
+    ///     .optimize(30, |trial: &mut optimizer::Trial| {
+    ///         let xv = x.suggest(trial)?;
+    ///         let yv = y.suggest(trial)?;
+    ///         Ok::<_, optimizer::Error>(xv * xv + 0.1 * yv)
+    ///     })
+    ///     .unwrap();
+    ///
+    /// let result = study.fanova().unwrap();
+    /// assert!(!result.main_effects.is_empty());
+    /// ```
+    pub fn fanova(&self) -> crate::Result<crate::fanova::FanovaResult> {
+        self.fanova_with_config(&crate::fanova::FanovaConfig::default())
+    }
+
+    /// Compute parameter importance using fANOVA with custom configuration.
+    ///
+    /// See [`Self::fanova`] for details. The [`FanovaConfig`](crate::fanova::FanovaConfig)
+    /// allows tuning the number of trees, tree depth, and random seed.
+    ///
+    /// # Errors
+    ///
+    /// Returns [`crate::Error::NoCompletedTrials`] if fewer than 2 trials have completed.
+    #[allow(clippy::cast_precision_loss)]
+    pub fn fanova_with_config(
+        &self,
+        config: &crate::fanova::FanovaConfig,
+    ) -> crate::Result<crate::fanova::FanovaResult> {
+        use std::collections::BTreeSet;
+
+        use crate::fanova::compute_fanova;
+        use crate::param::ParamValue;
+        use crate::types::TrialState;
+
+        let trials = self.storage.trials_arc().read();
+        let complete: Vec<_> = trials
+            .iter()
+            .filter(|t| t.state == TrialState::Complete)
+            .collect();
+
+        if complete.len() < 2 {
+            return Err(crate::Error::NoCompletedTrials);
+        }
+
+        // Collect all parameter IDs in a stable order.
+        let all_param_ids: Vec<_> = {
+            let set: BTreeSet<_> = complete.iter().flat_map(|t| t.params.keys()).collect();
+            set.into_iter().collect()
+        };
+
+        if all_param_ids.is_empty() {
+            return Ok(crate::fanova::FanovaResult {
+                main_effects: Vec::new(),
+                interactions: Vec::new(),
+            });
+        }
+
+        // Build feature matrix (only trials that have all parameters).
+        let mut data = Vec::with_capacity(complete.len());
+        let mut targets = Vec::with_capacity(complete.len());
+
+        for trial in &complete {
+            let mut row = Vec::with_capacity(all_param_ids.len());
+            let mut has_all = true;
+
+            for &pid in &all_param_ids {
+                if let Some(pv) = trial.params.get(pid) {
+                    row.push(match *pv {
+                        ParamValue::Float(v) => v,
+                        ParamValue::Int(v) => v as f64,
+                        ParamValue::Categorical(v) => v as f64,
+                    });
+                } else {
+                    has_all = false;
+                    break;
+                }
+            }
+
+            if has_all {
+                data.push(row);
+                targets.push(trial.value.clone().into());
+            }
+        }
+
+        if data.len() < 2 {
+            return Err(crate::Error::NoCompletedTrials);
+        }
+
+        // Build feature names from parameter labels.
+        let feature_names: Vec<String> = all_param_ids
+            .iter()
+            .map(|&pid| {
+                complete
+                    .iter()
+                    .find_map(|t| t.param_labels.get(pid))
+                    .map_or_else(|| pid.to_string(), Clone::clone)
+            })
+            .collect();
+
+        Ok(compute_fanova(&data, &targets, &feature_names, config))
+    }
+}

src/study/async_impl.rs

@@ -0,0 +1,285 @@
+use core::ops::ControlFlow;
+use std::sync::Arc;
+
+use crate::trial::Trial;
+use crate::types::TrialState;
+
+use super::{Study, is_trial_pruned};
+
+impl<V> Study<V>
+where
+    V: PartialOrd,
+{
+    /// Run async optimization with an objective.
+    ///
+    /// Like [`optimize`](Self::optimize), but each evaluation is wrapped in
+    /// [`spawn_blocking`](tokio::task::spawn_blocking), keeping the async
+    /// runtime responsive for CPU-bound objectives. Trials run sequentially.
+    ///
+    /// Accepts any [`Objective`](crate::Objective) implementation, including
+    /// plain closures. Struct-based objectives can override lifecycle hooks.
+    ///
+    /// # Errors
+    ///
+    /// Returns `Error::NoCompletedTrials` if no trials completed successfully.
+    /// Returns `Error::TaskError` if a spawned blocking task panics.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::parameter::{FloatParam, Parameter};
+    /// use optimizer::sampler::random::RandomSampler;
+    /// use optimizer::{Direction, Study};
+    ///
+    /// # #[cfg(feature = "async")]
+    /// # async fn example() -> optimizer::Result<()> {
+    /// let sampler = RandomSampler::with_seed(42);
+    /// let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+    /// let x_param = FloatParam::new(-10.0, 10.0);
+    ///
+    /// study
+    ///     .optimize_async(10, move |trial: &mut optimizer::Trial| {
+    ///         let x = x_param.suggest(trial)?;
+    ///         Ok::<_, optimizer::Error>(x * x)
+    ///     })
+    ///     .await?;
+    ///
+    /// assert!(study.n_trials() > 0);
+    /// # Ok(())
+    /// # }
+    /// ```
+    pub async fn optimize_async<O>(&self, n_trials: usize, objective: O) -> crate::Result<()>
+    where
+        O: crate::objective::Objective<V> + Send + Sync + 'static,
+        O::Error: Send,
+        V: Clone + Default + Send + 'static,
+    {
+        #[cfg(feature = "tracing")]
+        let _span =
+            tracing::info_span!("optimize_async", n_trials, direction = ?self.direction).entered();
+
+        let objective = Arc::new(objective);
+
+        for _ in 0..n_trials {
+            if let ControlFlow::Break(()) = objective.before_trial(self) {
+                break;
+            }
+
+            let obj = Arc::clone(&objective);
+            let mut trial = self.create_trial();
+            let result = tokio::task::spawn_blocking(move || {
+                let res = obj.evaluate(&mut trial);
+                (trial, res)
+            })
+            .await
+            .map_err(|e| crate::Error::TaskError(e.to_string()))?;
+
+            match result {
+                (t, Ok(value)) => {
+                    #[cfg(feature = "tracing")]
+                    let trial_id = t.id();
+
+                    let completed = t.into_completed(value, TrialState::Complete);
+                    let flow = objective.after_trial(self, &completed);
+                    self.storage.push(completed);
+                    trace_info!(trial_id, "trial completed");
+
+                    if let ControlFlow::Break(()) = flow {
+                        return Ok(());
+                    }
+                }
+                (t, Err(e)) if is_trial_pruned(&e) => {
+                    #[cfg(feature = "tracing")]
+                    let trial_id = t.id();
+                    self.prune_trial(t);
+                    trace_info!(trial_id, "trial pruned");
+                }
+                (t, Err(e)) => {
+                    #[cfg(feature = "tracing")]
+                    let trial_id = t.id();
+                    self.fail_trial(t, e.to_string());
+                    trace_debug!(trial_id, "trial failed");
+                }
+            }
+        }
+
+        let has_complete = self
+            .storage
+            .trials_arc()
+            .read()
+            .iter()
+            .any(|t| t.state == TrialState::Complete);
+        if !has_complete {
+            return Err(crate::Error::NoCompletedTrials);
+        }
+
+        Ok(())
+    }
+
+    /// Run parallel optimization with an objective.
+    ///
+    /// Spawns up to `concurrency` evaluations concurrently using
+    /// [`spawn_blocking`](tokio::task::spawn_blocking). Results are
+    /// collected via a [`JoinSet`](tokio::task::JoinSet).
+    ///
+    /// Accepts any [`Objective`](crate::Objective) implementation, including
+    /// plain closures. The [`after_trial`](crate::Objective::after_trial)
+    /// hook fires as each result arrives — returning `Break` stops spawning
+    /// new trials while in-flight tasks drain.
+    ///
+    /// # Errors
+    ///
+    /// Returns `Error::NoCompletedTrials` if no trials completed successfully.
+    /// Returns `Error::TaskError` if the semaphore is closed or a spawned task panics.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::parameter::{FloatParam, Parameter};
+    /// use optimizer::sampler::random::RandomSampler;
+    /// use optimizer::{Direction, Study};
+    ///
+    /// # #[cfg(feature = "async")]
+    /// # async fn example() -> optimizer::Result<()> {
+    /// let sampler = RandomSampler::with_seed(42);
+    /// let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+    /// let x_param = FloatParam::new(-10.0, 10.0);
+    ///
+    /// study
+    ///     .optimize_parallel(10, 4, move |trial: &mut optimizer::Trial| {
+    ///         let x = x_param.suggest(trial)?;
+    ///         Ok::<_, optimizer::Error>(x * x)
+    ///     })
+    ///     .await?;
+    ///
+    /// assert_eq!(study.n_trials(), 10);
+    /// # Ok(())
+    /// # }
+    /// ```
+    #[allow(clippy::missing_panics_doc, clippy::too_many_lines)]
+    pub async fn optimize_parallel<O>(
+        &self,
+        n_trials: usize,
+        concurrency: usize,
+        objective: O,
+    ) -> crate::Result<()>
+    where
+        O: crate::objective::Objective<V> + Send + Sync + 'static,
+        O::Error: Send,
+        V: Clone + Default + Send + 'static,
+    {
+        use tokio::sync::Semaphore;
+        use tokio::task::JoinSet;
+
+        assert!(concurrency > 0, "concurrency must be at least 1");
+
+        #[cfg(feature = "tracing")]
+        let _span = tracing::info_span!("optimize_parallel", n_trials, concurrency, direction = ?self.direction).entered();
+
+        let objective = Arc::new(objective);
+        let semaphore = Arc::new(Semaphore::new(concurrency));
+        let mut join_set: JoinSet<(Trial, Result<V, O::Error>)> = JoinSet::new();
+        let mut spawned = 0;
+
+        'spawn: while spawned < n_trials {
+            if let ControlFlow::Break(()) = objective.before_trial(self) {
+                break;
+            }
+
+            // If the join set is full, drain one result to free a slot.
+            while join_set.len() >= concurrency {
+                let result = join_set
+                    .join_next()
+                    .await
+                    .expect("join_set should not be empty")
+                    .map_err(|e| crate::Error::TaskError(e.to_string()))?;
+                match result {
+                    (t, Ok(value)) => {
+                        #[cfg(feature = "tracing")]
+                        let trial_id = t.id();
+
+                        let completed = t.into_completed(value, TrialState::Complete);
+                        let flow = objective.after_trial(self, &completed);
+                        self.storage.push(completed);
+                        trace_info!(trial_id, "trial completed");
+
+                        if let ControlFlow::Break(()) = flow {
+                            break 'spawn;
+                        }
+                    }
+                    (t, Err(e)) => {
+                        #[cfg(feature = "tracing")]
+                        let trial_id = t.id();
+                        if is_trial_pruned(&e) {
+                            self.prune_trial(t);
+                            trace_info!(trial_id, "trial pruned");
+                        } else {
+                            self.fail_trial(t, e.to_string());
+                            trace_debug!(trial_id, "trial failed");
+                        }
+                    }
+                }
+            }
+
+            let permit = semaphore
+                .clone()
+                .acquire_owned()
+                .await
+                .map_err(|e| crate::Error::TaskError(e.to_string()))?;
+
+            let mut trial = self.create_trial();
+            let obj = Arc::clone(&objective);
+            join_set.spawn(async move {
+                let result = tokio::task::spawn_blocking(move || {
+                    let res = obj.evaluate(&mut trial);
+                    (trial, res)
+                })
+                .await
+                .expect("spawn_blocking should not panic");
+                drop(permit);
+                result
+            });
+            spawned += 1;
+        }
+
+        // Drain remaining in-flight tasks.
+        while let Some(result) = join_set.join_next().await {
+            let result = result.map_err(|e| crate::Error::TaskError(e.to_string()))?;
+            match result {
+                (t, Ok(value)) => {
+                    #[cfg(feature = "tracing")]
+                    let trial_id = t.id();
+
+                    let completed = t.into_completed(value, TrialState::Complete);
+                    // Still fire after_trial for bookkeeping, but don't break — we're draining.
+                    let _ = objective.after_trial(self, &completed);
+                    self.storage.push(completed);
+                    trace_info!(trial_id, "trial completed");
+                }
+                (t, Err(e)) => {
+                    #[cfg(feature = "tracing")]
+                    let trial_id = t.id();
+                    if is_trial_pruned(&e) {
+                        self.prune_trial(t);
+                        trace_info!(trial_id, "trial pruned");
+                    } else {
+                        self.fail_trial(t, e.to_string());
+                        trace_debug!(trial_id, "trial failed");
+                    }
+                }
+            }
+        }
+
+        let has_complete = self
+            .storage
+            .trials_arc()
+            .read()
+            .iter()
+            .any(|t| t.state == TrialState::Complete);
+        if !has_complete {
+            return Err(crate::Error::NoCompletedTrials);
+        }
+
+        Ok(())
+    }
+}

src/study/builder.rs

@@ -0,0 +1,135 @@
+use core::marker::PhantomData;
+use std::collections::VecDeque;
+use std::sync::Arc;
+
+use parking_lot::Mutex;
+
+use crate::pruner::{NopPruner, Pruner};
+use crate::sampler::Sampler;
+use crate::sampler::random::RandomSampler;
+use crate::types::Direction;
+
+use super::Study;
+
+/// A builder for constructing [`Study`] instances with a fluent API.
+///
+/// Created via [`Study::builder()`]. Collects sampler, pruner, direction,
+/// and storage options before constructing the study.
+///
+/// # Defaults
+///
+/// - Direction: [`Minimize`](Direction::Minimize)
+/// - Sampler: [`RandomSampler`]
+/// - Pruner: [`NopPruner`]
+/// - Storage: [`MemoryStorage`](crate::storage::MemoryStorage)
+///
+/// # Examples
+///
+/// ```
+/// use optimizer::prelude::*;
+///
+/// let study: Study<f64> = Study::builder()
+///     .maximize()
+///     .sampler(TpeSampler::new())
+///     .pruner(MedianPruner::new(Direction::Maximize).n_warmup_steps(5))
+///     .build();
+///
+/// assert_eq!(study.direction(), Direction::Maximize);
+/// ```
+pub struct StudyBuilder<V: PartialOrd = f64> {
+    direction: Direction,
+    sampler: Option<Box<dyn Sampler>>,
+    pruner: Option<Box<dyn Pruner>>,
+    storage: Option<Box<dyn crate::storage::Storage<V>>>,
+    _marker: PhantomData<V>,
+}
+
+impl<V: PartialOrd> StudyBuilder<V> {
+    /// Create a new builder with default settings.
+    pub(super) fn new() -> Self {
+        Self {
+            direction: Direction::Minimize,
+            sampler: None,
+            pruner: None,
+            storage: None,
+            _marker: PhantomData,
+        }
+    }
+
+    /// Set the optimization direction to minimize (the default).
+    #[must_use]
+    pub fn minimize(mut self) -> Self {
+        self.direction = Direction::Minimize;
+        self
+    }
+
+    /// Set the optimization direction to maximize.
+    #[must_use]
+    pub fn maximize(mut self) -> Self {
+        self.direction = Direction::Maximize;
+        self
+    }
+
+    /// Set the optimization direction explicitly.
+    #[must_use]
+    pub fn direction(mut self, direction: Direction) -> Self {
+        self.direction = direction;
+        self
+    }
+
+    /// Set the sampler used for parameter suggestions.
+    ///
+    /// Defaults to [`RandomSampler`] if not specified.
+    #[must_use]
+    pub fn sampler(mut self, sampler: impl Sampler + 'static) -> Self {
+        self.sampler = Some(Box::new(sampler));
+        self
+    }
+
+    /// Set the pruner used for early stopping of trials.
+    ///
+    /// Defaults to [`NopPruner`] (no pruning) if not specified.
+    #[must_use]
+    pub fn pruner(mut self, pruner: impl Pruner + 'static) -> Self {
+        self.pruner = Some(Box::new(pruner));
+        self
+    }
+
+    /// Set a custom storage backend.
+    ///
+    /// Defaults to [`MemoryStorage`](crate::storage::MemoryStorage) if not specified.
+    #[must_use]
+    pub fn storage(mut self, storage: impl crate::storage::Storage<V> + 'static) -> Self {
+        self.storage = Some(Box::new(storage));
+        self
+    }
+
+    /// Build the [`Study`] with the configured options.
+    #[must_use]
+    pub fn build(self) -> Study<V>
+    where
+        V: Send + Sync + 'static,
+    {
+        let sampler = self
+            .sampler
+            .unwrap_or_else(|| Box::new(RandomSampler::new()));
+        let pruner = self.pruner.unwrap_or_else(|| Box::new(NopPruner));
+        let storage = self
+            .storage
+            .unwrap_or_else(|| Box::new(crate::storage::MemoryStorage::<V>::new()));
+
+        let sampler: Arc<dyn Sampler> = Arc::from(sampler);
+        let pruner: Arc<dyn Pruner> = Arc::from(pruner);
+        let storage: Arc<dyn crate::storage::Storage<V>> = Arc::from(storage);
+        let trial_factory = Study::make_trial_factory(&sampler, &storage, &pruner);
+
+        Study {
+            direction: self.direction,
+            sampler,
+            pruner,
+            storage,
+            trial_factory,
+            enqueued_params: Arc::new(Mutex::new(VecDeque::new())),
+        }
+    }
+}

src/study/export.rs

@@ -0,0 +1,274 @@
+use core::fmt;
+
+use crate::parameter::ParamId;
+use crate::types::{Direction, TrialState};
+
+use super::Study;
+
+impl<V> Study<V>
+where
+    V: PartialOrd + Clone + fmt::Display,
+{
+    /// Write completed trials to a writer in CSV format.
+    ///
+    /// Columns: `trial_id`, `value`, `state`, then one column per unique
+    /// parameter label, then one column per unique user-attribute key.
+    ///
+    /// Parameters without labels use a generated name (`param_<id>`).
+    /// Pruned trials have an empty `value` cell.
+    ///
+    /// # Errors
+    ///
+    /// Returns an I/O error if writing fails.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::parameter::{FloatParam, Parameter};
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let study: Study<f64> = Study::new(Direction::Minimize);
+    /// let x = FloatParam::new(0.0, 10.0).name("x");
+    ///
+    /// let mut trial = study.create_trial();
+    /// let _ = x.suggest(&mut trial);
+    /// study.complete_trial(trial, 0.42);
+    ///
+    /// let mut buf = Vec::new();
+    /// study.to_csv(&mut buf).unwrap();
+    /// let csv = String::from_utf8(buf).unwrap();
+    /// assert!(csv.contains("trial_id"));
+    /// ```
+    pub fn to_csv(&self, mut writer: impl std::io::Write) -> std::io::Result<()> {
+        use std::collections::BTreeMap;
+
+        let trials = self.storage.trials_arc().read();
+
+        // Collect all unique parameter labels (sorted for deterministic column order).
+        let mut param_columns: BTreeMap<ParamId, String> = BTreeMap::new();
+        for trial in trials.iter() {
+            for &id in trial.params.keys() {
+                param_columns.entry(id).or_insert_with(|| {
+                    trial
+                        .param_labels
+                        .get(&id)
+                        .cloned()
+                        .unwrap_or_else(|| id.to_string())
+                });
+            }
+        }
+        // Fill in labels from other trials that might have better labels.
+        for trial in trials.iter() {
+            for (&id, label) in &trial.param_labels {
+                param_columns.entry(id).or_insert_with(|| label.clone());
+            }
+        }
+
+        // Collect all unique attribute keys (sorted).
+        let mut attr_keys: Vec<String> = Vec::new();
+        for trial in trials.iter() {
+            for key in trial.user_attrs.keys() {
+                if !attr_keys.contains(key) {
+                    attr_keys.push(key.clone());
+                }
+            }
+        }
+        attr_keys.sort();
+
+        let param_ids: Vec<ParamId> = param_columns.keys().copied().collect();
+
+        // Write header.
+        write!(writer, "trial_id,value,state")?;
+        for id in &param_ids {
+            write!(writer, ",{}", csv_escape(&param_columns[id]))?;
+        }
+        for key in &attr_keys {
+            write!(writer, ",{}", csv_escape(key))?;
+        }
+        writeln!(writer)?;
+
+        // Write one row per trial.
+        for trial in trials.iter() {
+            write!(writer, "{}", trial.id)?;
+
+            // Value: empty for non-complete trials.
+            if trial.state == TrialState::Complete {
+                write!(writer, ",{}", trial.value)?;
+            } else {
+                write!(writer, ",")?;
+            }
+
+            write!(
+                writer,
+                ",{}",
+                match trial.state {
+                    TrialState::Complete => "Complete",
+                    TrialState::Pruned => "Pruned",
+                    TrialState::Failed => "Failed",
+                    TrialState::Running => "Running",
+                }
+            )?;
+
+            for id in &param_ids {
+                if let Some(pv) = trial.params.get(id) {
+                    write!(writer, ",{pv}")?;
+                } else {
+                    write!(writer, ",")?;
+                }
+            }
+
+            for key in &attr_keys {
+                if let Some(attr) = trial.user_attrs.get(key) {
+                    write!(writer, ",{}", csv_escape(&format_attr(attr)))?;
+                } else {
+                    write!(writer, ",")?;
+                }
+            }
+
+            writeln!(writer)?;
+        }
+
+        Ok(())
+    }
+
+    /// Export completed trials to a CSV file at the given path.
+    ///
+    /// Convenience wrapper around [`to_csv`](Self::to_csv) that creates a
+    /// buffered file writer.
+    ///
+    /// # Errors
+    ///
+    /// Returns an I/O error if the file cannot be created or written.
+    pub fn export_csv(&self, path: impl AsRef<std::path::Path>) -> std::io::Result<()> {
+        let file = std::fs::File::create(path)?;
+        self.to_csv(std::io::BufWriter::new(file))
+    }
+
+    /// Return a human-readable summary of the study.
+    ///
+    /// The summary includes:
+    /// - Optimization direction and total trial count
+    /// - Breakdown by state (complete, pruned) when applicable
+    /// - Best trial value and parameters (if any completed trials exist)
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::parameter::{FloatParam, Parameter};
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let study: Study<f64> = Study::new(Direction::Minimize);
+    /// let x = FloatParam::new(0.0, 10.0).name("x");
+    ///
+    /// let mut trial = study.create_trial();
+    /// let _ = x.suggest(&mut trial).unwrap();
+    /// study.complete_trial(trial, 0.42);
+    ///
+    /// let summary = study.summary();
+    /// assert!(summary.contains("Minimize"));
+    /// assert!(summary.contains("0.42"));
+    /// ```
+    #[must_use]
+    pub fn summary(&self) -> String {
+        use fmt::Write;
+
+        let trials = self.storage.trials_arc().read();
+        let n_complete = trials
+            .iter()
+            .filter(|t| t.state == TrialState::Complete)
+            .count();
+        let n_pruned = trials
+            .iter()
+            .filter(|t| t.state == TrialState::Pruned)
+            .count();
+
+        let direction_str = match self.direction {
+            Direction::Minimize => "Minimize",
+            Direction::Maximize => "Maximize",
+        };
+
+        let mut s = format!("Study: {direction_str} | {n} trials", n = trials.len());
+        if n_pruned > 0 {
+            let _ = write!(s, " ({n_complete} complete, {n_pruned} pruned)");
+        }
+
+        drop(trials);
+
+        if let Ok(best) = self.best_trial() {
+            let _ = write!(s, "\nBest value: {} (trial #{})", best.value, best.id);
+            if !best.params.is_empty() {
+                s.push_str("\nBest parameters:");
+                let mut params: Vec<_> = best.params.iter().collect();
+                params.sort_by_key(|(id, _)| *id);
+                for (id, value) in params {
+                    let label = best.param_labels.get(id).map_or("?", String::as_str);
+                    let _ = write!(s, "\n  {label} = {value}");
+                }
+            }
+        }
+
+        s
+    }
+}
+
+impl<V> fmt::Display for Study<V>
+where
+    V: PartialOrd + Clone + fmt::Display,
+{
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        f.write_str(&self.summary())
+    }
+}
+
+#[cfg(feature = "serde")]
+impl<V: PartialOrd + Clone + serde::Serialize> Study<V> {
+    /// Export trials as a pretty-printed JSON array to a file.
+    ///
+    /// Each element in the array is a serialized [`CompletedTrial`](crate::sampler::CompletedTrial).
+    /// Requires the `serde` feature.
+    ///
+    /// # Errors
+    ///
+    /// Returns an I/O error if the file cannot be created or written.
+    pub fn export_json(&self, path: impl AsRef<std::path::Path>) -> std::io::Result<()> {
+        let file = std::fs::File::create(path)?;
+        let trials = self.trials();
+        serde_json::to_writer_pretty(file, &trials).map_err(std::io::Error::other)
+    }
+}
+
+impl Study<f64> {
+    /// Generate an HTML report with interactive Plotly.js charts.
+    ///
+    /// Create a self-contained HTML file that can be opened in any browser.
+    /// See [`generate_html_report`](crate::visualization::generate_html_report)
+    /// for details on the included charts.
+    ///
+    /// # Errors
+    ///
+    /// Returns an I/O error if the file cannot be created or written.
+    pub fn export_html(&self, path: impl AsRef<std::path::Path>) -> std::io::Result<()> {
+        crate::visualization::generate_html_report(self, path)
+    }
+}
+
+/// Escape a string for CSV output. If the value contains a comma, quote, or
+/// newline, wrap it in double-quotes and double any embedded quotes.
+fn csv_escape(s: &str) -> String {
+    if s.contains(',') || s.contains('"') || s.contains('\n') || s.contains('\r') {
+        format!("\"{}\"", s.replace('"', "\"\""))
+    } else {
+        s.to_string()
+    }
+}
+
+/// Format an `AttrValue` as a string for CSV cells.
+fn format_attr(attr: &crate::trial::AttrValue) -> String {
+    use crate::trial::AttrValue;
+    match attr {
+        AttrValue::Float(v) => v.to_string(),
+        AttrValue::Int(v) => v.to_string(),
+        AttrValue::String(v) => v.clone(),
+        AttrValue::Bool(v) => v.to_string(),
+    }
+}

src/study/iter.rs

@@ -0,0 +1,43 @@
+use crate::sampler::CompletedTrial;
+
+use super::Study;
+
+impl<V> Study<V>
+where
+    V: PartialOrd + Clone,
+{
+    /// Return an iterator over all completed trials.
+    ///
+    /// This clones the internal trial list, so it is suitable for
+    /// analysis and iteration but not for hot paths.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let study: Study<f64> = Study::new(Direction::Minimize);
+    /// let trial = study.create_trial();
+    /// study.complete_trial(trial, 1.0);
+    ///
+    /// for t in study.iter() {
+    ///     println!("Trial {} → {}", t.id, t.value);
+    /// }
+    /// ```
+    #[must_use]
+    pub fn iter(&self) -> std::vec::IntoIter<CompletedTrial<V>> {
+        self.trials().into_iter()
+    }
+}
+
+impl<V> IntoIterator for &Study<V>
+where
+    V: PartialOrd + Clone,
+{
+    type Item = CompletedTrial<V>;
+    type IntoIter = std::vec::IntoIter<CompletedTrial<V>>;
+
+    fn into_iter(self) -> Self::IntoIter {
+        self.iter()
+    }
+}

src/study/mod.rs

@@ -0,0 +1,772 @@
+//! Study implementation for managing optimization trials.
+
+use core::any::Any;
+use std::collections::{HashMap, VecDeque};
+use std::sync::Arc;
+
+use parking_lot::{Mutex, RwLock};
+
+use crate::param::ParamValue;
+use crate::parameter::ParamId;
+use crate::pruner::{NopPruner, Pruner};
+use crate::sampler::random::RandomSampler;
+use crate::sampler::{CompletedTrial, Sampler};
+use crate::trial::Trial;
+use crate::types::{Direction, TrialState};
+
+mod analysis;
+mod builder;
+mod export;
+mod iter;
+mod optimize;
+mod persistence;
+
+#[cfg(feature = "async")]
+mod async_impl;
+
+pub use builder::StudyBuilder;
+#[cfg(feature = "serde")]
+pub use persistence::StudySnapshot;
+
+/// A study manages the optimization process, tracking trials and their results.
+///
+/// The study is parameterized by the objective value type `V`, which defaults to `f64`.
+/// The only constraint on `V` is `PartialOrd`, allowing comparison of objective values
+/// to determine which trial is best.
+///
+/// When `V = f64`, the study passes trial history to the sampler for informed
+/// parameter suggestions (e.g., TPE sampler uses history to guide sampling).
+///
+/// # Examples
+///
+/// ```
+/// use optimizer::{Direction, Study};
+///
+/// // Create a study to minimize an objective function
+/// let study: Study<f64> = Study::new(Direction::Minimize);
+/// assert_eq!(study.direction(), Direction::Minimize);
+/// ```
+pub struct Study<V = f64>
+where
+    V: PartialOrd,
+{
+    /// The optimization direction.
+    pub(crate) direction: Direction,
+    /// The sampler used to generate parameter values.
+    pub(crate) sampler: Arc<dyn Sampler>,
+    /// The pruner used to decide whether to stop trials early.
+    pub(crate) pruner: Arc<dyn Pruner>,
+    /// Trial storage backend (default: [`MemoryStorage`](crate::storage::MemoryStorage)).
+    pub(crate) storage: Arc<dyn crate::storage::Storage<V>>,
+    /// Optional factory for creating sampler-aware trials.
+    /// Set automatically for `Study<f64>` so that `create_trial()` and all
+    /// optimization methods use the sampler without requiring `_with_sampler` suffixes.
+    pub(crate) trial_factory: Option<Arc<dyn Fn(u64) -> Trial + Send + Sync>>,
+    /// Queue of parameter configurations to evaluate next.
+    pub(crate) enqueued_params: Arc<Mutex<VecDeque<HashMap<ParamId, ParamValue>>>>,
+}
+
+impl<V> Study<V>
+where
+    V: PartialOrd,
+{
+    /// Create a new study with the given optimization direction.
+    ///
+    /// Uses the default `RandomSampler` for parameter sampling.
+    ///
+    /// # Arguments
+    ///
+    /// * `direction` - Whether to minimize or maximize the objective function.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let study: Study<f64> = Study::new(Direction::Minimize);
+    /// assert_eq!(study.direction(), Direction::Minimize);
+    /// ```
+    #[must_use]
+    pub fn new(direction: Direction) -> Self
+    where
+        V: Send + Sync + 'static,
+    {
+        Self::with_sampler(direction, RandomSampler::new())
+    }
+
+    /// Return a [`StudyBuilder`] for constructing a study with a fluent API.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::prelude::*;
+    ///
+    /// let study: Study<f64> = Study::builder()
+    ///     .minimize()
+    ///     .sampler(TpeSampler::new())
+    ///     .pruner(NopPruner)
+    ///     .build();
+    /// ```
+    #[must_use]
+    pub fn builder() -> StudyBuilder<V> {
+        StudyBuilder::new()
+    }
+
+    /// Create a study that minimizes the objective value.
+    ///
+    /// This is a shorthand for `Study::with_sampler(Direction::Minimize, sampler)`.
+    ///
+    /// # Arguments
+    ///
+    /// * `sampler` - The sampler to use for parameter sampling.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::Study;
+    /// use optimizer::sampler::tpe::TpeSampler;
+    ///
+    /// let study: Study<f64> = Study::minimize(TpeSampler::new());
+    /// assert_eq!(study.direction(), optimizer::Direction::Minimize);
+    /// ```
+    #[must_use]
+    pub fn minimize(sampler: impl Sampler + 'static) -> Self
+    where
+        V: Send + Sync + 'static,
+    {
+        Self::with_sampler(Direction::Minimize, sampler)
+    }
+
+    /// Create a study that maximizes the objective value.
+    ///
+    /// This is a shorthand for `Study::with_sampler(Direction::Maximize, sampler)`.
+    ///
+    /// # Arguments
+    ///
+    /// * `sampler` - The sampler to use for parameter sampling.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::Study;
+    /// use optimizer::sampler::tpe::TpeSampler;
+    ///
+    /// let study: Study<f64> = Study::maximize(TpeSampler::new());
+    /// assert_eq!(study.direction(), optimizer::Direction::Maximize);
+    /// ```
+    #[must_use]
+    pub fn maximize(sampler: impl Sampler + 'static) -> Self
+    where
+        V: Send + Sync + 'static,
+    {
+        Self::with_sampler(Direction::Maximize, sampler)
+    }
+
+    /// Create a new study with a custom sampler.
+    ///
+    /// # Arguments
+    ///
+    /// * `direction` - Whether to minimize or maximize the objective function.
+    /// * `sampler` - The sampler to use for parameter sampling.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::sampler::random::RandomSampler;
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let sampler = RandomSampler::with_seed(42);
+    /// let study: Study<f64> = Study::with_sampler(Direction::Maximize, sampler);
+    /// assert_eq!(study.direction(), Direction::Maximize);
+    /// ```
+    pub fn with_sampler(direction: Direction, sampler: impl Sampler + 'static) -> Self
+    where
+        V: Send + Sync + 'static,
+    {
+        Self::with_sampler_and_storage(
+            direction,
+            sampler,
+            crate::storage::MemoryStorage::<V>::new(),
+        )
+    }
+
+    /// Build a trial factory for sampler integration when `V = f64`.
+    pub(crate) fn make_trial_factory(
+        sampler: &Arc<dyn Sampler>,
+        storage: &Arc<dyn crate::storage::Storage<V>>,
+        pruner: &Arc<dyn Pruner>,
+    ) -> Option<Arc<dyn Fn(u64) -> Trial + Send + Sync>>
+    where
+        V: 'static,
+    {
+        // Try to downcast the storage's trial buffer to the f64 specialization.
+        // This succeeds only when V = f64, enabling automatic sampler integration.
+        let trials_arc = storage.trials_arc();
+        let any_ref: &dyn Any = trials_arc;
+        let f64_trials: Option<&Arc<RwLock<Vec<CompletedTrial<f64>>>>> = any_ref.downcast_ref();
+
+        f64_trials.map(|trials| {
+            let sampler = Arc::clone(sampler);
+            let trials = Arc::clone(trials);
+            let pruner = Arc::clone(pruner);
+            let factory: Arc<dyn Fn(u64) -> Trial + Send + Sync> = Arc::new(move |id| {
+                Trial::with_sampler(
+                    id,
+                    Arc::clone(&sampler),
+                    Arc::clone(&trials),
+                    Arc::clone(&pruner),
+                )
+            });
+            factory
+        })
+    }
+
+    /// Create a study with a custom sampler and storage backend.
+    ///
+    /// This is the most general constructor — all other constructors
+    /// delegate to this one. Use it when you need a non-default storage
+    /// backend (e.g., `JournalStorage`).
+    ///
+    /// # Arguments
+    ///
+    /// * `direction` - Whether to minimize or maximize the objective function.
+    /// * `sampler` - The sampler to use for parameter sampling.
+    /// * `storage` - The storage backend for completed trials.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::sampler::random::RandomSampler;
+    /// use optimizer::storage::MemoryStorage;
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let storage = MemoryStorage::<f64>::new();
+    /// let study = Study::with_sampler_and_storage(Direction::Minimize, RandomSampler::new(), storage);
+    /// ```
+    pub fn with_sampler_and_storage(
+        direction: Direction,
+        sampler: impl Sampler + 'static,
+        storage: impl crate::storage::Storage<V> + 'static,
+    ) -> Self
+    where
+        V: 'static,
+    {
+        let sampler: Arc<dyn Sampler> = Arc::new(sampler);
+        let pruner: Arc<dyn Pruner> = Arc::new(NopPruner);
+        let storage: Arc<dyn crate::storage::Storage<V>> = Arc::new(storage);
+        let trial_factory = Self::make_trial_factory(&sampler, &storage, &pruner);
+
+        Self {
+            direction,
+            sampler,
+            pruner,
+            storage,
+            trial_factory,
+            enqueued_params: Arc::new(Mutex::new(VecDeque::new())),
+        }
+    }
+
+    /// Return the optimization direction.
+    #[must_use]
+    pub fn direction(&self) -> Direction {
+        self.direction
+    }
+
+    /// Creates a study with a custom sampler and pruner.
+    ///
+    /// Uses the default [`MemoryStorage`](crate::storage::MemoryStorage) backend.
+    ///
+    /// # Arguments
+    ///
+    /// * `direction` - Whether to minimize or maximize the objective function.
+    /// * `sampler` - The sampler to use for parameter sampling.
+    /// * `pruner` - The pruner to use for trial pruning.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::pruner::NopPruner;
+    /// use optimizer::sampler::random::RandomSampler;
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let sampler = RandomSampler::with_seed(42);
+    /// let study: Study<f64> = Study::with_sampler_and_pruner(Direction::Minimize, sampler, NopPruner);
+    /// ```
+    pub fn with_sampler_and_pruner(
+        direction: Direction,
+        sampler: impl Sampler + 'static,
+        pruner: impl Pruner + 'static,
+    ) -> Self
+    where
+        V: Send + Sync + 'static,
+    {
+        let sampler: Arc<dyn Sampler> = Arc::new(sampler);
+        let pruner: Arc<dyn Pruner> = Arc::new(pruner);
+        let storage: Arc<dyn crate::storage::Storage<V>> =
+            Arc::new(crate::storage::MemoryStorage::<V>::new());
+        let trial_factory = Self::make_trial_factory(&sampler, &storage, &pruner);
+
+        Self {
+            direction,
+            sampler,
+            pruner,
+            storage,
+            trial_factory,
+            enqueued_params: Arc::new(Mutex::new(VecDeque::new())),
+        }
+    }
+
+    /// Replace the sampler used for future parameter suggestions.
+    ///
+    /// The new sampler takes effect for all subsequent calls to
+    /// [`create_trial`](Self::create_trial), [`ask`](Self::ask), and the
+    /// `optimize*` family. Already-completed trials are unaffected.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::sampler::tpe::TpeSampler;
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let mut study: Study<f64> = Study::new(Direction::Minimize);
+    /// study.set_sampler(TpeSampler::new());
+    /// ```
+    pub fn set_sampler(&mut self, sampler: impl Sampler + 'static)
+    where
+        V: 'static,
+    {
+        self.sampler = Arc::new(sampler);
+        self.trial_factory = Self::make_trial_factory(&self.sampler, &self.storage, &self.pruner);
+    }
+
+    /// Replace the pruner used for future trials.
+    ///
+    /// The new pruner takes effect for all trials created after this call.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::prelude::*;
+    ///
+    /// let mut study: Study<f64> = Study::new(Direction::Minimize);
+    /// study.set_pruner(MedianPruner::new(Direction::Minimize));
+    /// ```
+    pub fn set_pruner(&mut self, pruner: impl Pruner + 'static)
+    where
+        V: 'static,
+    {
+        self.pruner = Arc::new(pruner);
+        self.trial_factory = Self::make_trial_factory(&self.sampler, &self.storage, &self.pruner);
+    }
+
+    /// Return a reference to the study's current pruner.
+    #[must_use]
+    pub fn pruner(&self) -> &dyn Pruner {
+        &*self.pruner
+    }
+
+    /// Enqueue a specific parameter configuration to be evaluated next.
+    ///
+    /// The next call to [`ask()`](Self::ask) or the next trial in [`optimize()`](Self::optimize)
+    /// will use these exact parameters instead of sampling from the sampler.
+    ///
+    /// Multiple configurations can be enqueued; they are evaluated in FIFO order.
+    /// If an enqueued configuration is missing a parameter that the objective calls
+    /// `suggest()` on, that parameter falls back to normal sampling.
+    ///
+    /// # Arguments
+    ///
+    /// * `params` - A map from parameter IDs to the values to use.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use std::collections::HashMap;
+    ///
+    /// use optimizer::parameter::{FloatParam, IntParam, ParamValue, Parameter};
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let study: Study<f64> = Study::new(Direction::Minimize);
+    /// let x = FloatParam::new(0.0, 10.0);
+    /// let y = IntParam::new(1, 100);
+    ///
+    /// // Evaluate these specific configurations first
+    /// study.enqueue(HashMap::from([
+    ///     (x.id(), ParamValue::Float(0.001)),
+    ///     (y.id(), ParamValue::Int(3)),
+    /// ]));
+    ///
+    /// // Next trial will use x=0.001, y=3
+    /// let mut trial = study.ask();
+    /// assert_eq!(x.suggest(&mut trial).unwrap(), 0.001);
+    /// assert_eq!(y.suggest(&mut trial).unwrap(), 3);
+    /// ```
+    pub fn enqueue(&self, params: HashMap<ParamId, ParamValue>) {
+        self.enqueued_params.lock().push_back(params);
+    }
+
+    /// Return the trial ID of the current best trial from the given slice.
+    #[cfg(feature = "tracing")]
+    pub(crate) fn best_id(&self, trials: &[CompletedTrial<V>]) -> Option<u64> {
+        let direction = self.direction;
+        trials
+            .iter()
+            .filter(|t| t.state == TrialState::Complete)
+            .max_by(|a, b| Self::compare_trials(a, b, direction))
+            .map(|t| t.id)
+    }
+
+    /// Return the number of enqueued parameter configurations.
+    ///
+    /// See [`enqueue`](Self::enqueue) for how to add configurations.
+    #[must_use]
+    pub fn n_enqueued(&self) -> usize {
+        self.enqueued_params.lock().len()
+    }
+
+    /// Generate the next unique trial ID.
+    pub(crate) fn next_trial_id(&self) -> u64 {
+        self.storage.next_trial_id()
+    }
+
+    /// Create a new trial with a unique ID.
+    ///
+    /// The trial starts in the `Running` state and can be used to suggest
+    /// parameter values. After the objective function is evaluated, call
+    /// `complete_trial` or `fail_trial` to record the result.
+    ///
+    /// For `Study<f64>`, this method automatically integrates with the study's
+    /// sampler and trial history, so there is no need to call a separate
+    /// `create_trial_with_sampler()` method.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let study: Study<f64> = Study::new(Direction::Minimize);
+    /// let trial = study.create_trial();
+    /// assert_eq!(trial.id(), 0);
+    ///
+    /// let trial2 = study.create_trial();
+    /// assert_eq!(trial2.id(), 1);
+    /// ```
+    #[must_use]
+    pub fn create_trial(&self) -> Trial {
+        self.storage.refresh();
+
+        let id = self.next_trial_id();
+        let mut trial = if let Some(factory) = &self.trial_factory {
+            factory(id)
+        } else {
+            Trial::new(id)
+        };
+
+        // If there are enqueued params, inject them into this trial
+        if let Some(fixed_params) = self.enqueued_params.lock().pop_front() {
+            trial.set_fixed_params(fixed_params);
+        }
+
+        trial
+    }
+
+    /// Record a completed trial with its objective value.
+    ///
+    /// This method stores the trial's parameters, distributions, and objective
+    /// value in the study's history. The stored data is used by samplers to
+    /// inform future parameter suggestions.
+    ///
+    /// # Arguments
+    ///
+    /// * `trial` - The trial that was evaluated.
+    /// * `value` - The objective value returned by the objective function.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::parameter::{FloatParam, Parameter};
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let study: Study<f64> = Study::new(Direction::Minimize);
+    /// let x_param = FloatParam::new(0.0, 1.0);
+    /// let mut trial = study.create_trial();
+    /// let x = x_param.suggest(&mut trial).unwrap();
+    /// let objective_value = x * x;
+    /// study.complete_trial(trial, objective_value);
+    ///
+    /// assert_eq!(study.n_trials(), 1);
+    /// ```
+    pub fn complete_trial(&self, trial: Trial, value: V) {
+        let completed = trial.into_completed(value, TrialState::Complete);
+        self.storage.push(completed);
+    }
+
+    /// Record a failed trial with an error message.
+    ///
+    /// Failed trials are not stored in the study's history and do not
+    /// contribute to future sampling decisions. This method is useful
+    /// when the objective function raises an error that should not stop
+    /// the optimization process.
+    ///
+    /// # Arguments
+    ///
+    /// * `trial` - The trial that failed.
+    /// * `_error` - An error message describing why the trial failed.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let study: Study<f64> = Study::new(Direction::Minimize);
+    /// let trial = study.create_trial();
+    /// study.fail_trial(trial, "objective function raised an exception");
+    ///
+    /// // Failed trials are not counted
+    /// assert_eq!(study.n_trials(), 0);
+    /// ```
+    pub fn fail_trial(&self, mut trial: Trial, _error: impl ToString) {
+        trial.set_failed();
+        // Failed trials are not stored in completed_trials
+        // They could be stored in a separate list for debugging if needed
+    }
+
+    /// Request a new trial with suggested parameters.
+    ///
+    /// This is the first half of the ask-and-tell interface. After calling
+    /// `ask()`, use parameter types to suggest values on the returned trial,
+    /// evaluate your objective externally, then pass the trial back to
+    /// [`tell()`](Self::tell) with the result.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::parameter::{FloatParam, Parameter};
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let study: Study<f64> = Study::new(Direction::Minimize);
+    /// let x = FloatParam::new(0.0, 10.0);
+    ///
+    /// let mut trial = study.ask();
+    /// let x_val = x.suggest(&mut trial).unwrap();
+    /// let value = x_val * x_val;
+    /// study.tell(trial, Ok::<_, &str>(value));
+    /// ```
+    #[must_use]
+    pub fn ask(&self) -> Trial {
+        self.create_trial()
+    }
+
+    /// Report the result of a trial obtained from [`ask()`](Self::ask).
+    ///
+    /// Pass `Ok(value)` for a successful evaluation or `Err(reason)` for a
+    /// failure. Failed trials are not stored in the study's history.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let study: Study<f64> = Study::new(Direction::Minimize);
+    ///
+    /// let trial = study.ask();
+    /// study.tell(trial, Ok::<_, &str>(42.0));
+    /// assert_eq!(study.n_trials(), 1);
+    ///
+    /// let trial = study.ask();
+    /// study.tell(trial, Err::<f64, _>("evaluation failed"));
+    /// assert_eq!(study.n_trials(), 1); // failed trials not counted
+    /// ```
+    pub fn tell(&self, trial: Trial, value: core::result::Result<V, impl ToString>) {
+        match value {
+            Ok(v) => self.complete_trial(trial, v),
+            Err(e) => self.fail_trial(trial, e),
+        }
+    }
+
+    /// Record a pruned trial, preserving its intermediate values.
+    ///
+    /// Pruned trials are stored alongside completed trials so that samplers
+    /// can optionally learn from partial evaluations. The trial's state is
+    /// set to [`Pruned`](crate::TrialState::Pruned).
+    ///
+    /// In practice you rarely call this directly — returning
+    /// `Err(TrialPruned)` from an objective function handles pruning
+    /// automatically.
+    ///
+    /// # Arguments
+    ///
+    /// * `trial` - The trial that was pruned.
+    pub fn prune_trial(&self, trial: Trial)
+    where
+        V: Default,
+    {
+        let completed = trial.into_completed(V::default(), TrialState::Pruned);
+        self.storage.push(completed);
+    }
+
+    /// Return all completed trials as a `Vec`.
+    ///
+    /// The returned vector contains clones of `CompletedTrial` values, which contain
+    /// the trial's parameters, distributions, and objective value.
+    ///
+    /// Note: This method acquires a read lock on the completed trials, so the
+    /// returned vector is a clone of the internal storage.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::parameter::{FloatParam, Parameter};
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let study: Study<f64> = Study::new(Direction::Minimize);
+    /// let x_param = FloatParam::new(0.0, 1.0);
+    /// let mut trial = study.create_trial();
+    /// let _ = x_param.suggest(&mut trial);
+    /// study.complete_trial(trial, 0.5);
+    ///
+    /// for completed in study.trials() {
+    ///     println!("Trial {} has value {:?}", completed.id, completed.value);
+    /// }
+    /// ```
+    #[must_use]
+    pub fn trials(&self) -> Vec<CompletedTrial<V>>
+    where
+        V: Clone,
+    {
+        self.storage.trials_arc().read().clone()
+    }
+
+    /// Return the number of completed trials.
+    ///
+    /// Pruned and failed trials are not counted. Use
+    /// [`n_pruned_trials()`](Self::n_pruned_trials) for the pruned count.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::parameter::{FloatParam, Parameter};
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let study: Study<f64> = Study::new(Direction::Minimize);
+    /// assert_eq!(study.n_trials(), 0);
+    ///
+    /// let x_param = FloatParam::new(0.0, 1.0);
+    /// let mut trial = study.create_trial();
+    /// let _ = x_param.suggest(&mut trial);
+    /// study.complete_trial(trial, 0.5);
+    /// assert_eq!(study.n_trials(), 1);
+    /// ```
+    #[must_use]
+    pub fn n_trials(&self) -> usize {
+        self.storage
+            .trials_arc()
+            .read()
+            .iter()
+            .filter(|t| t.state == TrialState::Complete)
+            .count()
+    }
+
+    /// Return the number of pruned trials.
+    ///
+    /// Pruned trials are those that were stopped early by the pruner.
+    #[must_use]
+    pub fn n_pruned_trials(&self) -> usize {
+        self.storage
+            .trials_arc()
+            .read()
+            .iter()
+            .filter(|t| t.state == TrialState::Pruned)
+            .count()
+    }
+
+    /// Compare two completed trials using constraint-aware ranking.
+    ///
+    /// 1. Feasible trials always rank above infeasible trials.
+    /// 2. Among feasible trials, rank by objective value (respecting direction).
+    /// 3. Among infeasible trials, rank by total constraint violation (lower is better).
+    pub(crate) fn compare_trials(
+        a: &CompletedTrial<V>,
+        b: &CompletedTrial<V>,
+        direction: Direction,
+    ) -> core::cmp::Ordering {
+        match (a.is_feasible(), b.is_feasible()) {
+            (true, false) => core::cmp::Ordering::Greater,
+            (false, true) => core::cmp::Ordering::Less,
+            (false, false) => {
+                let va: f64 = a.constraints.iter().map(|c| c.max(0.0)).sum();
+                let vb: f64 = b.constraints.iter().map(|c| c.max(0.0)).sum();
+                vb.partial_cmp(&va).unwrap_or(core::cmp::Ordering::Equal)
+            }
+            (true, true) => {
+                let ordering = a.value.partial_cmp(&b.value);
+                match direction {
+                    Direction::Minimize => {
+                        ordering.map_or(core::cmp::Ordering::Equal, core::cmp::Ordering::reverse)
+                    }
+                    Direction::Maximize => ordering.unwrap_or(core::cmp::Ordering::Equal),
+                }
+            }
+        }
+    }
+}
+
+impl<V: PartialOrd + Send + Sync + 'static> Study<V> {
+    /// Create a study with a custom sampler, pruner, and storage backend.
+    ///
+    /// The most flexible constructor, allowing full control over all components.
+    ///
+    /// # Arguments
+    ///
+    /// * `direction` - Whether to minimize or maximize the objective function.
+    /// * `sampler` - The sampler to use for parameter sampling.
+    /// * `pruner` - The pruner to use for trial pruning.
+    /// * `storage` - The storage backend for completed trials.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::prelude::*;
+    /// use optimizer::storage::MemoryStorage;
+    ///
+    /// let study = Study::with_sampler_pruner_and_storage(
+    ///     Direction::Minimize,
+    ///     TpeSampler::new(),
+    ///     MedianPruner::new(Direction::Minimize),
+    ///     MemoryStorage::<f64>::new(),
+    /// );
+    /// ```
+    pub fn with_sampler_pruner_and_storage(
+        direction: Direction,
+        sampler: impl Sampler + 'static,
+        pruner: impl Pruner + 'static,
+        storage: impl crate::storage::Storage<V> + 'static,
+    ) -> Self {
+        let sampler: Arc<dyn Sampler> = Arc::new(sampler);
+        let pruner: Arc<dyn Pruner> = Arc::new(pruner);
+        let storage: Arc<dyn crate::storage::Storage<V>> = Arc::new(storage);
+        let trial_factory = Self::make_trial_factory(&sampler, &storage, &pruner);
+
+        Self {
+            direction,
+            sampler,
+            pruner,
+            storage,
+            trial_factory,
+            enqueued_params: Arc::new(Mutex::new(VecDeque::new())),
+        }
+    }
+}
+
+/// Returns `true` if the error represents a pruned trial.
+///
+/// Checks via `Any` downcasting whether `e` is `Error::TrialPruned` or
+/// the standalone `TrialPruned` struct.
+pub(super) fn is_trial_pruned<E: 'static>(e: &E) -> bool {
+    let any: &dyn Any = e;
+    if let Some(err) = any.downcast_ref::<crate::Error>() {
+        matches!(err, crate::Error::TrialPruned)
+    } else {
+        any.downcast_ref::<crate::error::TrialPruned>().is_some()
+    }
+}

src/study/optimize.rs

@@ -0,0 +1,123 @@
+use core::ops::ControlFlow;
+
+use crate::types::TrialState;
+
+use super::{Study, is_trial_pruned};
+
+impl<V> Study<V>
+where
+    V: PartialOrd,
+{
+    /// Run optimization with an objective.
+    ///
+    /// Accepts any [`Objective`](crate::Objective) implementation, including
+    /// plain closures (`Fn(&mut Trial) -> Result<V, E>`) thanks to the
+    /// blanket impl. Struct-based objectives can override
+    /// [`before_trial`](crate::Objective::before_trial) and
+    /// [`after_trial`](crate::Objective::after_trial) for early stopping.
+    ///
+    /// Runs up to `n_trials` evaluations sequentially.
+    ///
+    /// # Errors
+    ///
+    /// Returns `Error::NoCompletedTrials` if no trials completed successfully.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::parameter::{FloatParam, Parameter};
+    /// use optimizer::sampler::random::RandomSampler;
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let sampler = RandomSampler::with_seed(42);
+    /// let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+    /// let x_param = FloatParam::new(-10.0, 10.0);
+    ///
+    /// study
+    ///     .optimize(10, |trial: &mut optimizer::Trial| {
+    ///         let x = x_param.suggest(trial)?;
+    ///         Ok::<_, optimizer::Error>(x * x)
+    ///     })
+    ///     .unwrap();
+    ///
+    /// assert!(study.n_trials() > 0);
+    /// assert!(study.best_value().unwrap() >= 0.0);
+    /// ```
+    #[allow(clippy::needless_pass_by_value)]
+    pub fn optimize(
+        &self,
+        n_trials: usize,
+        objective: impl crate::objective::Objective<V>,
+    ) -> crate::Result<()>
+    where
+        V: Clone + Default,
+    {
+        #[cfg(feature = "tracing")]
+        let _span =
+            tracing::info_span!("optimize", n_trials, direction = ?self.direction).entered();
+
+        for _ in 0..n_trials {
+            if let ControlFlow::Break(()) = objective.before_trial(self) {
+                break;
+            }
+
+            let mut trial = self.create_trial();
+            match objective.evaluate(&mut trial) {
+                Ok(value) => {
+                    #[cfg(feature = "tracing")]
+                    let trial_id = trial.id();
+
+                    let completed = trial.into_completed(value, TrialState::Complete);
+
+                    // Fire after_trial hook before pushing to storage
+                    let flow = objective.after_trial(self, &completed);
+                    self.storage.push(completed);
+
+                    #[cfg(feature = "tracing")]
+                    {
+                        tracing::info!(trial_id, "trial completed");
+                        let trials = self.storage.trials_arc().read();
+                        if trials
+                            .iter()
+                            .filter(|t| t.state == TrialState::Complete)
+                            .count()
+                            == 1
+                            || trials.last().map(|t| t.id) == self.best_id(&trials)
+                        {
+                            tracing::info!(trial_id, "new best value found");
+                        }
+                    }
+
+                    if let ControlFlow::Break(()) = flow {
+                        return Ok(());
+                    }
+                }
+                Err(e) if is_trial_pruned(&e) => {
+                    #[cfg(feature = "tracing")]
+                    let trial_id = trial.id();
+                    self.prune_trial(trial);
+                    trace_info!(trial_id, "trial pruned");
+                }
+                Err(e) => {
+                    #[cfg(feature = "tracing")]
+                    let trial_id = trial.id();
+                    self.fail_trial(trial, e.to_string());
+                    trace_debug!(trial_id, "trial failed");
+                }
+            }
+        }
+
+        // Return error if no trials completed successfully
+        let has_complete = self
+            .storage
+            .trials_arc()
+            .read()
+            .iter()
+            .any(|t| t.state == TrialState::Complete);
+        if !has_complete {
+            return Err(crate::Error::NoCompletedTrials);
+        }
+
+        Ok(())
+    }
+}

src/study/persistence.rs

@@ -0,0 +1,138 @@
+#[cfg(feature = "serde")]
+use std::collections::HashMap;
+
+#[cfg(feature = "serde")]
+use crate::sampler::CompletedTrial;
+#[cfg(feature = "serde")]
+use crate::types::Direction;
+
+#[cfg(feature = "serde")]
+use super::Study;
+
+/// A serializable snapshot of a study's state.
+///
+/// Since [`Study`] contains non-serializable fields (samplers, atomics, etc.),
+/// this struct captures the essential state needed to save and restore a study.
+///
+/// # Schema versioning
+///
+/// The `version` field enables future schema evolution without breaking existing files.
+/// The current version is `1`.
+///
+/// # Sampler state
+///
+/// Sampler state is **not** included in the snapshot. After loading, the study
+/// uses a default `RandomSampler`. Call [`Study::set_sampler`] to restore
+/// the desired sampler configuration.
+#[cfg(feature = "serde")]
+#[derive(serde::Serialize, serde::Deserialize)]
+pub struct StudySnapshot<V> {
+    /// Schema version for forward compatibility.
+    pub version: u32,
+    /// The optimization direction.
+    pub direction: Direction,
+    /// All completed (and pruned) trials.
+    pub trials: Vec<CompletedTrial<V>>,
+    /// The next trial ID to assign.
+    pub next_trial_id: u64,
+    /// Optional metadata (creation timestamp, sampler description, etc.).
+    pub metadata: HashMap<String, String>,
+}
+
+#[cfg(feature = "serde")]
+impl<V: PartialOrd + Clone + serde::Serialize> Study<V> {
+    /// Save the study state to a JSON file.
+    ///
+    /// # Errors
+    ///
+    /// Returns an I/O error if the file cannot be created or written.
+    pub fn save(&self, path: impl AsRef<std::path::Path>) -> std::io::Result<()> {
+        let path = path.as_ref();
+        let trials = self.trials();
+        let next_trial_id = self.storage.peek_next_trial_id();
+        let snapshot = StudySnapshot {
+            version: 1,
+            direction: self.direction,
+            trials,
+            next_trial_id,
+            metadata: HashMap::new(),
+        };
+
+        // Atomic write: write to a temp file in the same directory, then rename.
+        // This prevents corrupt files if the process crashes mid-write.
+        let parent = path.parent().unwrap_or(std::path::Path::new("."));
+        let tmp_path = parent.join(format!(
+            ".{}.tmp",
+            path.file_name().unwrap_or_default().to_string_lossy()
+        ));
+        let file = std::fs::File::create(&tmp_path)?;
+        serde_json::to_writer_pretty(file, &snapshot).map_err(std::io::Error::other)?;
+        std::fs::rename(&tmp_path, path)
+    }
+}
+
+#[cfg(feature = "serde")]
+impl<V: PartialOrd + Send + Sync + Clone + serde::de::DeserializeOwned + 'static> Study<V> {
+    /// Load a study from a JSON file.
+    ///
+    /// The loaded study uses a `RandomSampler` by default. Call
+    /// [`set_sampler()`](Self::set_sampler) to restore the original sampler
+    /// configuration.
+    ///
+    /// # Errors
+    ///
+    /// Returns an I/O error if the file cannot be read or parsed.
+    pub fn load(path: impl AsRef<std::path::Path>) -> std::io::Result<Self> {
+        use crate::sampler::random::RandomSampler;
+
+        let file = std::fs::File::open(path)?;
+        let snapshot: StudySnapshot<V> = serde_json::from_reader(file)
+            .map_err(|e| std::io::Error::new(std::io::ErrorKind::InvalidData, e))?;
+        let storage = crate::storage::MemoryStorage::with_trials(snapshot.trials);
+        Ok(Self::with_sampler_and_storage(
+            snapshot.direction,
+            RandomSampler::new(),
+            storage,
+        ))
+    }
+}
+
+#[cfg(feature = "journal")]
+impl<V> Study<V>
+where
+    V: PartialOrd + Send + Sync + serde::Serialize + serde::de::DeserializeOwned + 'static,
+{
+    /// Create a study backed by a JSONL journal file.
+    ///
+    /// Any existing trials in the file are loaded into memory and the
+    /// trial ID counter is set to one past the highest stored ID. New
+    /// trials are written through to the file on completion.
+    ///
+    /// # Arguments
+    ///
+    /// * `direction` - Whether to minimize or maximize the objective function.
+    /// * `sampler` - The sampler to use for parameter sampling.
+    /// * `path` - Path to the JSONL journal file (created if absent).
+    ///
+    /// # Errors
+    ///
+    /// Returns a [`Storage`](crate::Error::Storage) error if loading fails.
+    ///
+    /// # Examples
+    ///
+    /// ```no_run
+    /// use optimizer::sampler::tpe::TpeSampler;
+    /// use optimizer::{Direction, Study};
+    ///
+    /// let study: Study<f64> =
+    ///     Study::with_journal(Direction::Minimize, TpeSampler::new(), "trials.jsonl").unwrap();
+    /// ```
+    pub fn with_journal(
+        direction: Direction,
+        sampler: impl crate::sampler::Sampler + 'static,
+        path: impl AsRef<std::path::Path>,
+    ) -> crate::Result<Self> {
+        let storage = crate::storage::JournalStorage::<V>::open(path)?;
+        Ok(Self::with_sampler_and_storage(direction, sampler, storage))
+    }
+}

src/trial.rs

@@ -1,4 +1,23 @@
-//! Trial implementation for tracking sampled parameters and trial state.
+//! Trial lifecycle management for optimization runs.
+//!
+//! A [`Trial`] represents a single evaluation of the objective function. The study
+//! creates trials, the objective function samples parameters from them via
+//! [`Parameter::suggest`](crate::parameter::Parameter::suggest), and reports
+//! intermediate values for pruning decisions.
+//!
+//! # Lifecycle
+//!
+//! 1. **Created** — `Study` creates a trial with [`Trial::new`] or internally via
+//!    `Trial::with_sampler`.
+//! 2. **Running** — The objective calls [`Trial::suggest_param`] to sample parameters
+//!    and optionally [`Trial::report`] / [`Trial::should_prune`] for early stopping.
+//! 3. **Completed / Failed / Pruned** — The study marks the trial's final state.
+//!
+//! # User Attributes
+//!
+//! Trials support arbitrary key-value metadata via [`Trial::set_user_attr`] and
+//! [`Trial::user_attr`], useful for logging hyperparameters, hardware info, or
+//! debug notes alongside the optimization results.
 
 use std::collections::HashMap;
 use std::sync::Arc;
@@ -7,6 +26,7 @@ use parking_lot::RwLock;
 
 use crate::distribution::Distribution;
 use crate::error::{Error, Result};
+use crate::multi_objective::MultiObjectiveTrial;
 use crate::param::ParamValue;
 use crate::parameter::{ParamId, Parameter};
 use crate::pruner::Pruner;
@@ -57,14 +77,26 @@ impl From<bool> for AttrValue {
     }
 }
 
-/// A trial represents a single evaluation of the objective function.
+/// A single evaluation of the objective function.
 ///
 /// Each trial has a unique ID and stores the sampled parameters along with
-/// their distributions. The trial progresses through states: Running -> Complete/Failed.
+/// their distributions. The trial progresses through states:
+/// `Running` → `Complete` / `Failed` / `Pruned`.
 ///
-/// Trials use a sampler to generate parameter values. When created through
-/// `Study::create_trial()`, the trial receives the study's sampler and access
-/// to the history of completed trials for informed sampling.
+/// Trials use a [`Sampler`](crate::sampler::Sampler) to generate parameter
+/// values. When created through [`Study::create_trial`](crate::Study::create_trial),
+/// the trial receives the study's sampler and access to the history of
+/// completed trials for informed sampling.
+///
+/// # Examples
+///
+/// ```
+/// use optimizer::Trial;
+/// use optimizer::parameter::{FloatParam, Parameter};
+///
+/// let mut trial = Trial::new(0);
+/// let x = FloatParam::new(-5.0, 5.0).suggest(&mut trial).unwrap();
+/// ```
 #[derive(Clone)]
 pub struct Trial {
     /// Unique identifier for this trial.
@@ -113,13 +145,14 @@ impl core::fmt::Debug for Trial {
 }
 
 impl Trial {
-    /// Creates a new trial with the given ID.
+    /// Create a new trial with the given ID.
     ///
     /// The trial starts in the `Running` state with no parameters sampled.
-    /// This constructor creates a trial without a sampler, which will use
-    /// local random sampling for suggest methods.
+    /// This constructor creates a trial without a sampler, which will fall
+    /// back to random sampling for [`suggest_param`](Self::suggest_param) calls.
     ///
-    /// For trials that use the study's sampler, use `Trial::with_sampler` instead.
+    /// For trials that use the study's sampler, the study creates them
+    /// internally via `Trial::with_sampler`.
     ///
     /// # Arguments
     ///
@@ -151,10 +184,10 @@ impl Trial {
         }
     }
 
-    /// Creates a new trial with a sampler and access to trial history.
+    /// Create a new trial with a sampler and access to trial history.
     ///
-    /// This constructor is used by `Study::create_trial()` to create trials
-    /// that use the study's sampler for informed parameter suggestions.
+    /// Used internally by `Study::create_trial()` to create trials that use
+    /// the study's sampler for informed parameter suggestions.
     ///
     /// # Arguments
     ///
@@ -183,19 +216,19 @@ impl Trial {
         }
     }
 
-    /// Sets pre-filled parameters on this trial.
+    /// Set pre-filled parameters on this trial.
     ///
-    /// When `suggest_param` is called for a parameter that has a fixed value,
-    /// the fixed value is used instead of sampling.
+    /// When [`suggest_param`](Self::suggest_param) is called for a parameter
+    /// that has a fixed value, the fixed value is used instead of sampling.
     pub(crate) fn set_fixed_params(&mut self, params: HashMap<ParamId, ParamValue>) {
         self.fixed_params = params;
     }
 
-    /// Samples a value from the given distribution using the sampler.
+    /// Sample a value from the given distribution using the sampler.
     ///
-    /// If the trial has a sampler, it delegates to the sampler's sample method
-    /// with the history of completed trials. Otherwise, it uses the `RandomSampler`
-    /// as a fallback.
+    /// If the trial has a sampler, delegates to the sampler's sample method
+    /// with the history of completed trials. Otherwise, falls back to
+    /// [`RandomSampler`](crate::sampler::random::RandomSampler).
     fn sample_value(&self, distribution: &Distribution) -> ParamValue {
         if let (Some(sampler), Some(history)) = (&self.sampler, &self.history) {
             let history_guard = history.read();
@@ -208,40 +241,55 @@ impl Trial {
         }
     }
 
-    /// Returns the unique ID of this trial.
+    /// Return the unique ID of this trial.
     #[must_use]
     pub fn id(&self) -> u64 {
         self.id
     }
 
-    /// Returns the current state of this trial.
+    /// Return the current state of this trial.
     #[must_use]
     pub fn state(&self) -> TrialState {
         self.state
     }
 
-    /// Returns a reference to the sampled parameters.
+    /// Return a reference to the sampled parameters, keyed by [`ParamId`](crate::parameter::ParamId).
     #[must_use]
     pub fn params(&self) -> &HashMap<ParamId, ParamValue> {
         &self.params
     }
 
-    /// Returns a reference to the parameter distributions.
+    /// Return a reference to the parameter distributions, keyed by [`ParamId`](crate::parameter::ParamId).
     #[must_use]
     pub fn distributions(&self) -> &HashMap<ParamId, Distribution> {
         &self.distributions
     }
 
-    /// Returns a reference to the parameter labels.
+    /// Return a reference to the parameter labels, keyed by [`ParamId`](crate::parameter::ParamId).
     #[must_use]
     pub fn param_labels(&self) -> &HashMap<ParamId, String> {
         &self.param_labels
     }
 
-    /// Reports an intermediate objective value at a given step.
+    /// Report an intermediate objective value at a given step.
     ///
-    /// Steps should be monotonically increasing (e.g., epoch number).
-    /// Duplicate steps overwrite the previous value.
+    /// Call this during iterative training (e.g., once per epoch) so the
+    /// [`Pruner`](crate::pruner::Pruner) can decide whether to stop the trial
+    /// early. Steps should be monotonically increasing; duplicate steps
+    /// overwrite the previous value.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::Trial;
+    ///
+    /// let mut trial = Trial::new(0);
+    /// for epoch in 0..10 {
+    ///     let loss = 1.0 / (epoch as f64 + 1.0);
+    ///     trial.report(epoch, loss);
+    /// }
+    /// assert_eq!(trial.intermediate_values().len(), 10);
+    /// ```
     pub fn report(&mut self, step: u64, value: f64) {
         if let Some(entry) = self
             .intermediate_values
@@ -256,8 +304,11 @@ impl Trial {
 
     /// Ask whether this trial should be pruned at the current step.
     ///
-    /// Returns `true` if the pruner recommends stopping this trial.
-    /// The caller should return `Err(TrialPruned)` from the objective.
+    /// Return `true` if the pruner recommends stopping this trial based on
+    /// the intermediate values reported so far. When `true`, the objective
+    /// should return early with `Err(TrialPruned)?`.
+    ///
+    /// Always returns `false` when no pruner is configured.
     #[must_use]
     pub fn should_prune(&self) -> bool {
         let (Some(pruner), Some(history)) = (&self.pruner, &self.history) else {
@@ -274,59 +325,77 @@ impl Trial {
         prune
     }
 
-    /// Returns all intermediate values reported so far.
+    /// Return all intermediate values reported so far as `(step, value)` pairs.
     #[must_use]
     pub fn intermediate_values(&self) -> &[(u64, f64)] {
         &self.intermediate_values
     }
 
-    /// Sets a user attribute on this trial.
+    /// Set a user attribute on this trial.
+    ///
+    /// User attributes are arbitrary key-value pairs for logging, debugging,
+    /// or analysis. Values can be `f64`, `i64`, `String`, `&str`, or `bool`
+    /// (anything implementing `Into<AttrValue>`).
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::Trial;
+    ///
+    /// let mut trial = Trial::new(0);
+    /// trial.set_user_attr("gpu", "A100");
+    /// trial.set_user_attr("batch_size", 64_i64);
+    /// trial.set_user_attr("accuracy", 0.95);
+    /// ```
     pub fn set_user_attr(&mut self, key: impl Into<String>, value: impl Into<AttrValue>) {
         self.user_attrs.insert(key.into(), value.into());
     }
 
-    /// Gets a user attribute by key.
+    /// Return a user attribute by key, or `None` if it does not exist.
     #[must_use]
     pub fn user_attr(&self, key: &str) -> Option<&AttrValue> {
         self.user_attrs.get(key)
     }
 
-    /// Returns all user attributes.
+    /// Return all user attributes as a map.
     #[must_use]
     pub fn user_attrs(&self) -> &HashMap<String, AttrValue> {
         &self.user_attrs
     }
 
-    /// Sets constraint values for this trial.
+    /// Set constraint values for this trial.
     ///
-    /// Each value represents a constraint; a value <= 0.0 means the constraint
-    /// is satisfied (feasible). A value > 0.0 means the constraint is violated.
+    /// Each element represents one constraint. A value ≤ 0.0 means the
+    /// constraint is satisfied (feasible); a value > 0.0 means violated.
+    /// Constrained samplers (e.g., NSGA-II with constraints) use these values
+    /// to prefer feasible solutions.
+    ///
+    /// # Examples
+    ///
+    /// ```
+    /// use optimizer::Trial;
+    ///
+    /// let mut trial = Trial::new(0);
+    /// // Two constraints: first satisfied, second violated
+    /// trial.set_constraints(vec![-0.5, 0.3]);
+    /// assert_eq!(trial.constraint_values(), &[-0.5, 0.3]);
+    /// ```
     pub fn set_constraints(&mut self, values: Vec<f64>) {
         self.constraint_values = values;
     }
 
-    /// Returns the constraint values for this trial.
+    /// Return the constraint values for this trial.
     #[must_use]
     pub fn constraint_values(&self) -> &[f64] {
         &self.constraint_values
     }
 
-    /// Sets the trial state to Complete.
-    pub(crate) fn set_complete(&mut self) {
-        self.state = TrialState::Complete;
-    }
-
-    /// Sets the trial state to Failed.
+    /// Set the trial state to `Failed`.
     pub(crate) fn set_failed(&mut self) {
         self.state = TrialState::Failed;
     }
 
-    /// Sets the trial state to Pruned.
-    pub(crate) fn set_pruned(&mut self) {
-        self.state = TrialState::Pruned;
-    }
-
-    /// Suggests a parameter value using a [`Parameter`] definition.
+    /// Suggest a parameter value using a [`Parameter`] definition.
     ///
     /// This is the primary entry point for sampling parameters. It handles
     /// validation, caching, conflict detection, sampling, and conversion.
@@ -404,4 +473,201 @@ impl Trial {
 
         Ok(result)
     }
+
+    /// Consume this trial and move its fields into a [`CompletedTrial`].
+    ///
+    /// This avoids cloning the trial's `HashMap`s and `Vec`s by moving
+    /// ownership directly into the completed trial.
+    pub(crate) fn into_completed<V>(self, value: V, state: TrialState) -> CompletedTrial<V> {
+        CompletedTrial {
+            id: self.id,
+            params: self.params,
+            distributions: self.distributions,
+            param_labels: self.param_labels,
+            value,
+            intermediate_values: self.intermediate_values,
+            state,
+            user_attrs: self.user_attrs,
+            constraints: self.constraint_values,
+        }
+    }
+
+    /// Consume this trial and move its fields into a [`MultiObjectiveTrial`].
+    pub(crate) fn into_multi_objective_trial(
+        self,
+        values: Vec<f64>,
+        state: TrialState,
+    ) -> MultiObjectiveTrial {
+        MultiObjectiveTrial {
+            id: self.id,
+            params: self.params,
+            distributions: self.distributions,
+            param_labels: self.param_labels,
+            values,
+            state,
+            user_attrs: self.user_attrs,
+            constraints: self.constraint_values,
+        }
+    }
+}
+
+#[cfg(test)]
+#[allow(clippy::float_cmp)]
+mod tests {
+    use crate::parameter::{BoolParam, CategoricalParam, FloatParam, IntParam, Parameter};
+    use crate::types::TrialState;
+
+    #[test]
+    fn trial_state() {
+        // from test_trial_state (L643 of integration.rs)
+        let trial = super::Trial::new(0);
+        assert_eq!(trial.state(), TrialState::Running);
+    }
+
+    #[test]
+    fn trial_params_access() {
+        // from test_trial_params_access (L651)
+        let x_param = FloatParam::new(0.0, 1.0);
+        let n_param = IntParam::new(1, 10);
+        let mut trial = super::Trial::new(0);
+
+        x_param.suggest(&mut trial).unwrap();
+        n_param.suggest(&mut trial).unwrap();
+
+        let params = trial.params();
+        assert_eq!(params.len(), 2);
+    }
+
+    #[test]
+    fn trial_debug_format() {
+        // from test_trial_debug_format (L792)
+        let param = FloatParam::new(0.0, 1.0);
+        let mut trial = super::Trial::new(42);
+        param.suggest(&mut trial).unwrap();
+
+        let debug_str = format!("{trial:?}");
+
+        assert!(debug_str.contains("Trial"));
+        assert!(debug_str.contains("42"));
+        assert!(debug_str.contains("has_sampler"));
+    }
+
+    #[test]
+    fn distributions_access() {
+        // from test_distributions_access (L960)
+        let x_param = FloatParam::new(0.0, 1.0);
+        let n_param = IntParam::new(1, 10);
+        let opt_param = CategoricalParam::new(vec!["a", "b", "c"]);
+        let mut trial = super::Trial::new(0);
+
+        x_param.suggest(&mut trial).unwrap();
+        n_param.suggest(&mut trial).unwrap();
+        opt_param.suggest(&mut trial).unwrap();
+
+        let dists = trial.distributions();
+        assert_eq!(dists.len(), 3);
+    }
+
+    #[test]
+    fn multiple_parameters_independent_caching() {
+        // from test_multiple_parameters_independent_caching (L356)
+        let x_param = FloatParam::new(0.0, 1.0);
+        let y_param = FloatParam::new(0.0, 1.0);
+        let n_param = IntParam::new(1, 10);
+        let opt_param = CategoricalParam::new(vec!["a", "b"]);
+        let mut trial = super::Trial::new(0);
+
+        let x = x_param.suggest(&mut trial).unwrap();
+        let y = y_param.suggest(&mut trial).unwrap();
+        let n = n_param.suggest(&mut trial).unwrap();
+        let opt = opt_param.suggest(&mut trial).unwrap();
+
+        assert_eq!(x, x_param.suggest(&mut trial).unwrap());
+        assert_eq!(y, y_param.suggest(&mut trial).unwrap());
+        assert_eq!(n, n_param.suggest(&mut trial).unwrap());
+        assert_eq!(opt, opt_param.suggest(&mut trial).unwrap());
+    }
+
+    #[test]
+    fn suggest_bool_multiple_parameters() {
+        // from test_suggest_bool_multiple_parameters (L1131)
+        let dropout_param = BoolParam::new();
+        let batchnorm_param = BoolParam::new();
+        let skip_param = BoolParam::new();
+        let mut trial = super::Trial::new(0);
+
+        let a = dropout_param.suggest(&mut trial).unwrap();
+        let b = batchnorm_param.suggest(&mut trial).unwrap();
+        let c = skip_param.suggest(&mut trial).unwrap();
+
+        assert_eq!(a, dropout_param.suggest(&mut trial).unwrap());
+        assert_eq!(b, batchnorm_param.suggest(&mut trial).unwrap());
+        assert_eq!(c, skip_param.suggest(&mut trial).unwrap());
+    }
+
+    #[test]
+    fn param_name() {
+        // from test_param_name (L1312)
+        let param = FloatParam::new(0.0, 1.0).name("learning_rate");
+        let mut trial = super::Trial::new(0);
+        param.suggest(&mut trial).unwrap();
+
+        let labels = trial.param_labels();
+        let label = labels.values().next().unwrap();
+        assert_eq!(label, "learning_rate");
+    }
+
+    #[test]
+    fn step_float_snaps_to_grid() {
+        // from test_step_float_snaps_to_grid (L676)
+        let param = FloatParam::new(0.0, 1.0).step(0.25);
+        let mut trial = super::Trial::new(0);
+
+        let x = param.suggest(&mut trial).unwrap();
+
+        let valid_values = [0.0, 0.25, 0.5, 0.75, 1.0];
+        let is_valid = valid_values.iter().any(|&v| (x - v).abs() < 1e-10);
+        assert!(is_valid, "stepped float {x} should snap to grid");
+    }
+
+    #[test]
+    fn step_int_snaps_to_grid() {
+        // from test_step_int_snaps_to_grid (L689)
+        let param = IntParam::new(0, 100).step(25);
+        let mut trial = super::Trial::new(0);
+
+        let n = param.suggest(&mut trial).unwrap();
+
+        assert!(
+            n % 25 == 0 && (0..=100).contains(&n),
+            "stepped int {n} should snap to grid"
+        );
+    }
+
+    #[test]
+    fn int_bounds_with_low_equals_high() {
+        // from test_int_bounds_with_low_equals_high (L1086)
+        let mut trial = super::Trial::new(0);
+
+        let n_param = IntParam::new(5, 5);
+        let n = n_param.suggest(&mut trial).unwrap();
+        assert_eq!(n, 5);
+
+        let x_param = FloatParam::new(3.0, 3.0);
+        let x = x_param.suggest(&mut trial).unwrap();
+        assert_eq!(x, 3.0);
+    }
+
+    #[test]
+    fn single_value_float_range() {
+        // from test_single_value_float_range (L1296)
+        let param = FloatParam::new(4.2, 4.2);
+        let mut trial = super::Trial::new(0);
+
+        let x = param.suggest(&mut trial).unwrap();
+        assert!(
+            (x - 4.2).abs() < f64::EPSILON,
+            "single-value range should return that value"
+        );
+    }
 }

src/visualization.rs

@@ -0,0 +1,500 @@
+//! HTML report generation for optimization visualization.
+//!
+//! Generate self-contained HTML files with embedded
+//! [Plotly.js](https://plotly.com/javascript/) charts for offline
+//! visualization of optimization results. No feature flag is required —
+//! this module is always available.
+//!
+//! # Charts included
+//!
+//! | Chart | Description |
+//! |---|---|
+//! | **Optimization history** | Objective value vs trial number with best-so-far line |
+//! | **Slice plots** | Objective value vs each parameter (1D scatter per param) |
+//! | **Parallel coordinates** | Multi-parameter relationship view (color = objective) |
+//! | **Parameter importance** | Horizontal bar chart of Spearman-based importance |
+//! | **Trial timeline** | Duration/index of each trial, color-coded by state |
+//! | **Intermediate values** | Per-trial learning curves (if pruning data available) |
+//!
+//! # Usage
+//!
+//! Call [`Study::export_html()`](crate::Study::export_html) or
+//! [`generate_html_report()`] directly:
+//!
+//! ```no_run
+//! use optimizer::prelude::*;
+//!
+//! let study: Study<f64> = Study::new(Direction::Minimize);
+//! # let x = FloatParam::new(0.0, 1.0);
+//! # study.optimize(10, |trial: &mut optimizer::Trial| {
+//! #     let v = x.suggest(trial)?;
+//! #     Ok::<_, optimizer::Error>(v * v)
+//! # }).unwrap();
+//! study.export_html("report.html").unwrap();
+//! ```
+//!
+//! The output is a single HTML file that can be opened in any browser.
+//! An internet connection is needed on first load to fetch `Plotly.js`
+//! from a CDN.
+
+use core::fmt::Write as _;
+use std::collections::BTreeMap;
+use std::path::Path;
+
+use crate::param::ParamValue;
+use crate::parameter::ParamId;
+use crate::sampler::CompletedTrial;
+use crate::types::{Direction, TrialState};
+
+/// Generate an HTML report with interactive Plotly.js charts.
+///
+/// Create a self-contained HTML file at `path` containing up to six
+/// interactive charts. Charts that require data not present in the study
+/// (e.g., intermediate values) are automatically omitted.
+///
+/// The report includes: optimization history, slice plots, parallel
+/// coordinates, parameter importance, trial timeline, and intermediate
+/// values (when available).
+///
+/// This is also available as [`Study::export_html()`](crate::Study::export_html).
+///
+/// # Errors
+///
+/// Return an I/O error if the file cannot be created or written.
+pub fn generate_html_report(
+    study: &crate::Study<f64>,
+    path: impl AsRef<Path>,
+) -> std::io::Result<()> {
+    let trials = study.trials();
+    let direction = study.direction();
+    let importance = study.param_importance();
+
+    let html = build_html(&trials, direction, &importance);
+    std::fs::write(path, html)
+}
+
+fn build_html(
+    trials: &[CompletedTrial<f64>],
+    direction: Direction,
+    importance: &[(String, f64)],
+) -> String {
+    let mut html = String::with_capacity(8192);
+
+    let dir_label = match direction {
+        Direction::Minimize => "Minimize",
+        Direction::Maximize => "Maximize",
+    };
+
+    // Collect parameter metadata.
+    let param_info = collect_param_info(trials);
+    let has_intermediate = trials.iter().any(|t| !t.intermediate_values.is_empty());
+
+    let _ = write!(
+        html,
+        r#"<!DOCTYPE html>
+<html lang="en">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>Optimization Report</title>
+<script src="https://cdn.plot.ly/plotly-2.35.2.min.js"></script>
+<style>
+  * {{ margin: 0; padding: 0; box-sizing: border-box; }}
+  body {{ font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, sans-serif;
+         background: #f5f6fa; color: #2c3e50; padding: 24px; }}
+  h1 {{ text-align: center; margin-bottom: 8px; font-size: 1.8em; }}
+  .subtitle {{ text-align: center; color: #7f8c8d; margin-bottom: 24px; }}
+  .chart {{ background: #fff; border-radius: 8px; box-shadow: 0 2px 8px rgba(0,0,0,0.08);
+            margin-bottom: 24px; padding: 16px; }}
+  .chart-title {{ font-size: 1.1em; font-weight: 600; margin-bottom: 8px; }}
+</style>
+</head>
+<body>
+<h1>Optimization Report</h1>
+<p class="subtitle">{dir_label} &middot; {n} trials</p>
+"#,
+        n = trials.len(),
+    );
+
+    // Optimization history chart.
+    html.push_str("<div class=\"chart\"><div class=\"chart-title\">Optimization History</div><div id=\"history\"></div></div>\n");
+    write_history_chart(&mut html, trials, direction);
+
+    // Slice plots.
+    if !param_info.is_empty() {
+        html.push_str("<div class=\"chart\"><div class=\"chart-title\">Slice Plots</div><div id=\"slices\"></div></div>\n");
+        write_slice_charts(&mut html, trials, &param_info);
+    }
+
+    // Parallel coordinates.
+    if param_info.len() >= 2 {
+        html.push_str("<div class=\"chart\"><div class=\"chart-title\">Parallel Coordinates</div><div id=\"parcoords\"></div></div>\n");
+        write_parallel_coordinates(&mut html, trials, &param_info, direction);
+    }
+
+    // Parameter importance.
+    if !importance.is_empty() {
+        html.push_str("<div class=\"chart\"><div class=\"chart-title\">Parameter Importance</div><div id=\"importance\"></div></div>\n");
+        write_importance_chart(&mut html, importance);
+    }
+
+    // Trial timeline.
+    html.push_str("<div class=\"chart\"><div class=\"chart-title\">Trial Timeline</div><div id=\"timeline\"></div></div>\n");
+    write_timeline_chart(&mut html, trials);
+
+    // Intermediate values.
+    if has_intermediate {
+        html.push_str("<div class=\"chart\"><div class=\"chart-title\">Intermediate Values</div><div id=\"intermediate\"></div></div>\n");
+        write_intermediate_chart(&mut html, trials);
+    }
+
+    html.push_str("</body>\n</html>\n");
+    html
+}
+
+/// Metadata about each parameter seen across trials.
+struct ParamMeta {
+    label: String,
+}
+
+/// Collect parameter labels and distributions across all trials.
+fn collect_param_info(trials: &[CompletedTrial<f64>]) -> BTreeMap<ParamId, ParamMeta> {
+    let mut info: BTreeMap<ParamId, ParamMeta> = BTreeMap::new();
+    for trial in trials {
+        for &id in trial.params.keys() {
+            info.entry(id).or_insert_with(|| {
+                let label = trial
+                    .param_labels
+                    .get(&id)
+                    .cloned()
+                    .unwrap_or_else(|| id.to_string());
+                ParamMeta { label }
+            });
+        }
+    }
+    info
+}
+
+// ---------------------------------------------------------------------------
+// Chart generators
+// ---------------------------------------------------------------------------
+
+fn write_history_chart(html: &mut String, trials: &[CompletedTrial<f64>], direction: Direction) {
+    let complete: Vec<_> = trials
+        .iter()
+        .filter(|t| t.state == TrialState::Complete)
+        .collect();
+    if complete.is_empty() {
+        return;
+    }
+
+    let mut ids = Vec::with_capacity(complete.len());
+    let mut vals = Vec::with_capacity(complete.len());
+    let mut best_vals = Vec::with_capacity(complete.len());
+    let mut best = complete[0].value;
+    for t in &complete {
+        ids.push(t.id);
+        vals.push(t.value);
+        best = match direction {
+            Direction::Minimize => best.min(t.value),
+            Direction::Maximize => best.max(t.value),
+        };
+        best_vals.push(best);
+    }
+
+    let _ = write!(
+        html,
+        r##"<script>
+Plotly.newPlot("history", [
+  {{ x: {ids:?}, y: {vals:?}, mode: "markers", name: "Objective", type: "scatter",
+     marker: {{ color: "#3498db", size: 6 }} }},
+  {{ x: {ids:?}, y: {best_vals:?}, mode: "lines", name: "Best so far", type: "scatter",
+     line: {{ color: "#e74c3c", width: 2 }} }}
+], {{ xaxis: {{ title: "Trial" }}, yaxis: {{ title: "Objective Value" }},
+     margin: {{ t: 10 }}, legend: {{ x: 1, xanchor: "right", y: 1 }} }},
+   {{ responsive: true }});
+</script>
+"##,
+    );
+}
+
+fn write_slice_charts(
+    html: &mut String,
+    trials: &[CompletedTrial<f64>],
+    param_info: &BTreeMap<ParamId, ParamMeta>,
+) {
+    let complete: Vec<_> = trials
+        .iter()
+        .filter(|t| t.state == TrialState::Complete)
+        .collect();
+    if complete.is_empty() {
+        return;
+    }
+
+    let n_params = param_info.len();
+    let cols = if n_params <= 2 { n_params } else { 2 };
+    let rows = n_params.div_ceil(cols);
+
+    // Build subplot titles and data.
+    let mut subplot_titles = Vec::new();
+    let mut traces = String::new();
+    for (i, (id, meta)) in param_info.iter().enumerate() {
+        subplot_titles.push(format!("\"{}\"", escape_js(&meta.label)));
+
+        let mut x_vals = Vec::new();
+        let mut y_vals = Vec::new();
+        for t in &complete {
+            if let Some(pv) = t.params.get(id) {
+                x_vals.push(param_value_to_f64(pv));
+                y_vals.push(t.value);
+            }
+        }
+
+        let subplot_idx = i + 1;
+        let xa = if subplot_idx == 1 {
+            "x".to_string()
+        } else {
+            format!("x{subplot_idx}")
+        };
+        let ya = if subplot_idx == 1 {
+            "y".to_string()
+        } else {
+            format!("y{subplot_idx}")
+        };
+        let _ = write!(
+            traces,
+            r##"{{ x: {x_vals:?}, y: {y_vals:?}, mode: "markers", type: "scatter",
+               xaxis: "{xa}", yaxis: "{ya}",
+               marker: {{ color: "#3498db", size: 5 }}, showlegend: false }},"##,
+        );
+    }
+
+    let _ = write!(
+        html,
+        r#"<script>
+Plotly.newPlot("slices", [{traces}],
+  {{ grid: {{ rows: {rows}, columns: {cols}, pattern: "independent" }},
+     annotations: [{annotations}],
+     margin: {{ t: 30 }}, showlegend: false }},
+  {{ responsive: true }});
+</script>
+"#,
+        annotations = build_subplot_annotations(&subplot_titles, rows, cols),
+    );
+}
+
+fn write_parallel_coordinates(
+    html: &mut String,
+    trials: &[CompletedTrial<f64>],
+    param_info: &BTreeMap<ParamId, ParamMeta>,
+    direction: Direction,
+) {
+    let complete: Vec<_> = trials
+        .iter()
+        .filter(|t| t.state == TrialState::Complete)
+        .collect();
+    if complete.is_empty() {
+        return;
+    }
+
+    let mut dimensions = String::new();
+
+    // Add objective value as the first dimension.
+    let obj_vals: Vec<f64> = complete.iter().map(|t| t.value).collect();
+    let _ = write!(
+        dimensions,
+        r#"{{ label: "Objective", values: {obj_vals:?} }},"#,
+    );
+
+    // Add each parameter as a dimension.
+    for (id, meta) in param_info {
+        let vals: Vec<f64> = complete
+            .iter()
+            .map(|t| t.params.get(id).map_or(f64::NAN, param_value_to_f64))
+            .collect();
+        let _ = write!(
+            dimensions,
+            r#"{{ label: "{label}", values: {vals:?} }},"#,
+            label = escape_js(&meta.label),
+        );
+    }
+
+    // Color by objective value: green = good, red = bad.
+    let (cmin, cmax) = min_max(&obj_vals);
+    let colorscale = match direction {
+        Direction::Minimize => r##"[[0,"#2ecc71"],[1,"#e74c3c"]]"##,
+        Direction::Maximize => r##"[[0,"#e74c3c"],[1,"#2ecc71"]]"##,
+    };
+
+    let _ = write!(
+        html,
+        r#"<script>
+Plotly.newPlot("parcoords", [{{
+  type: "parcoords",
+  line: {{ color: {obj_vals:?}, colorscale: {colorscale},
+           cmin: {cmin}, cmax: {cmax}, showscale: true }},
+  dimensions: [{dimensions}]
+}}], {{ margin: {{ t: 10 }} }}, {{ responsive: true }});
+</script>
+"#,
+    );
+}
+
+fn write_importance_chart(html: &mut String, importance: &[(String, f64)]) {
+    let names: Vec<_> = importance
+        .iter()
+        .map(|(n, _)| format!("\"{}\"", escape_js(n)))
+        .collect();
+    let values: Vec<f64> = importance.iter().map(|(_, v)| *v).collect();
+
+    let _ = write!(
+        html,
+        r##"<script>
+Plotly.newPlot("importance", [{{
+  x: {values:?}, y: [{names}], type: "bar", orientation: "h",
+  marker: {{ color: "#9b59b6" }}
+}}], {{ xaxis: {{ title: "Importance (|Spearman correlation|)" }},
+       yaxis: {{ automargin: true }}, margin: {{ t: 10, l: 120 }} }},
+   {{ responsive: true }});
+</script>
+"##,
+        names = names.join(","),
+    );
+}
+
+fn write_timeline_chart(html: &mut String, trials: &[CompletedTrial<f64>]) {
+    let mut ids = Vec::with_capacity(trials.len());
+    let mut colors = Vec::with_capacity(trials.len());
+    let mut labels = Vec::with_capacity(trials.len());
+
+    for t in trials {
+        ids.push(format!("\"Trial {}\"", t.id));
+        let (color, label) = match t.state {
+            TrialState::Complete => ("#2ecc71", "Complete"),
+            TrialState::Pruned => ("#f39c12", "Pruned"),
+            TrialState::Failed => ("#e74c3c", "Failed"),
+            TrialState::Running => ("#3498db", "Running"),
+        };
+        colors.push(format!("\"{color}\""));
+        labels.push(format!("\"{label}\""));
+    }
+
+    // Use trial index as a proxy for duration (no wallclock data available).
+    let indices: Vec<usize> = (0..trials.len()).collect();
+
+    let _ = write!(
+        html,
+        r#"<script>
+Plotly.newPlot("timeline", [{{
+  y: [{ids}], x: {indices:?}, type: "bar", orientation: "h",
+  text: [{labels}], textposition: "auto",
+  marker: {{ color: [{colors}] }}
+}}], {{ xaxis: {{ title: "Trial Index" }}, yaxis: {{ automargin: true, autorange: "reversed" }},
+       margin: {{ t: 10, l: 80 }}, showlegend: false }},
+   {{ responsive: true }});
+</script>
+"#,
+        ids = ids.join(","),
+        colors = colors.join(","),
+        labels = labels.join(","),
+    );
+}
+
+fn write_intermediate_chart(html: &mut String, trials: &[CompletedTrial<f64>]) {
+    let trials_with_iv: Vec<_> = trials
+        .iter()
+        .filter(|t| !t.intermediate_values.is_empty())
+        .collect();
+    if trials_with_iv.is_empty() {
+        return;
+    }
+
+    let mut traces = String::new();
+    for t in &trials_with_iv {
+        let steps: Vec<u64> = t.intermediate_values.iter().map(|(s, _)| *s).collect();
+        let values: Vec<f64> = t.intermediate_values.iter().map(|(_, v)| *v).collect();
+        let color = match t.state {
+            TrialState::Pruned => "#f39c12",
+            _ => "#3498db",
+        };
+        let _ = write!(
+            traces,
+            r#"{{ x: {steps:?}, y: {values:?}, mode: "lines+markers", name: "Trial {id}",
+               line: {{ color: "{color}", width: 1 }}, marker: {{ size: 3 }} }},"#,
+            id = t.id,
+        );
+    }
+
+    let _ = write!(
+        html,
+        r#"<script>
+Plotly.newPlot("intermediate", [{traces}],
+  {{ xaxis: {{ title: "Step" }}, yaxis: {{ title: "Intermediate Value" }},
+     margin: {{ t: 10 }}, showlegend: true }},
+  {{ responsive: true }});
+</script>
+"#,
+    );
+}
+
+// ---------------------------------------------------------------------------
+// Helpers
+// ---------------------------------------------------------------------------
+
+#[allow(clippy::cast_precision_loss)]
+fn param_value_to_f64(pv: &ParamValue) -> f64 {
+    match *pv {
+        ParamValue::Float(v) => v,
+        ParamValue::Int(v) => v as f64,
+        ParamValue::Categorical(v) => v as f64,
+    }
+}
+
+fn escape_js(s: &str) -> String {
+    s.replace('\\', "\\\\")
+        .replace('"', "\\\"")
+        .replace('\n', "\\n")
+}
+
+fn min_max(vals: &[f64]) -> (f64, f64) {
+    let mut mn = f64::INFINITY;
+    let mut mx = f64::NEG_INFINITY;
+    for &v in vals {
+        if v.is_nan() {
+            continue;
+        }
+        if v < mn {
+            mn = v;
+        }
+        if v > mx {
+            mx = v;
+        }
+    }
+    // If all values were NaN, return 0.0..1.0 as a safe fallback.
+    if mn > mx {
+        return (0.0, 1.0);
+    }
+    (mn, mx)
+}
+
+/// Build Plotly annotation objects to act as subplot titles.
+#[allow(clippy::cast_precision_loss)]
+fn build_subplot_annotations(titles: &[String], rows: usize, cols: usize) -> String {
+    let mut anns = Vec::new();
+    for (i, title) in titles.iter().enumerate() {
+        let row = i / cols;
+        let col = i % cols;
+        // Compute x/y anchor in paper coordinates.
+        let x = if cols == 1 {
+            0.5
+        } else {
+            (f64::from(u32::try_from(col).unwrap_or(0))) / (cols as f64 - 1.0)
+        };
+        let y = 1.0 - (f64::from(u32::try_from(row).unwrap_or(0))) / (rows as f64).max(1.0) + 0.02;
+        anns.push(format!(
+            r#"{{ text: {title}, x: {x:.3}, y: {y:.3}, xref: "paper", yref: "paper",
+               showarrow: false, font: {{ size: 12 }} }}"#,
+        ));
+    }
+    anns.join(",")
+}

tests/async_tests.rs

@@ -4,6 +4,8 @@
 
 #![cfg(feature = "async")]
 
+use std::sync::Arc;
+
 use optimizer::parameter::{FloatParam, Parameter};
 use optimizer::sampler::random::RandomSampler;
 use optimizer::sampler::tpe::TpeSampler;
@@ -17,12 +19,9 @@ async fn test_optimize_async_basic() {
     let x_param = FloatParam::new(-10.0, 10.0);
 
     study
-        .optimize_async(10, move |mut trial| {
-            let x_param = x_param.clone();
-            async move {
-                let x = x_param.suggest(&mut trial)?;
-                Ok::<_, Error>((trial, x * x))
-            }
+        .optimize_async(10, move |trial: &mut optimizer::Trial| {
+            let x = x_param.suggest(trial)?;
+            Ok::<_, Error>(x * x)
         })
         .await
         .expect("async optimization should succeed");
@@ -45,12 +44,9 @@ async fn test_optimize_async_with_tpe() {
     let x_param = FloatParam::new(-5.0, 5.0);
 
     study
-        .optimize_async(15, move |mut trial| {
-            let x_param = x_param.clone();
-            async move {
-                let x = x_param.suggest(&mut trial)?;
-                Ok::<_, Error>((trial, x * x))
-            }
+        .optimize_async(15, move |trial: &mut optimizer::Trial| {
+            let x = x_param.suggest(trial)?;
+            Ok::<_, Error>(x * x)
         })
         .await
         .expect("async optimization with sampler should succeed");
@@ -68,12 +64,9 @@ async fn test_optimize_parallel() {
     let x_param = FloatParam::new(-10.0, 10.0);
 
     study
-        .optimize_parallel(20, 4, move |mut trial| {
-            let x_param = x_param.clone();
-            async move {
-                let x = x_param.suggest(&mut trial)?;
-                Ok::<_, Error>((trial, x * x))
-            }
+        .optimize_parallel(20, 4, move |trial: &mut optimizer::Trial| {
+            let x = x_param.suggest(trial)?;
+            Ok::<_, Error>(x * x)
         })
         .await
         .expect("parallel optimization should succeed");
@@ -95,14 +88,10 @@ async fn test_optimize_parallel_with_tpe() {
     let y_param = FloatParam::new(-5.0, 5.0);
 
     study
-        .optimize_parallel(15, 3, move |mut trial| {
-            let x_param = x_param.clone();
-            let y_param = y_param.clone();
-            async move {
-                let x = x_param.suggest(&mut trial)?;
-                let y = y_param.suggest(&mut trial)?;
-                Ok::<_, Error>((trial, x * x + y * y))
-            }
+        .optimize_parallel(15, 3, move |trial: &mut optimizer::Trial| {
+            let x = x_param.suggest(trial)?;
+            let y = y_param.suggest(trial)?;
+            Ok::<_, Error>(x * x + y * y)
         })
         .await
         .expect("parallel optimization with sampler should succeed");
@@ -115,27 +104,8 @@ async fn test_optimize_async_all_failures() {
     let study: Study<f64> = Study::new(Direction::Minimize);
 
     let result = study
-        .optimize_async(5, |trial| async move {
-            let _ = trial;
-            Err::<(_, f64), &str>("always fails")
-        })
-        .await;
-
-    assert!(
-        matches!(result, Err(Error::NoCompletedTrials)),
-        "should return NoCompletedTrials when all trials fail"
-    );
-}
-
-#[tokio::test]
-#[allow(deprecated)]
-async fn test_optimize_async_with_sampler_all_failures() {
-    let study: Study<f64> = Study::new(Direction::Minimize);
-
-    let result = study
-        .optimize_async_with_sampler(5, |trial| async move {
-            let _ = trial;
-            Err::<(_, f64), &str>("always fails")
+        .optimize_async(5, |_trial: &mut optimizer::Trial| {
+            Err::<f64, &str>("always fails")
         })
         .await;
 
@@ -150,27 +120,8 @@ async fn test_optimize_parallel_all_failures() {
     let study: Study<f64> = Study::new(Direction::Minimize);
 
     let result = study
-        .optimize_parallel(5, 2, |trial| async move {
-            let _ = trial;
-            Err::<(_, f64), &str>("always fails")
-        })
-        .await;
-
-    assert!(
-        matches!(result, Err(Error::NoCompletedTrials)),
-        "should return NoCompletedTrials when all trials fail"
-    );
-}
-
-#[tokio::test]
-#[allow(deprecated)]
-async fn test_optimize_parallel_with_sampler_all_failures() {
-    let study: Study<f64> = Study::new(Direction::Minimize);
-
-    let result = study
-        .optimize_parallel_with_sampler(5, 2, |trial| async move {
-            let _ = trial;
-            Err::<(_, f64), &str>("always fails")
+        .optimize_parallel(5, 2, |_trial: &mut optimizer::Trial| {
+            Err::<f64, &str>("always fails")
         })
         .await;
 
@@ -190,16 +141,13 @@ async fn test_optimize_async_partial_failures() {
     let x_param = FloatParam::new(0.0, 10.0);
 
     study
-        .optimize_async(10, move |mut trial| {
+        .optimize_async(10, move |trial: &mut optimizer::Trial| {
             let count = counter.fetch_add(1, std::sync::atomic::Ordering::SeqCst);
-            let x_param = x_param.clone();
-            async move {
-                if count.is_multiple_of(2) {
-                    let x = x_param.suggest(&mut trial)?;
-                    Ok::<_, Error>((trial, x))
-                } else {
-                    Err(Error::NoCompletedTrials) // Use as error type
-                }
+            if count.is_multiple_of(2) {
+                let x = x_param.suggest(trial)?;
+                Ok::<_, Error>(x)
+            } else {
+                Err(Error::NoCompletedTrials) // Use as error type
             }
         })
         .await
@@ -218,12 +166,9 @@ async fn test_optimize_parallel_high_concurrency() {
 
     // Run with concurrency higher than n_trials
     study
-        .optimize_parallel(5, 10, move |mut trial| {
-            let x_param = x_param.clone();
-            async move {
-                let x = x_param.suggest(&mut trial)?;
-                Ok::<_, Error>((trial, x))
-            }
+        .optimize_parallel(5, 10, move |trial: &mut optimizer::Trial| {
+            let x = x_param.suggest(trial)?;
+            Ok::<_, Error>(x)
         })
         .await
         .expect("should handle high concurrency");
@@ -240,15 +185,139 @@ async fn test_optimize_parallel_single_concurrency() {
 
     // Run with concurrency of 1 (sequential)
     study
-        .optimize_parallel(10, 1, move |mut trial| {
-            let x_param = x_param.clone();
-            async move {
-                let x = x_param.suggest(&mut trial)?;
-                Ok::<_, Error>((trial, x))
-            }
+        .optimize_parallel(10, 1, move |trial: &mut optimizer::Trial| {
+            let x = x_param.suggest(trial)?;
+            Ok::<_, Error>(x)
         })
         .await
         .expect("should work with single concurrency");
 
     assert_eq!(study.n_trials(), 10);
 }
+
+#[tokio::test]
+async fn test_parallel_executes_concurrently() {
+    use std::sync::atomic::{AtomicUsize, Ordering};
+
+    let sampler = RandomSampler::with_seed(42);
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+
+    let x_param = FloatParam::new(0.0, 10.0);
+    let active = Arc::new(AtomicUsize::new(0));
+    let max_active = Arc::new(AtomicUsize::new(0));
+
+    let active_c = Arc::clone(&active);
+    let max_active_c = Arc::clone(&max_active);
+
+    study
+        .optimize_parallel(4, 4, move |trial: &mut optimizer::Trial| {
+            let x = x_param.suggest(trial)?;
+
+            let current = active_c.fetch_add(1, Ordering::SeqCst) + 1;
+            max_active_c.fetch_max(current, Ordering::SeqCst);
+
+            std::thread::sleep(std::time::Duration::from_millis(50));
+
+            active_c.fetch_sub(1, Ordering::SeqCst);
+            Ok::<_, Error>(x)
+        })
+        .await
+        .expect("parallel optimization should succeed");
+
+    assert_eq!(study.n_trials(), 4);
+    let max = max_active.load(Ordering::SeqCst);
+    // With 4 trials and concurrency=4, all should run concurrently
+    assert!(
+        max >= 2,
+        "expected at least 2 concurrent workers, but max was {max}"
+    );
+}
+
+#[tokio::test]
+async fn test_parallel_max_concurrency_reached() {
+    use std::sync::atomic::{AtomicUsize, Ordering};
+
+    let sampler = RandomSampler::with_seed(42);
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+
+    let x_param = FloatParam::new(0.0, 10.0);
+    let active = Arc::new(AtomicUsize::new(0));
+    let max_active = Arc::new(AtomicUsize::new(0));
+
+    let active_c = Arc::clone(&active);
+    let max_active_c = Arc::clone(&max_active);
+
+    study
+        .optimize_parallel(8, 4, move |trial: &mut optimizer::Trial| {
+            let x = x_param.suggest(trial)?;
+
+            let current = active_c.fetch_add(1, Ordering::SeqCst) + 1;
+            // Update max_active if this is the highest seen so far
+            max_active_c.fetch_max(current, Ordering::SeqCst);
+
+            std::thread::sleep(std::time::Duration::from_millis(50));
+
+            active_c.fetch_sub(1, Ordering::SeqCst);
+            Ok::<_, Error>(x)
+        })
+        .await
+        .expect("parallel optimization should succeed");
+
+    assert_eq!(study.n_trials(), 8);
+    let max = max_active.load(Ordering::SeqCst);
+    assert!(
+        max >= 2,
+        "expected at least 2 concurrent workers, but max was {max}"
+    );
+}
+
+#[tokio::test]
+async fn test_parallel_panic_returns_task_error() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+
+    let result = study
+        .optimize_parallel(3, 2, |_trial: &mut optimizer::Trial| {
+            panic!("boom");
+            #[allow(unreachable_code)]
+            Ok::<_, Error>(0.0)
+        })
+        .await;
+
+    match result {
+        Err(Error::TaskError(msg)) => {
+            assert!(
+                msg.contains("panic"),
+                "expected error message to contain 'panic', got: {msg}"
+            );
+        }
+        other => panic!("expected TaskError, got {other:?}"),
+    }
+}
+
+#[tokio::test]
+async fn test_parallel_partial_failures_trial_count() {
+    use std::sync::atomic::{AtomicUsize, Ordering};
+
+    let sampler = RandomSampler::with_seed(42);
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+
+    let x_param = FloatParam::new(0.0, 10.0);
+    let counter = Arc::new(AtomicUsize::new(0));
+
+    let counter_c = Arc::clone(&counter);
+
+    study
+        .optimize_parallel(10, 3, move |trial: &mut optimizer::Trial| {
+            let idx = counter_c.fetch_add(1, Ordering::SeqCst);
+            if idx % 2 == 1 {
+                return Err(Error::NoCompletedTrials);
+            }
+            let x = x_param.suggest(trial)?;
+            Ok::<_, Error>(x)
+        })
+        .await
+        .expect("should succeed with partial failures");
+
+    // Even indices succeed (0, 2, 4, 6, 8), odd indices fail
+    assert_eq!(study.n_trials(), 5);
+}

tests/export_tests.rs

@@ -0,0 +1,230 @@
+use optimizer::parameter::{FloatParam, IntParam, Parameter};
+use optimizer::sampler::random::RandomSampler;
+use optimizer::{Direction, Study};
+
+#[test]
+fn csv_empty_study_produces_header_only() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+    let mut buf = Vec::new();
+    study.to_csv(&mut buf).unwrap();
+    let csv = String::from_utf8(buf).unwrap();
+    assert_eq!(csv, "trial_id,value,state\n");
+}
+
+#[test]
+fn csv_includes_all_trial_data() {
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, RandomSampler::with_seed(42));
+    let x = FloatParam::new(0.0, 10.0).name("x");
+    let y = IntParam::new(1, 5).name("y");
+
+    study
+        .optimize(3, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            let yv = y.suggest(trial)?;
+            Ok::<_, optimizer::Error>(xv + yv as f64)
+        })
+        .unwrap();
+
+    let mut buf = Vec::new();
+    study.to_csv(&mut buf).unwrap();
+    let csv = String::from_utf8(buf).unwrap();
+
+    let lines: Vec<&str> = csv.lines().collect();
+    // Header + 3 data rows.
+    assert_eq!(lines.len(), 4);
+
+    // Header should contain our parameter names.
+    let header = lines[0];
+    assert!(header.starts_with("trial_id,value,state"));
+    assert!(header.contains("x"));
+    assert!(header.contains("y"));
+
+    // Each data row should have the right number of columns.
+    let n_cols = header.split(',').count();
+    for line in &lines[1..] {
+        assert_eq!(line.split(',').count(), n_cols);
+    }
+
+    // All rows should have "Complete" state.
+    for line in &lines[1..] {
+        assert!(line.contains("Complete"));
+    }
+}
+
+#[test]
+fn csv_handles_pruned_trials() {
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, RandomSampler::with_seed(42));
+    let x = FloatParam::new(0.0, 10.0).name("x");
+
+    // First trial: complete
+    let mut trial = study.create_trial();
+    let _ = x.suggest(&mut trial).unwrap();
+    study.complete_trial(trial, 1.0);
+
+    // Second trial: pruned
+    let mut trial = study.create_trial();
+    let _ = x.suggest(&mut trial).unwrap();
+    study.prune_trial(trial);
+
+    let mut buf = Vec::new();
+    study.to_csv(&mut buf).unwrap();
+    let csv = String::from_utf8(buf).unwrap();
+
+    let lines: Vec<&str> = csv.lines().collect();
+    assert_eq!(lines.len(), 3); // header + 2 data rows
+
+    // Pruned trial should have empty value.
+    let pruned_line = lines[2];
+    assert!(pruned_line.contains("Pruned"));
+    // The value field (second column) should be empty.
+    let cols: Vec<&str> = pruned_line.split(',').collect();
+    assert_eq!(cols[2], "Pruned");
+    assert_eq!(cols[1], ""); // empty value for pruned
+}
+
+#[test]
+fn csv_handles_different_parameter_sets() {
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, RandomSampler::with_seed(42));
+    let x = FloatParam::new(0.0, 10.0).name("x");
+    let y = FloatParam::new(0.0, 10.0).name("y");
+
+    // First trial: only x
+    let mut trial = study.create_trial();
+    let xv = x.suggest(&mut trial).unwrap();
+    study.complete_trial(trial, xv);
+
+    // Second trial: only y
+    let mut trial = study.create_trial();
+    let yv = y.suggest(&mut trial).unwrap();
+    study.complete_trial(trial, yv);
+
+    let mut buf = Vec::new();
+    study.to_csv(&mut buf).unwrap();
+    let csv = String::from_utf8(buf).unwrap();
+
+    let lines: Vec<&str> = csv.lines().collect();
+    assert_eq!(lines.len(), 3);
+
+    // Both x and y columns should exist.
+    let header = lines[0];
+    assert!(header.contains("x"));
+    assert!(header.contains("y"));
+
+    // Each row has the right column count (missing params are empty).
+    let n_cols = header.split(',').count();
+    for line in &lines[1..] {
+        assert_eq!(line.split(',').count(), n_cols);
+    }
+}
+
+#[test]
+fn csv_output_is_parseable() {
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, RandomSampler::with_seed(42));
+    let lr = FloatParam::new(0.001, 0.1).name("learning_rate");
+    let layers = IntParam::new(1, 5).name("n_layers");
+
+    study
+        .optimize(5, |trial: &mut optimizer::Trial| {
+            let l = lr.suggest(trial)?;
+            let n = layers.suggest(trial)?;
+            Ok::<_, optimizer::Error>(l * n as f64)
+        })
+        .unwrap();
+
+    let mut buf = Vec::new();
+    study.to_csv(&mut buf).unwrap();
+    let csv = String::from_utf8(buf).unwrap();
+
+    // Parse each row: every value field should be a valid f64 for complete trials.
+    let lines: Vec<&str> = csv.lines().collect();
+    for line in &lines[1..] {
+        let cols: Vec<&str> = line.split(',').collect();
+        // trial_id should be a number
+        cols[0].parse::<u64>().unwrap();
+        // value should be parseable as f64
+        cols[1].parse::<f64>().unwrap();
+        // state should be a known value
+        assert!(["Complete", "Pruned", "Failed", "Running"].contains(&cols[2]));
+    }
+}
+
+#[test]
+fn export_csv_writes_file() {
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, RandomSampler::with_seed(42));
+    let x = FloatParam::new(0.0, 10.0).name("x");
+    study
+        .optimize(3, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            Ok::<_, optimizer::Error>(xv * xv)
+        })
+        .unwrap();
+
+    let dir = std::env::temp_dir().join("optimizer_export_test");
+    std::fs::create_dir_all(&dir).unwrap();
+    let path = dir.join("test_export.csv");
+
+    study.export_csv(&path).unwrap();
+
+    let contents = std::fs::read_to_string(&path).unwrap();
+    assert!(contents.starts_with("trial_id,value,state"));
+    assert!(contents.lines().count() == 4); // header + 3 rows
+
+    // Clean up.
+    let _ = std::fs::remove_dir_all(&dir);
+}
+
+#[cfg(feature = "serde")]
+#[test]
+fn export_json_writes_file() {
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, RandomSampler::with_seed(42));
+    let x = FloatParam::new(0.0, 10.0).name("x");
+    study
+        .optimize(3, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            Ok::<_, optimizer::Error>(xv * xv)
+        })
+        .unwrap();
+
+    let dir = std::env::temp_dir().join("optimizer_json_export_test");
+    std::fs::create_dir_all(&dir).unwrap();
+    let path = dir.join("test_export.json");
+
+    study.export_json(&path).unwrap();
+
+    let contents = std::fs::read_to_string(&path).unwrap();
+    let parsed: serde_json::Value = serde_json::from_str(&contents).unwrap();
+    let arr = parsed.as_array().unwrap();
+    assert_eq!(arr.len(), 3);
+
+    // Each entry should have the expected fields.
+    for entry in arr {
+        assert!(entry.get("id").is_some());
+        assert!(entry.get("value").is_some());
+        assert!(entry.get("state").is_some());
+        assert!(entry.get("params").is_some());
+    }
+
+    // Clean up.
+    let _ = std::fs::remove_dir_all(&dir);
+}
+
+#[test]
+fn csv_includes_user_attributes() {
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, RandomSampler::with_seed(42));
+    let x = FloatParam::new(0.0, 10.0).name("x");
+
+    study
+        .optimize(2, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            trial.set_user_attr("training_time_secs", 45.2);
+            Ok::<_, optimizer::Error>(xv * xv)
+        })
+        .unwrap();
+
+    let mut buf = Vec::new();
+    study.to_csv(&mut buf).unwrap();
+    let csv = String::from_utf8(buf).unwrap();
+
+    let header = csv.lines().next().unwrap();
+    assert!(header.contains("training_time_secs"));
+}

tests/fanova_tests.rs

@@ -0,0 +1,86 @@
+//! Integration tests for fANOVA parameter importance.
+
+use optimizer::prelude::*;
+
+#[test]
+fn fanova_dominant_parameter() {
+    // f(x, y) = x^2 — x should dominate
+    let x = FloatParam::new(0.0, 10.0).name("x");
+    let y = FloatParam::new(0.0, 10.0).name("y");
+
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, RandomSampler::with_seed(42));
+    study
+        .optimize(50, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            let _yv = y.suggest(trial)?;
+            Ok::<_, Error>(xv * xv)
+        })
+        .unwrap();
+
+    let result = study.fanova().unwrap();
+    assert_eq!(result.main_effects[0].0, "x");
+    assert!(
+        result.main_effects[0].1 > 0.7,
+        "x importance = {}",
+        result.main_effects[0].1
+    );
+}
+
+#[test]
+fn fanova_interaction() {
+    // f(x, y) = x * y — both matter and interact
+    let x = FloatParam::new(0.0, 10.0).name("x");
+    let y = FloatParam::new(0.0, 10.0).name("y");
+
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, RandomSampler::with_seed(7));
+    study
+        .optimize(100, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            let yv = y.suggest(trial)?;
+            Ok::<_, Error>(xv * yv)
+        })
+        .unwrap();
+
+    let config = FanovaConfig {
+        n_trees: 128,
+        ..FanovaConfig::default()
+    };
+    let result = study.fanova_with_config(&config).unwrap();
+
+    // Should detect interaction
+    assert!(
+        !result.interactions.is_empty(),
+        "should detect x*y interaction"
+    );
+}
+
+#[test]
+fn fanova_consistent_with_correlation() {
+    // f(x, y) = 3*x + 0.5*y — x should rank higher in both methods
+    let x = FloatParam::new(0.0, 10.0).name("x");
+    let y = FloatParam::new(0.0, 10.0).name("y");
+
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, RandomSampler::with_seed(99));
+    study
+        .optimize(80, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            let yv = y.suggest(trial)?;
+            Ok::<_, Error>(3.0 * xv + 0.5 * yv)
+        })
+        .unwrap();
+
+    let corr = study.param_importance();
+    let fanova = study.fanova().unwrap();
+
+    // Both methods should rank x above y
+    assert_eq!(corr[0].0, "x", "correlation should rank x first");
+    assert_eq!(fanova.main_effects[0].0, "x", "fanova should rank x first");
+}
+
+#[test]
+fn fanova_too_few_trials() {
+    let study: Study<f64> = Study::new(Direction::Minimize);
+
+    let result = study.fanova();
+    assert!(result.is_err(), "should error with no trials");
+}

tests/importance_tests.rs

@@ -1,9 +1,10 @@
 use optimizer::parameter::{CategoricalParam, FloatParam, IntParam, Parameter};
+use optimizer::sampler::random::RandomSampler;
 use optimizer::{Direction, Study};
 
 #[test]
 fn known_perfect_correlation() {
-    let study: Study<f64> = Study::new(Direction::Minimize);
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, RandomSampler::with_seed(42));
     let x = FloatParam::new(0.0, 100.0).name("x");
 
     // Objective = x, so perfect correlation.
@@ -24,7 +25,7 @@ fn known_perfect_correlation() {
 
 #[test]
 fn no_effect_parameter() {
-    let study: Study<f64> = Study::new(Direction::Minimize);
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, RandomSampler::with_seed(42));
     let x = FloatParam::new(0.0, 100.0).name("x");
     let noise = FloatParam::new(0.0, 100.0).name("noise");
 
@@ -51,7 +52,7 @@ fn no_effect_parameter() {
 
 #[test]
 fn multiple_parameters_varying_importance() {
-    let study: Study<f64> = Study::new(Direction::Minimize);
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, RandomSampler::with_seed(42));
     let x = FloatParam::new(0.0, 10.0).name("x");
     let y = FloatParam::new(0.0, 10.0).name("y");
 
@@ -86,7 +87,7 @@ fn fewer_than_two_trials_returns_empty() {
 
 #[test]
 fn int_parameter() {
-    let study: Study<f64> = Study::new(Direction::Minimize);
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, RandomSampler::with_seed(42));
     let n = IntParam::new(1, 100).name("n");
 
     for _ in 0..30 {
@@ -103,7 +104,7 @@ fn int_parameter() {
 
 #[test]
 fn categorical_parameter() {
-    let study: Study<f64> = Study::new(Direction::Minimize);
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, RandomSampler::with_seed(42));
     let cat = CategoricalParam::new(vec!["a", "b", "c"]).name("cat");
     let x = FloatParam::new(0.0, 100.0).name("x");
 
@@ -123,7 +124,7 @@ fn categorical_parameter() {
 
 #[test]
 fn normalization_sums_to_one() {
-    let study: Study<f64> = Study::new(Direction::Minimize);
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, RandomSampler::with_seed(42));
     let x = FloatParam::new(0.0, 10.0).name("x");
     let y = FloatParam::new(0.0, 10.0).name("y");
     let z = FloatParam::new(0.0, 10.0).name("z");
@@ -146,7 +147,7 @@ fn normalization_sums_to_one() {
 
 #[test]
 fn label_when_unnamed_uses_debug() {
-    let study: Study<f64> = Study::new(Direction::Minimize);
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, RandomSampler::with_seed(42));
     // No .name() call → label defaults to Debug representation.
     let x = FloatParam::new(0.0, 10.0);
 

tests/journal_tests.rs

@@ -0,0 +1,540 @@
+//! Integration tests for the journal storage backend.
+
+use std::collections::HashMap;
+use std::sync::Arc;
+
+use std::io::Write;
+
+use optimizer::parameter::{FloatParam, Parameter};
+use optimizer::sampler::CompletedTrial;
+use optimizer::sampler::random::RandomSampler;
+use optimizer::storage::{JournalStorage, Storage};
+use optimizer::{Direction, Study};
+
+fn temp_path() -> std::path::PathBuf {
+    use std::sync::atomic::{AtomicU64, Ordering};
+    static COUNTER: AtomicU64 = AtomicU64::new(0);
+
+    let mut path = std::env::temp_dir();
+    path.push(format!(
+        "optimizer_journal_test_{}_{}.jsonl",
+        std::process::id(),
+        COUNTER.fetch_add(1, Ordering::Relaxed)
+    ));
+    path
+}
+
+fn sample_trial(id: u64, value: f64) -> CompletedTrial<f64> {
+    CompletedTrial::new(id, HashMap::new(), HashMap::new(), HashMap::new(), value)
+}
+
+#[test]
+fn roundtrip_single_trial() {
+    let path = temp_path();
+    let storage = JournalStorage::new(&path);
+
+    storage.push(sample_trial(0, 42.0));
+
+    let loaded = storage.trials_arc().read().clone();
+    assert_eq!(loaded.len(), 1);
+    assert_eq!(loaded[0].id, 0);
+    assert_eq!(loaded[0].value, 42.0);
+
+    // Also verify via a fresh open from disk
+    let storage2 = JournalStorage::<f64>::open(&path).unwrap();
+    let loaded2 = storage2.trials_arc().read().clone();
+    assert_eq!(loaded2.len(), 1);
+    assert_eq!(loaded2[0].value, 42.0);
+
+    std::fs::remove_file(&path).ok();
+}
+
+#[test]
+fn append_multiple_trials() {
+    let path = temp_path();
+    let storage = JournalStorage::new(&path);
+
+    for i in 0..5 {
+        storage.push(sample_trial(i, i as f64));
+    }
+
+    // Reload from disk
+    let storage2 = JournalStorage::<f64>::open(&path).unwrap();
+    let loaded = storage2.trials_arc().read().clone();
+    assert_eq!(loaded.len(), 5);
+    for (i, trial) in loaded.iter().enumerate() {
+        assert_eq!(trial.id, i as u64);
+        assert_eq!(trial.value, i as f64);
+    }
+
+    std::fs::remove_file(&path).ok();
+}
+
+#[test]
+fn missing_file_returns_empty() {
+    let path = temp_path();
+    let storage = JournalStorage::<f64>::open(&path).unwrap();
+
+    let loaded = storage.trials_arc().read().clone();
+    assert!(loaded.is_empty());
+}
+
+#[test]
+fn concurrent_writes() {
+    let path = temp_path();
+    let storage = Arc::new(JournalStorage::new(&path));
+
+    let mut handles = Vec::new();
+    for thread_id in 0..4u64 {
+        let s = Arc::clone(&storage);
+        handles.push(std::thread::spawn(move || {
+            for i in 0..25u64 {
+                let id = thread_id * 25 + i;
+                s.push(sample_trial(id, id as f64));
+            }
+        }));
+    }
+    for h in handles {
+        h.join().unwrap();
+    }
+
+    // Reload from disk to verify persistence
+    let storage2 = JournalStorage::<f64>::open(&path).unwrap();
+    let loaded = storage2.trials_arc().read().clone();
+    assert_eq!(loaded.len(), 100);
+
+    // Verify all IDs are present (order may vary)
+    let mut ids: Vec<u64> = loaded.iter().map(|t| t.id).collect();
+    ids.sort();
+    assert_eq!(ids, (0..100).collect::<Vec<_>>());
+
+    std::fs::remove_file(&path).ok();
+}
+
+#[test]
+fn study_with_journal_integration() {
+    let path = temp_path();
+    let x = FloatParam::new(-10.0, 10.0);
+
+    // First "process": run some trials
+    {
+        let study =
+            Study::with_journal(Direction::Minimize, RandomSampler::with_seed(1), &path).unwrap();
+        study
+            .optimize(5, |trial: &mut optimizer::Trial| {
+                let val = x.suggest(trial)?;
+                Ok::<_, optimizer::Error>(val * val)
+            })
+            .unwrap();
+        assert_eq!(study.n_trials(), 5);
+    }
+
+    // Second "process": loads the same file, sees existing trials
+    let study2 =
+        Study::with_journal(Direction::Minimize, RandomSampler::with_seed(2), &path).unwrap();
+    assert_eq!(study2.n_trials(), 5);
+
+    // Continue optimizing
+    study2
+        .optimize(5, |trial: &mut optimizer::Trial| {
+            let val = x.suggest(trial)?;
+            Ok::<_, optimizer::Error>(val * val)
+        })
+        .unwrap();
+    assert_eq!(study2.n_trials(), 10);
+
+    // Verify all 10 written to disk
+    let storage3 = JournalStorage::<f64>::open(&path).unwrap();
+    let loaded = storage3.trials_arc().read().clone();
+    assert_eq!(loaded.len(), 10);
+
+    std::fs::remove_file(&path).ok();
+}
+
+#[test]
+fn ids_are_unique_after_reload() {
+    let path = temp_path();
+
+    // First batch
+    {
+        let study =
+            Study::with_journal(Direction::Minimize, RandomSampler::with_seed(1), &path).unwrap();
+        study
+            .optimize(3, |trial: &mut optimizer::Trial| {
+                let _ = FloatParam::new(0.0, 1.0).suggest(trial)?;
+                Ok::<_, optimizer::Error>(1.0)
+            })
+            .unwrap();
+    }
+
+    // Second batch — IDs should continue from 3
+    let study =
+        Study::with_journal(Direction::Minimize, RandomSampler::with_seed(2), &path).unwrap();
+    study
+        .optimize(3, |trial: &mut optimizer::Trial| {
+            let _ = FloatParam::new(0.0, 1.0).suggest(trial)?;
+            Ok::<_, optimizer::Error>(1.0)
+        })
+        .unwrap();
+
+    let all = study.trials();
+    let mut ids: Vec<u64> = all.iter().map(|t| t.id).collect();
+    ids.sort();
+    // All 6 IDs should be unique
+    ids.dedup();
+    assert_eq!(ids.len(), 6);
+
+    std::fs::remove_file(&path).ok();
+}
+
+#[test]
+fn pruned_trials_are_stored() {
+    let path = temp_path();
+    let study =
+        Study::with_journal(Direction::Minimize, RandomSampler::with_seed(1), &path).unwrap();
+
+    // Complete one, prune one
+    let x = FloatParam::new(0.0, 1.0);
+    study
+        .optimize(3, |trial: &mut optimizer::Trial| {
+            let _ = x.suggest(trial)?;
+            if trial.id() == 1 {
+                Err(optimizer::TrialPruned)?;
+            }
+            Ok::<_, optimizer::Error>(1.0)
+        })
+        .unwrap();
+
+    let storage2 = JournalStorage::<f64>::open(&path).unwrap();
+    let loaded = storage2.trials_arc().read().clone();
+    assert_eq!(loaded.len(), 3);
+    assert!(
+        loaded
+            .iter()
+            .any(|t| t.state == optimizer::TrialState::Pruned)
+    );
+
+    std::fs::remove_file(&path).ok();
+}
+
+#[test]
+fn rejects_non_finite_values_in_journal() {
+    // serde_json rejects 1e999 ("number out of range"), so non-finite
+    // floats cannot sneak in through standard JSON.  Verify the overall
+    // loading path catches the error regardless of which layer rejects it.
+    let path = temp_path();
+    std::fs::write(
+        &path,
+        r#"{"id":0,"params":{},"distributions":{"0":{"Float":{"low":0.0,"high":1e999,"log_scale":false,"step":null}}},"param_labels":{},"value":1.0,"intermediate_values":[],"state":"Complete","user_attrs":{},"constraints":[]}"#,
+    )
+    .unwrap();
+
+    assert!(JournalStorage::<f64>::open(&path).is_err());
+    std::fs::remove_file(&path).ok();
+}
+
+#[test]
+fn validate_rejects_non_finite_distribution_bound() {
+    use optimizer::distribution::{Distribution, FloatDistribution};
+
+    let pid = FloatParam::new(0.0, 1.0).id();
+    let mut trial = sample_trial(0, 1.0);
+    trial.distributions.insert(
+        pid,
+        Distribution::Float(FloatDistribution {
+            low: 0.0,
+            high: f64::INFINITY,
+            log_scale: false,
+            step: None,
+        }),
+    );
+    let err = trial.validate().unwrap_err();
+    assert!(err.contains("non-finite"), "unexpected: {err}");
+}
+
+#[test]
+fn validate_rejects_nan_constraint() {
+    let mut trial = sample_trial(0, 1.0);
+    trial.constraints.push(f64::NAN);
+    let err = trial.validate().unwrap_err();
+    assert!(err.contains("non-finite"), "unexpected: {err}");
+}
+
+#[test]
+fn validate_rejects_non_finite_param_value() {
+    use optimizer::param::ParamValue;
+
+    let pid = FloatParam::new(0.0, 1.0).id();
+    let mut trial = sample_trial(0, 1.0);
+    trial
+        .params
+        .insert(pid, ParamValue::Float(f64::NEG_INFINITY));
+    let err = trial.validate().unwrap_err();
+    assert!(err.contains("non-finite"), "unexpected: {err}");
+}
+
+#[test]
+fn validate_rejects_nan_intermediate_value() {
+    let mut trial = sample_trial(0, 1.0);
+    trial.intermediate_values.push((0, f64::NAN));
+    let err = trial.validate().unwrap_err();
+    assert!(err.contains("non-finite"), "unexpected: {err}");
+}
+
+#[test]
+fn validate_accepts_valid_trial() {
+    use optimizer::distribution::{Distribution, FloatDistribution};
+    use optimizer::param::ParamValue;
+
+    let pid = FloatParam::new(0.0, 1.0).id();
+    let mut trial = sample_trial(0, 1.0);
+    trial.params.insert(pid, ParamValue::Float(0.5));
+    trial.distributions.insert(
+        pid,
+        Distribution::Float(FloatDistribution {
+            low: 0.0,
+            high: 1.0,
+            log_scale: false,
+            step: None,
+        }),
+    );
+    trial.constraints.push(-1.0);
+    trial.intermediate_values.push((0, 0.5));
+    assert!(trial.validate().is_ok());
+}
+
+#[test]
+fn accepts_valid_journal_with_distributions() {
+    let path = temp_path();
+    std::fs::write(
+        &path,
+        r#"{"id":0,"params":{"0":{"Float":0.5}},"distributions":{"0":{"Float":{"low":0.0,"high":1.0,"log_scale":false,"step":null}}},"param_labels":{},"value":0.25,"intermediate_values":[],"state":"Complete","user_attrs":{},"constraints":[-1.0]}"#,
+    )
+    .unwrap();
+
+    let storage = JournalStorage::<f64>::open(&path).unwrap();
+    let loaded = storage.trials_arc().read().clone();
+    assert_eq!(loaded.len(), 1);
+    assert_eq!(loaded[0].value, 0.25);
+
+    std::fs::remove_file(&path).ok();
+}
+
+#[test]
+fn refresh_skips_own_writes() {
+    let path = temp_path();
+    let storage = JournalStorage::new(&path);
+
+    for i in 0..5 {
+        storage.push(sample_trial(i, i as f64));
+        // Our own push advanced the offset, so refresh should find nothing new.
+        assert!(!storage.refresh(), "refresh returned true after push {i}");
+    }
+
+    assert_eq!(storage.trials_arc().read().len(), 5);
+    std::fs::remove_file(&path).ok();
+}
+
+#[test]
+fn refresh_picks_up_external_writes() {
+    let path = temp_path();
+    let storage = JournalStorage::new(&path);
+
+    // Push 3 trials through the storage (advances offset).
+    for i in 0..3 {
+        storage.push(sample_trial(i, i as f64));
+    }
+    assert_eq!(storage.trials_arc().read().len(), 3);
+
+    // Simulate an external process appending 2 more lines directly.
+    {
+        let mut file = std::fs::OpenOptions::new()
+            .append(true)
+            .open(&path)
+            .unwrap();
+        for i in 3..5u64 {
+            let trial = sample_trial(i, i as f64);
+            let line = serde_json::to_string(&trial).unwrap();
+            writeln!(file, "{line}").unwrap();
+        }
+        file.sync_all().unwrap();
+    }
+
+    // refresh() should pick up the 2 external trials.
+    assert!(storage.refresh(), "refresh should detect external writes");
+    assert_eq!(storage.trials_arc().read().len(), 5);
+
+    // A second refresh should be a no-op.
+    assert!(
+        !storage.refresh(),
+        "second refresh should return false (no new data)"
+    );
+    assert_eq!(storage.trials_arc().read().len(), 5);
+
+    std::fs::remove_file(&path).ok();
+}
+
+// ── Corrupted / malicious journal file tests ────────────────────────
+
+fn valid_trial_line_with_id(id: u64) -> String {
+    format!(
+        r#"{{"id":{id},"params":{{}},"distributions":{{}},"param_labels":{{}},"value":1.0,"intermediate_values":[],"state":"Complete","user_attrs":{{}},"constraints":[]}}"#
+    )
+}
+
+#[test]
+fn empty_file_loads_as_empty_storage() {
+    let path = temp_path();
+    std::fs::write(&path, "").unwrap();
+
+    let storage = JournalStorage::<f64>::open(&path).unwrap();
+    assert_eq!(storage.trials_arc().read().len(), 0);
+
+    std::fs::remove_file(&path).ok();
+}
+
+#[test]
+fn whitespace_only_lines_are_skipped() {
+    let path = temp_path();
+    std::fs::write(&path, "  \n\t\n\n").unwrap();
+
+    let storage = JournalStorage::<f64>::open(&path).unwrap();
+    assert_eq!(storage.trials_arc().read().len(), 0);
+
+    std::fs::remove_file(&path).ok();
+}
+
+#[test]
+fn truncated_json_line_returns_error() {
+    let path = temp_path();
+    std::fs::write(&path, r#"{"id":0,"params":{"#).unwrap();
+
+    assert!(JournalStorage::<f64>::open(&path).is_err());
+
+    std::fs::remove_file(&path).ok();
+}
+
+#[test]
+fn invalid_json_syntax_returns_error() {
+    let path = temp_path();
+    std::fs::write(&path, "not valid json\n").unwrap();
+
+    assert!(JournalStorage::<f64>::open(&path).is_err());
+
+    std::fs::remove_file(&path).ok();
+}
+
+#[test]
+fn missing_required_field_returns_error() {
+    let path = temp_path();
+    // Missing params, distributions, param_labels, etc.
+    std::fs::write(&path, r#"{"id":0,"value":1.0}"#).unwrap();
+
+    assert!(JournalStorage::<f64>::open(&path).is_err());
+
+    std::fs::remove_file(&path).ok();
+}
+
+#[test]
+fn extra_fields_are_ignored() {
+    let path = temp_path();
+    let line = r#"{"id":0,"params":{},"distributions":{},"param_labels":{},"value":0.5,"intermediate_values":[],"state":"Complete","user_attrs":{},"constraints":[],"foo":"bar","extra_number":42}"#;
+    std::fs::write(&path, format!("{line}\n")).unwrap();
+
+    let storage = JournalStorage::<f64>::open(&path).unwrap();
+    let loaded = storage.trials_arc().read().clone();
+    assert_eq!(loaded.len(), 1);
+    assert_eq!(loaded[0].id, 0);
+    assert_eq!(loaded[0].value, 0.5);
+
+    std::fs::remove_file(&path).ok();
+}
+
+#[test]
+fn out_of_bounds_categorical_index_loads() {
+    let path = temp_path();
+    // Categorical param with index 999, but distribution only has 3 choices.
+    // validate() does not check categorical bounds, so this should load.
+    let line = r#"{"id":0,"params":{"0":{"Categorical":999}},"distributions":{"0":{"Categorical":{"n_choices":3}}},"param_labels":{},"value":1.0,"intermediate_values":[],"state":"Complete","user_attrs":{},"constraints":[]}"#;
+    std::fs::write(&path, format!("{line}\n")).unwrap();
+
+    let storage = JournalStorage::<f64>::open(&path).unwrap();
+    let loaded = storage.trials_arc().read().clone();
+    assert_eq!(loaded.len(), 1);
+
+    std::fs::remove_file(&path).ok();
+}
+
+#[test]
+fn valid_lines_before_corruption_are_not_loaded() {
+    let path = temp_path();
+    let content = format!(
+        "{}\n{}\n{}\n",
+        valid_trial_line_with_id(0),
+        valid_trial_line_with_id(1),
+        "CORRUPTED LINE"
+    );
+    std::fs::write(&path, content).unwrap();
+
+    // load_trials_from_file is all-or-nothing: the corrupted third line
+    // makes the entire open() fail.
+    assert!(JournalStorage::<f64>::open(&path).is_err());
+
+    std::fs::remove_file(&path).ok();
+}
+
+#[test]
+fn refresh_rejects_corrupted_external_append() {
+    let path = temp_path();
+    let storage = JournalStorage::new(&path);
+
+    // Push 2 valid trials through the storage API.
+    storage.push(sample_trial(0, 1.0));
+    storage.push(sample_trial(1, 2.0));
+    assert_eq!(storage.trials_arc().read().len(), 2);
+
+    // Simulate an external process appending corrupted JSON.
+    {
+        let mut file = std::fs::OpenOptions::new()
+            .append(true)
+            .open(&path)
+            .unwrap();
+        writeln!(file, "CORRUPTED LINE").unwrap();
+        file.sync_all().unwrap();
+    }
+
+    // refresh() should reject the corrupted data and return false.
+    assert!(!storage.refresh());
+    // Memory still has only the original 2 trials.
+    assert_eq!(storage.trials_arc().read().len(), 2);
+
+    std::fs::remove_file(&path).ok();
+}
+
+#[test]
+fn refresh_rejects_truncated_external_append() {
+    let path = temp_path();
+    let storage = JournalStorage::new(&path);
+
+    // Push 2 valid trials through the storage API.
+    storage.push(sample_trial(0, 1.0));
+    storage.push(sample_trial(1, 2.0));
+    assert_eq!(storage.trials_arc().read().len(), 2);
+
+    // Simulate an external process appending truncated JSON.
+    {
+        let mut file = std::fs::OpenOptions::new()
+            .append(true)
+            .open(&path)
+            .unwrap();
+        writeln!(file, r#"{{"id":2,"params":{{"#).unwrap();
+        file.sync_all().unwrap();
+    }
+
+    // refresh() should reject the truncated data and return false.
+    assert!(!storage.refresh());
+    // Memory still has only the original 2 trials.
+    assert_eq!(storage.trials_arc().read().len(), 2);
+
+    std::fs::remove_file(&path).ok();
+}

tests/multi_objective_tests.rs

@@ -0,0 +1,739 @@
+//! Integration tests for multi-objective optimization.
+
+use optimizer::Direction;
+use optimizer::multi_objective::MultiObjectiveStudy;
+use optimizer::parameter::{CategoricalParam, FloatParam, Parameter};
+use optimizer::sampler::Decomposition;
+use optimizer::sampler::moead::MoeadSampler;
+use optimizer::sampler::nsga2::Nsga2Sampler;
+use optimizer::sampler::nsga3::Nsga3Sampler;
+
+// ---------------------------------------------------------------------------
+// Pareto utility tests (via public MultiObjectiveStudy)
+// ---------------------------------------------------------------------------
+
+#[test]
+fn test_basic_two_objective_random() {
+    let study = MultiObjectiveStudy::new(vec![Direction::Minimize, Direction::Minimize]);
+    let x = FloatParam::new(0.0, 1.0);
+
+    study
+        .optimize(30, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            Ok::<_, optimizer::Error>(vec![xv, 1.0 - xv])
+        })
+        .unwrap();
+
+    let front = study.pareto_front();
+    assert!(!front.is_empty(), "Pareto front should be non-empty");
+
+    // Verify no solution in the front dominates another
+    for a in &front {
+        for b in &front {
+            if core::ptr::eq(a, b) {
+                continue;
+            }
+            let a_dom_b = a.values[0] <= b.values[0]
+                && a.values[1] <= b.values[1]
+                && (a.values[0] < b.values[0] || a.values[1] < b.values[1]);
+            assert!(
+                !a_dom_b,
+                "Front solution {:?} dominates {:?}",
+                a.values, b.values
+            );
+        }
+    }
+}
+
+#[test]
+fn test_dimension_mismatch_error() {
+    let study = MultiObjectiveStudy::new(vec![Direction::Minimize, Direction::Minimize]);
+    let x = FloatParam::new(0.0, 1.0);
+
+    let result = study.optimize(1, |trial: &mut optimizer::Trial| {
+        let xv = x.suggest(trial)?;
+        // Return wrong number of values
+        Ok::<_, optimizer::Error>(vec![xv])
+    });
+
+    assert!(result.is_err());
+    let err = result.unwrap_err();
+    assert!(
+        matches!(
+            err,
+            optimizer::Error::ObjectiveDimensionMismatch {
+                expected: 2,
+                got: 1
+            }
+        ),
+        "Expected ObjectiveDimensionMismatch, got: {err}"
+    );
+}
+
+#[test]
+fn test_ask_tell() {
+    let study = MultiObjectiveStudy::new(vec![Direction::Minimize, Direction::Maximize]);
+    let x = FloatParam::new(0.0, 10.0);
+
+    for _ in 0..10 {
+        let mut trial = study.ask();
+        let xv = x.suggest(&mut trial).unwrap();
+        study
+            .tell(trial, Ok::<_, &str>(vec![xv, 10.0 - xv]))
+            .unwrap();
+    }
+
+    assert_eq!(study.n_trials(), 10);
+    let front = study.pareto_front();
+    assert!(!front.is_empty());
+}
+
+#[test]
+fn test_ask_tell_dimension_mismatch() {
+    let study = MultiObjectiveStudy::new(vec![Direction::Minimize, Direction::Minimize]);
+    let trial = study.ask();
+    let result = study.tell(trial, Ok::<_, &str>(vec![1.0, 2.0, 3.0]));
+    assert!(result.is_err());
+}
+
+#[test]
+fn test_n_trials_counting() {
+    let study = MultiObjectiveStudy::new(vec![Direction::Minimize, Direction::Minimize]);
+    assert_eq!(study.n_trials(), 0);
+
+    let x = FloatParam::new(0.0, 1.0);
+
+    study
+        .optimize(5, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            Ok::<_, optimizer::Error>(vec![xv, 1.0 - xv])
+        })
+        .unwrap();
+
+    assert_eq!(study.n_trials(), 5);
+}
+
+#[test]
+fn test_three_objectives() {
+    let study = MultiObjectiveStudy::new(vec![
+        Direction::Minimize,
+        Direction::Minimize,
+        Direction::Maximize,
+    ]);
+    let x = FloatParam::new(0.0, 1.0);
+    let y = FloatParam::new(0.0, 1.0);
+
+    study
+        .optimize(30, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            let yv = y.suggest(trial)?;
+            Ok::<_, optimizer::Error>(vec![xv, yv, 1.0 - xv - yv])
+        })
+        .unwrap();
+
+    let front = study.pareto_front();
+    assert!(!front.is_empty());
+    assert_eq!(study.n_objectives(), 3);
+}
+
+#[test]
+fn test_directions_accessor() {
+    let dirs = vec![Direction::Minimize, Direction::Maximize];
+    let study = MultiObjectiveStudy::new(dirs.clone());
+    assert_eq!(study.directions(), &dirs);
+    assert_eq!(study.n_objectives(), 2);
+}
+
+#[test]
+fn test_trials_accessor() {
+    let study = MultiObjectiveStudy::new(vec![Direction::Minimize, Direction::Minimize]);
+    let x = FloatParam::new(0.0, 1.0);
+
+    study
+        .optimize(3, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            Ok::<_, optimizer::Error>(vec![xv, 1.0 - xv])
+        })
+        .unwrap();
+
+    let trials = study.trials();
+    assert_eq!(trials.len(), 3);
+    for t in &trials {
+        assert_eq!(t.values.len(), 2);
+    }
+}
+
+// ---------------------------------------------------------------------------
+// NSGA-II sampler tests
+// ---------------------------------------------------------------------------
+
+#[test]
+fn test_nsga2_zdt1() {
+    // ZDT1 benchmark: minimize both objectives
+    let n_vars = 5;
+    let params: Vec<FloatParam> = (0..n_vars).map(|_| FloatParam::new(0.0, 1.0)).collect();
+
+    let sampler = Nsga2Sampler::builder().population_size(20).seed(42).build();
+    let study =
+        MultiObjectiveStudy::with_sampler(vec![Direction::Minimize, Direction::Minimize], sampler);
+
+    study
+        .optimize(200, |trial: &mut optimizer::Trial| {
+            let xs: Vec<f64> = params
+                .iter()
+                .map(|p| p.suggest(trial))
+                .collect::<Result<_, _>>()?;
+
+            let f1 = xs[0];
+            let g = 1.0 + 9.0 * xs[1..].iter().sum::<f64>() / (n_vars - 1) as f64;
+            let f2 = g * (1.0 - (f1 / g).sqrt());
+            Ok::<_, optimizer::Error>(vec![f1, f2])
+        })
+        .unwrap();
+
+    let front = study.pareto_front();
+    assert!(!front.is_empty(), "Pareto front should be non-empty");
+
+    // Verify no dominated solutions in the front
+    for a in &front {
+        for b in &front {
+            if core::ptr::eq(a, b) {
+                continue;
+            }
+            let a_dom_b = a.values[0] <= b.values[0]
+                && a.values[1] <= b.values[1]
+                && (a.values[0] < b.values[0] || a.values[1] < b.values[1]);
+            assert!(
+                !a_dom_b,
+                "Front solution {:?} dominates {:?}",
+                a.values, b.values
+            );
+        }
+    }
+}
+
+#[test]
+fn test_nsga2_with_seed_reproducible() {
+    let x = FloatParam::new(0.0, 1.0);
+    let y = FloatParam::new(0.0, 1.0);
+
+    let run = |seed: u64| -> Vec<Vec<f64>> {
+        let sampler = Nsga2Sampler::with_seed(seed);
+        let study = MultiObjectiveStudy::with_sampler(
+            vec![Direction::Minimize, Direction::Minimize],
+            sampler,
+        );
+        study
+            .optimize(30, |trial: &mut optimizer::Trial| {
+                let xv = x.suggest(trial)?;
+                let yv = y.suggest(trial)?;
+                Ok::<_, optimizer::Error>(vec![xv, yv])
+            })
+            .unwrap();
+        study.trials().iter().map(|t| t.values.clone()).collect()
+    };
+
+    let r1 = run(123);
+    let r2 = run(123);
+    assert_eq!(r1, r2, "Same seed should produce same results");
+
+    let r3 = run(456);
+    assert_ne!(r1, r3, "Different seeds should produce different results");
+}
+
+#[test]
+fn test_nsga2_builder() {
+    let sampler = Nsga2Sampler::builder()
+        .population_size(10)
+        .crossover_prob(0.8)
+        .crossover_eta(15.0)
+        .mutation_eta(25.0)
+        .seed(42)
+        .build();
+
+    let study =
+        MultiObjectiveStudy::with_sampler(vec![Direction::Minimize, Direction::Minimize], sampler);
+    let x = FloatParam::new(0.0, 1.0);
+
+    study
+        .optimize(30, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            Ok::<_, optimizer::Error>(vec![xv, 1.0 - xv])
+        })
+        .unwrap();
+
+    assert_eq!(study.n_trials(), 30);
+}
+
+#[test]
+fn test_nsga2_categorical_params() {
+    let sampler = Nsga2Sampler::with_seed(42);
+    let study =
+        MultiObjectiveStudy::with_sampler(vec![Direction::Minimize, Direction::Minimize], sampler);
+
+    let x = FloatParam::new(0.0, 1.0);
+    let cat = CategoricalParam::new(vec!["a", "b", "c"]);
+
+    study
+        .optimize(30, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            let cv = cat.suggest(trial)?;
+            let bonus = match cv {
+                "a" => 0.0,
+                "b" => 0.5,
+                _ => 1.0,
+            };
+            Ok::<_, optimizer::Error>(vec![xv + bonus, 1.0 - xv])
+        })
+        .unwrap();
+
+    assert_eq!(study.n_trials(), 30);
+    let front = study.pareto_front();
+    assert!(!front.is_empty());
+}
+
+#[test]
+fn test_nsga2_constraints() {
+    let sampler = Nsga2Sampler::with_seed(42);
+    let study =
+        MultiObjectiveStudy::with_sampler(vec![Direction::Minimize, Direction::Minimize], sampler);
+
+    let x = FloatParam::new(0.0, 1.0);
+
+    study
+        .optimize(50, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            // Constraint: x >= 0.3 (i.e. 0.3 - x <= 0)
+            trial.set_constraints(vec![0.3 - xv]);
+            Ok::<_, optimizer::Error>(vec![xv, 1.0 - xv])
+        })
+        .unwrap();
+
+    let front = study.pareto_front();
+    assert!(!front.is_empty());
+
+    // Check that feasible solutions exist on the front
+    let feasible_count = front.iter().filter(|t| t.is_feasible()).count();
+    assert!(
+        feasible_count > 0,
+        "Should have feasible solutions on front"
+    );
+}
+
+#[test]
+fn test_multi_objective_trial_get() {
+    let study = MultiObjectiveStudy::new(vec![Direction::Minimize, Direction::Minimize]);
+    let x = FloatParam::new(0.0, 10.0).name("x");
+
+    study
+        .optimize(5, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            Ok::<_, optimizer::Error>(vec![xv, 10.0 - xv])
+        })
+        .unwrap();
+
+    let front = study.pareto_front();
+    for t in &front {
+        let xv: f64 = t.get(&x).unwrap();
+        assert!((0.0..=10.0).contains(&xv));
+    }
+}
+
+#[test]
+fn test_multi_objective_trial_is_feasible() {
+    let study = MultiObjectiveStudy::new(vec![Direction::Minimize, Direction::Minimize]);
+    let x = FloatParam::new(0.0, 1.0);
+
+    study
+        .optimize(10, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            trial.set_constraints(vec![0.5 - xv]); // feasible if x >= 0.5
+            Ok::<_, optimizer::Error>(vec![xv, 1.0 - xv])
+        })
+        .unwrap();
+
+    let trials = study.trials();
+    for t in &trials {
+        let xv = t.values[0];
+        if xv >= 0.5 {
+            assert!(t.is_feasible());
+        } else {
+            assert!(!t.is_feasible());
+        }
+    }
+}
+
+#[test]
+fn test_multi_objective_trial_user_attrs() {
+    let study = MultiObjectiveStudy::new(vec![Direction::Minimize, Direction::Minimize]);
+    let x = FloatParam::new(0.0, 1.0);
+
+    study
+        .optimize(3, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            trial.set_user_attr("iteration", 42_i64);
+            Ok::<_, optimizer::Error>(vec![xv, 1.0 - xv])
+        })
+        .unwrap();
+
+    let trials = study.trials();
+    for t in &trials {
+        assert!(t.user_attr("iteration").is_some());
+    }
+}
+
+#[test]
+fn test_tell_with_failure() {
+    let study = MultiObjectiveStudy::new(vec![Direction::Minimize, Direction::Minimize]);
+
+    let trial = study.ask();
+    study
+        .tell(trial, Err::<Vec<f64>, _>("evaluation failed"))
+        .unwrap();
+
+    // Failed trial not counted
+    assert_eq!(study.n_trials(), 0);
+}
+
+// ---------------------------------------------------------------------------
+// NSGA-III sampler tests
+// ---------------------------------------------------------------------------
+
+#[test]
+fn test_nsga3_zdt1() {
+    let n_vars = 5;
+    let params: Vec<FloatParam> = (0..n_vars).map(|_| FloatParam::new(0.0, 1.0)).collect();
+
+    let sampler = Nsga3Sampler::builder().population_size(20).seed(42).build();
+    let study =
+        MultiObjectiveStudy::with_sampler(vec![Direction::Minimize, Direction::Minimize], sampler);
+
+    study
+        .optimize(200, |trial: &mut optimizer::Trial| {
+            let xs: Vec<f64> = params
+                .iter()
+                .map(|p| p.suggest(trial))
+                .collect::<Result<_, _>>()?;
+
+            let f1 = xs[0];
+            let g = 1.0 + 9.0 * xs[1..].iter().sum::<f64>() / (n_vars - 1) as f64;
+            let f2 = g * (1.0 - (f1 / g).sqrt());
+            Ok::<_, optimizer::Error>(vec![f1, f2])
+        })
+        .unwrap();
+
+    let front = study.pareto_front();
+    assert!(
+        !front.is_empty(),
+        "NSGA-III Pareto front should be non-empty"
+    );
+
+    // Verify no dominated solutions in the front
+    for a in &front {
+        for b in &front {
+            if core::ptr::eq(a, b) {
+                continue;
+            }
+            let a_dom_b = a.values[0] <= b.values[0]
+                && a.values[1] <= b.values[1]
+                && (a.values[0] < b.values[0] || a.values[1] < b.values[1]);
+            assert!(
+                !a_dom_b,
+                "Front solution {:?} dominates {:?}",
+                a.values, b.values
+            );
+        }
+    }
+}
+
+#[test]
+fn test_nsga3_four_objectives() {
+    // DTLZ2 with 4 objectives
+    let n_obj = 4;
+    let n_vars = n_obj + 4; // k = 5 decision variables beyond the first (n_obj-1)
+    let params: Vec<FloatParam> = (0..n_vars).map(|_| FloatParam::new(0.0, 1.0)).collect();
+
+    let sampler = Nsga3Sampler::builder().population_size(50).seed(42).build();
+    let directions = vec![Direction::Minimize; n_obj];
+    let study = MultiObjectiveStudy::with_sampler(directions, sampler);
+
+    study
+        .optimize(500, |trial: &mut optimizer::Trial| {
+            let xs: Vec<f64> = params
+                .iter()
+                .map(|p| p.suggest(trial))
+                .collect::<Result<_, _>>()?;
+
+            // DTLZ2 formulation
+            let g: f64 = xs[n_obj - 1..]
+                .iter()
+                .map(|&xi| (xi - 0.5).powi(2))
+                .sum::<f64>();
+
+            let mut objectives = vec![0.0_f64; n_obj];
+            for i in 0..n_obj {
+                let mut f = 1.0 + g;
+                for xj in &xs[..(n_obj - 1 - i)] {
+                    f *= (xj * core::f64::consts::FRAC_PI_2).cos();
+                }
+                if i > 0 {
+                    f *= (xs[n_obj - 1 - i] * core::f64::consts::FRAC_PI_2).sin();
+                }
+                objectives[i] = f;
+            }
+
+            Ok::<_, optimizer::Error>(objectives)
+        })
+        .unwrap();
+
+    let front = study.pareto_front();
+    assert!(!front.is_empty(), "4-objective front should be non-empty");
+    // All front solutions should have 4 objectives
+    for t in &front {
+        assert_eq!(t.values.len(), 4);
+    }
+}
+
+#[test]
+fn test_nsga3_reproducible() {
+    let x = FloatParam::new(0.0, 1.0);
+    let y = FloatParam::new(0.0, 1.0);
+
+    let run = |seed: u64| -> Vec<Vec<f64>> {
+        let sampler = Nsga3Sampler::with_seed(seed);
+        let study = MultiObjectiveStudy::with_sampler(
+            vec![Direction::Minimize, Direction::Minimize],
+            sampler,
+        );
+        study
+            .optimize(30, |trial: &mut optimizer::Trial| {
+                let xv = x.suggest(trial)?;
+                let yv = y.suggest(trial)?;
+                Ok::<_, optimizer::Error>(vec![xv, yv])
+            })
+            .unwrap();
+        study.trials().iter().map(|t| t.values.clone()).collect()
+    };
+
+    let r1 = run(123);
+    let r2 = run(123);
+    assert_eq!(r1, r2, "Same seed should produce same results");
+
+    let r3 = run(456);
+    assert_ne!(r1, r3, "Different seeds should produce different results");
+}
+
+#[test]
+fn test_nsga3_builder() {
+    let sampler = Nsga3Sampler::builder()
+        .population_size(12)
+        .n_divisions(4)
+        .crossover_prob(0.9)
+        .crossover_eta(20.0)
+        .mutation_eta(20.0)
+        .seed(42)
+        .build();
+
+    let study =
+        MultiObjectiveStudy::with_sampler(vec![Direction::Minimize, Direction::Minimize], sampler);
+    let x = FloatParam::new(0.0, 1.0);
+
+    study
+        .optimize(30, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            Ok::<_, optimizer::Error>(vec![xv, 1.0 - xv])
+        })
+        .unwrap();
+
+    assert_eq!(study.n_trials(), 30);
+}
+
+#[test]
+fn test_nsga3_constraints() {
+    let sampler = Nsga3Sampler::with_seed(42);
+    let study =
+        MultiObjectiveStudy::with_sampler(vec![Direction::Minimize, Direction::Minimize], sampler);
+
+    let x = FloatParam::new(0.0, 1.0);
+
+    study
+        .optimize(50, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            trial.set_constraints(vec![0.3 - xv]);
+            Ok::<_, optimizer::Error>(vec![xv, 1.0 - xv])
+        })
+        .unwrap();
+
+    let front = study.pareto_front();
+    assert!(!front.is_empty());
+
+    let feasible_count = front.iter().filter(|t| t.is_feasible()).count();
+    assert!(
+        feasible_count > 0,
+        "Should have feasible solutions on front"
+    );
+}
+
+// ---------------------------------------------------------------------------
+// MOEA/D sampler tests
+// ---------------------------------------------------------------------------
+
+#[test]
+fn test_moead_zdt1_tchebycheff() {
+    let n_vars = 5;
+    let params: Vec<FloatParam> = (0..n_vars).map(|_| FloatParam::new(0.0, 1.0)).collect();
+
+    let sampler = MoeadSampler::builder().population_size(20).seed(42).build();
+    let study =
+        MultiObjectiveStudy::with_sampler(vec![Direction::Minimize, Direction::Minimize], sampler);
+
+    study
+        .optimize(200, |trial: &mut optimizer::Trial| {
+            let xs: Vec<f64> = params
+                .iter()
+                .map(|p| p.suggest(trial))
+                .collect::<Result<_, _>>()?;
+
+            let f1 = xs[0];
+            let g = 1.0 + 9.0 * xs[1..].iter().sum::<f64>() / (n_vars - 1) as f64;
+            let f2 = g * (1.0 - (f1 / g).sqrt());
+            Ok::<_, optimizer::Error>(vec![f1, f2])
+        })
+        .unwrap();
+
+    let front = study.pareto_front();
+    assert!(!front.is_empty(), "MOEA/D Pareto front should be non-empty");
+
+    for a in &front {
+        for b in &front {
+            if core::ptr::eq(a, b) {
+                continue;
+            }
+            let a_dom_b = a.values[0] <= b.values[0]
+                && a.values[1] <= b.values[1]
+                && (a.values[0] < b.values[0] || a.values[1] < b.values[1]);
+            assert!(
+                !a_dom_b,
+                "Front solution {:?} dominates {:?}",
+                a.values, b.values
+            );
+        }
+    }
+}
+
+#[test]
+fn test_moead_zdt1_weighted_sum() {
+    let n_vars = 3;
+    let params: Vec<FloatParam> = (0..n_vars).map(|_| FloatParam::new(0.0, 1.0)).collect();
+
+    let sampler = MoeadSampler::builder()
+        .population_size(20)
+        .decomposition(Decomposition::WeightedSum)
+        .seed(42)
+        .build();
+    let study =
+        MultiObjectiveStudy::with_sampler(vec![Direction::Minimize, Direction::Minimize], sampler);
+
+    study
+        .optimize(200, |trial: &mut optimizer::Trial| {
+            let xs: Vec<f64> = params
+                .iter()
+                .map(|p| p.suggest(trial))
+                .collect::<Result<_, _>>()?;
+
+            let f1 = xs[0];
+            let g = 1.0 + 9.0 * xs[1..].iter().sum::<f64>() / (n_vars - 1) as f64;
+            let f2 = g * (1.0 - (f1 / g).sqrt());
+            Ok::<_, optimizer::Error>(vec![f1, f2])
+        })
+        .unwrap();
+
+    let front = study.pareto_front();
+    assert!(!front.is_empty());
+}
+
+#[test]
+fn test_moead_zdt1_pbi() {
+    let n_vars = 3;
+    let params: Vec<FloatParam> = (0..n_vars).map(|_| FloatParam::new(0.0, 1.0)).collect();
+
+    let sampler = MoeadSampler::builder()
+        .population_size(20)
+        .decomposition(Decomposition::Pbi { theta: 5.0 })
+        .seed(42)
+        .build();
+    let study =
+        MultiObjectiveStudy::with_sampler(vec![Direction::Minimize, Direction::Minimize], sampler);
+
+    study
+        .optimize(200, |trial: &mut optimizer::Trial| {
+            let xs: Vec<f64> = params
+                .iter()
+                .map(|p| p.suggest(trial))
+                .collect::<Result<_, _>>()?;
+
+            let f1 = xs[0];
+            let g = 1.0 + 9.0 * xs[1..].iter().sum::<f64>() / (n_vars - 1) as f64;
+            let f2 = g * (1.0 - (f1 / g).sqrt());
+            Ok::<_, optimizer::Error>(vec![f1, f2])
+        })
+        .unwrap();
+
+    let front = study.pareto_front();
+    assert!(!front.is_empty());
+}
+
+#[test]
+fn test_moead_reproducible() {
+    let x = FloatParam::new(0.0, 1.0);
+    let y = FloatParam::new(0.0, 1.0);
+
+    let run = |seed: u64| -> Vec<Vec<f64>> {
+        let sampler = MoeadSampler::with_seed(seed);
+        let study = MultiObjectiveStudy::with_sampler(
+            vec![Direction::Minimize, Direction::Minimize],
+            sampler,
+        );
+        study
+            .optimize(30, |trial: &mut optimizer::Trial| {
+                let xv = x.suggest(trial)?;
+                let yv = y.suggest(trial)?;
+                Ok::<_, optimizer::Error>(vec![xv, yv])
+            })
+            .unwrap();
+        study.trials().iter().map(|t| t.values.clone()).collect()
+    };
+
+    let r1 = run(123);
+    let r2 = run(123);
+    assert_eq!(r1, r2, "Same seed should produce same results");
+
+    let r3 = run(456);
+    assert_ne!(r1, r3, "Different seeds should produce different results");
+}
+
+#[test]
+fn test_moead_builder() {
+    let sampler = MoeadSampler::builder()
+        .population_size(15)
+        .neighborhood_size(5)
+        .decomposition(Decomposition::Tchebycheff)
+        .crossover_prob(0.9)
+        .crossover_eta(20.0)
+        .mutation_eta(20.0)
+        .seed(42)
+        .build();
+
+    let study =
+        MultiObjectiveStudy::with_sampler(vec![Direction::Minimize, Direction::Minimize], sampler);
+    let x = FloatParam::new(0.0, 1.0);
+
+    study
+        .optimize(30, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            Ok::<_, optimizer::Error>(vec![xv, 1.0 - xv])
+        })
+        .unwrap();
+
+    assert_eq!(study.n_trials(), 30);
+}

tests/parameter_tests.rs

@@ -180,7 +180,7 @@ fn parameter_api_with_study() {
 
     let study: Study<f64> = Study::new(Direction::Minimize);
     study
-        .optimize(5, |trial| {
+        .optimize(5, |trial: &mut optimizer::Trial| {
             let x = x_param.suggest(trial)?;
             let n = n_param.suggest(trial)?;
             let dropout = dropout_param.suggest(trial)?;

tests/pruner/main.rs

@@ -0,0 +1,2 @@
+mod median;
+mod threshold;

tests/sampler/bohb.rs

@@ -27,7 +27,7 @@ fn bohb_converges_on_quadratic() {
     let x_param = FloatParam::new(-10.0, 10.0);
 
     study
-        .optimize(60, |trial| {
+        .optimize(60, |trial: &mut optimizer::Trial| {
             let x = x_param.suggest(trial)?;
 
             // Report intermediate values at budget steps 1, 3, 9
@@ -66,7 +66,7 @@ fn bohb_with_pruning() {
     let x_param = FloatParam::new(-5.0, 5.0);
 
     study
-        .optimize(40, |trial| {
+        .optimize(40, |trial: &mut optimizer::Trial| {
             let x = x_param.suggest(trial)?;
             let obj = x * x;
 
@@ -112,7 +112,7 @@ fn bohb_uses_budget_conditioned_history() {
     let x_param = FloatParam::new(0.0, 10.0);
 
     study
-        .optimize(30, |trial| {
+        .optimize(30, |trial: &mut optimizer::Trial| {
             let x = x_param.suggest(trial)?;
             // Intermediate values that guide optimization toward x=2
             trial.report(1, (x - 2.0).powi(2) + 1.0);

tests/sampler/cma_es.rs

@@ -1,5 +1,3 @@
-#![cfg(feature = "cma-es")]
-
 use optimizer::prelude::*;
 use optimizer::sampler::cma_es::CmaEsSampler;
 
@@ -12,7 +10,7 @@ fn sphere_function() {
     let y = FloatParam::new(-5.0, 5.0).name("y");
 
     study
-        .optimize(200, |trial| {
+        .optimize(200, |trial: &mut optimizer::Trial| {
             let xv = x.suggest(trial)?;
             let yv = y.suggest(trial)?;
             Ok::<_, Error>(xv * xv + yv * yv)
@@ -36,7 +34,7 @@ fn rosenbrock_function() {
     let y = FloatParam::new(-5.0, 5.0).name("y");
 
     study
-        .optimize(300, |trial| {
+        .optimize(300, |trial: &mut optimizer::Trial| {
             let xv = x.suggest(trial)?;
             let yv = y.suggest(trial)?;
             let val = (1.0 - xv).powi(2) + 100.0 * (yv - xv * xv).powi(2);
@@ -62,7 +60,7 @@ fn bounds_respected() {
     let y = FloatParam::new(0.0, 10.0).name("y");
 
     study
-        .optimize(100, |trial| {
+        .optimize(100, |trial: &mut optimizer::Trial| {
             let xv = x.suggest(trial)?;
             let yv = y.suggest(trial)?;
             Ok::<_, Error>(xv + yv)
@@ -86,7 +84,7 @@ fn mixed_params_float_and_categorical() {
     let cat = CategoricalParam::new(vec!["a", "b", "c"]).name("cat");
 
     study
-        .optimize(50, |trial| {
+        .optimize(50, |trial: &mut optimizer::Trial| {
             let xv = x.suggest(trial)?;
             let cv = cat.suggest(trial)?;
             let penalty = match cv {
@@ -116,7 +114,7 @@ fn seeded_reproducibility() {
         let sampler = CmaEsSampler::with_seed(seed);
         let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
         study
-            .optimize(50, |trial| {
+            .optimize(50, |trial: &mut optimizer::Trial| {
                 let xv = x.suggest(trial)?;
                 let yv = y.suggest(trial)?;
                 Ok::<_, Error>(xv * xv + yv * yv)
@@ -139,7 +137,7 @@ fn different_seeds_different_results() {
         let sampler = CmaEsSampler::with_seed(seed);
         let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
         study
-            .optimize(20, |trial| {
+            .optimize(20, |trial: &mut optimizer::Trial| {
                 let xv = x.suggest(trial)?;
                 let yv = y.suggest(trial)?;
                 Ok::<_, Error>(xv * xv + yv * yv)
@@ -164,7 +162,7 @@ fn single_dimension() {
     let x = FloatParam::new(-10.0, 10.0).name("x");
 
     study
-        .optimize(100, |trial| {
+        .optimize(100, |trial: &mut optimizer::Trial| {
             let xv = x.suggest(trial)?;
             Ok::<_, Error>((xv - 3.0).powi(2))
         })
@@ -186,7 +184,7 @@ fn integer_params() {
     let n = IntParam::new(1, 20).name("n");
 
     study
-        .optimize(100, |trial| {
+        .optimize(100, |trial: &mut optimizer::Trial| {
             let nv = n.suggest(trial)?;
             // Minimum at n = 10
             Ok::<_, Error>(((nv - 10) * (nv - 10)) as f64)
@@ -214,7 +212,7 @@ fn log_scale_params() {
     let lr = FloatParam::new(1e-5, 1.0).log_scale().name("lr");
 
     study
-        .optimize(100, |trial| {
+        .optimize(100, |trial: &mut optimizer::Trial| {
             let lrv = lr.suggest(trial)?;
             // Minimum at lr = 0.01
             Ok::<_, Error>((lrv.ln() - 0.01_f64.ln()).powi(2))
@@ -243,7 +241,7 @@ fn custom_population_size_and_sigma() {
     let y = FloatParam::new(-5.0, 5.0).name("y");
 
     study
-        .optimize(100, |trial| {
+        .optimize(100, |trial: &mut optimizer::Trial| {
             let xv = x.suggest(trial)?;
             let yv = y.suggest(trial)?;
             Ok::<_, Error>(xv * xv + yv * yv)

tests/sampler/differential_evolution.rs

@@ -0,0 +1,345 @@
+use optimizer::prelude::*;
+use optimizer::sampler::de::{DESampler, DEStrategy};
+
+#[test]
+fn sphere_function() {
+    let sampler = DESampler::with_seed(42);
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+
+    let x = FloatParam::new(-5.0, 5.0).name("x");
+    let y = FloatParam::new(-5.0, 5.0).name("y");
+
+    study
+        .optimize(200, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            let yv = y.suggest(trial)?;
+            Ok::<_, Error>(xv * xv + yv * yv)
+        })
+        .unwrap();
+
+    let best = study.best_trial().unwrap();
+    assert!(
+        best.value < 1.0,
+        "sphere best value should be < 1.0, got {}",
+        best.value
+    );
+}
+
+#[test]
+fn rosenbrock_function() {
+    let sampler = DESampler::builder().population_size(20).seed(42).build();
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+
+    let x = FloatParam::new(-5.0, 5.0).name("x");
+    let y = FloatParam::new(-5.0, 5.0).name("y");
+
+    study
+        .optimize(400, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            let yv = y.suggest(trial)?;
+            let val = (1.0 - xv).powi(2) + 100.0 * (yv - xv * xv).powi(2);
+            Ok::<_, Error>(val)
+        })
+        .unwrap();
+
+    let best = study.best_trial().unwrap();
+    assert!(
+        best.value < 50.0,
+        "rosenbrock best value should be < 50.0, got {}",
+        best.value
+    );
+}
+
+#[test]
+fn rastrigin_function() {
+    let sampler = DESampler::builder()
+        .population_size(30)
+        .mutation_factor(0.7)
+        .crossover_rate(0.9)
+        .seed(42)
+        .build();
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+
+    let x = FloatParam::new(-5.12, 5.12).name("x");
+    let y = FloatParam::new(-5.12, 5.12).name("y");
+
+    study
+        .optimize(500, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            let yv = y.suggest(trial)?;
+            let val = 20.0
+                + (xv * xv - 10.0 * (2.0 * std::f64::consts::PI * xv).cos())
+                + (yv * yv - 10.0 * (2.0 * std::f64::consts::PI * yv).cos());
+            Ok::<_, Error>(val)
+        })
+        .unwrap();
+
+    let best = study.best_trial().unwrap();
+    // Rastrigin is multimodal; just check reasonable convergence
+    assert!(
+        best.value < 10.0,
+        "rastrigin best value should be < 10.0, got {}",
+        best.value
+    );
+}
+
+#[test]
+fn bounds_respected() {
+    let sampler = DESampler::with_seed(123);
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+
+    let x = FloatParam::new(-2.0, 3.0).name("x");
+    let y = FloatParam::new(0.0, 10.0).name("y");
+
+    study
+        .optimize(100, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            let yv = y.suggest(trial)?;
+            Ok::<_, Error>(xv + yv)
+        })
+        .unwrap();
+
+    for trial in study.trials() {
+        let xv: f64 = trial.get(&x).unwrap();
+        let yv: f64 = trial.get(&y).unwrap();
+        assert!((-2.0..=3.0).contains(&xv), "x = {xv} out of bounds [-2, 3]");
+        assert!((0.0..=10.0).contains(&yv), "y = {yv} out of bounds [0, 10]");
+    }
+}
+
+#[test]
+fn strategy_best1() {
+    let sampler = DESampler::builder()
+        .strategy(DEStrategy::Best1)
+        .population_size(15)
+        .seed(42)
+        .build();
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+
+    let x = FloatParam::new(-5.0, 5.0).name("x");
+    let y = FloatParam::new(-5.0, 5.0).name("y");
+
+    study
+        .optimize(200, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            let yv = y.suggest(trial)?;
+            Ok::<_, Error>(xv * xv + yv * yv)
+        })
+        .unwrap();
+
+    let best = study.best_trial().unwrap();
+    assert!(
+        best.value < 5.0,
+        "Best1 strategy should converge, got {}",
+        best.value
+    );
+}
+
+#[test]
+fn strategy_current_to_best1() {
+    let sampler = DESampler::builder()
+        .strategy(DEStrategy::CurrentToBest1)
+        .population_size(15)
+        .seed(42)
+        .build();
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+
+    let x = FloatParam::new(-5.0, 5.0).name("x");
+    let y = FloatParam::new(-5.0, 5.0).name("y");
+
+    study
+        .optimize(200, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            let yv = y.suggest(trial)?;
+            Ok::<_, Error>(xv * xv + yv * yv)
+        })
+        .unwrap();
+
+    let best = study.best_trial().unwrap();
+    assert!(
+        best.value < 5.0,
+        "CurrentToBest1 strategy should converge, got {}",
+        best.value
+    );
+}
+
+#[test]
+fn mixed_params_float_and_categorical() {
+    let sampler = DESampler::with_seed(42);
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+
+    let x = FloatParam::new(-5.0, 5.0).name("x");
+    let cat = CategoricalParam::new(vec!["a", "b", "c"]).name("cat");
+
+    study
+        .optimize(100, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            let cv = cat.suggest(trial)?;
+            let penalty = match cv {
+                "a" => 0.0,
+                "b" => 1.0,
+                _ => 2.0,
+            };
+            Ok::<_, Error>(xv * xv + penalty)
+        })
+        .unwrap();
+
+    let best = study.best_trial().unwrap();
+    assert!(
+        best.value < 10.0,
+        "best value should be < 10.0, got {}",
+        best.value
+    );
+}
+
+#[test]
+fn seeded_reproducibility() {
+    let x = FloatParam::new(-5.0, 5.0).name("x");
+    let y = FloatParam::new(-5.0, 5.0).name("y");
+
+    let run = |seed: u64| {
+        let sampler = DESampler::with_seed(seed);
+        let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+        study
+            .optimize(50, |trial: &mut optimizer::Trial| {
+                let xv = x.suggest(trial)?;
+                let yv = y.suggest(trial)?;
+                Ok::<_, Error>(xv * xv + yv * yv)
+            })
+            .unwrap();
+        study.trials().iter().map(|t| t.value).collect::<Vec<_>>()
+    };
+
+    let results1 = run(42);
+    let results2 = run(42);
+    assert_eq!(results1, results2, "same seed should produce same results");
+}
+
+#[test]
+fn different_seeds_different_results() {
+    let x = FloatParam::new(-5.0, 5.0).name("x");
+    let y = FloatParam::new(-5.0, 5.0).name("y");
+
+    let run = |seed: u64| {
+        let sampler = DESampler::with_seed(seed);
+        let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+        study
+            .optimize(20, |trial: &mut optimizer::Trial| {
+                let xv = x.suggest(trial)?;
+                let yv = y.suggest(trial)?;
+                Ok::<_, Error>(xv * xv + yv * yv)
+            })
+            .unwrap();
+        study.trials().iter().map(|t| t.value).collect::<Vec<_>>()
+    };
+
+    let results1 = run(42);
+    let results2 = run(99);
+    assert_ne!(
+        results1, results2,
+        "different seeds should produce different results"
+    );
+}
+
+#[test]
+fn single_dimension() {
+    let sampler = DESampler::with_seed(42);
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+
+    let x = FloatParam::new(-10.0, 10.0).name("x");
+
+    study
+        .optimize(100, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            Ok::<_, Error>((xv - 3.0).powi(2))
+        })
+        .unwrap();
+
+    let best = study.best_trial().unwrap();
+    assert!(
+        best.value < 1.0,
+        "1-D optimization should converge, got {}",
+        best.value
+    );
+}
+
+#[test]
+fn integer_params() {
+    let sampler = DESampler::with_seed(42);
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+
+    let n = IntParam::new(1, 20).name("n");
+
+    study
+        .optimize(100, |trial: &mut optimizer::Trial| {
+            let nv = n.suggest(trial)?;
+            // Minimum at n = 10
+            Ok::<_, Error>(((nv - 10) * (nv - 10)) as f64)
+        })
+        .unwrap();
+
+    let best = study.best_trial().unwrap();
+    let best_n: i64 = best.get(&n).unwrap();
+    assert!(
+        (1..=20).contains(&best_n),
+        "integer value {best_n} out of bounds"
+    );
+    assert!(
+        best.value < 10.0,
+        "integer optimization should converge, got {}",
+        best.value
+    );
+}
+
+#[test]
+fn log_scale_params() {
+    let sampler = DESampler::with_seed(42);
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+
+    let lr = FloatParam::new(1e-5, 1.0).log_scale().name("lr");
+
+    study
+        .optimize(100, |trial: &mut optimizer::Trial| {
+            let lrv = lr.suggest(trial)?;
+            // Minimum at lr = 0.01
+            Ok::<_, Error>((lrv.ln() - 0.01_f64.ln()).powi(2))
+        })
+        .unwrap();
+
+    for trial in study.trials() {
+        let lrv: f64 = trial.get(&lr).unwrap();
+        assert!(
+            (1e-5..=1.0).contains(&lrv),
+            "log-scale value {lrv} out of bounds"
+        );
+    }
+}
+
+#[test]
+fn custom_mutation_and_crossover() {
+    let sampler = DESampler::builder()
+        .mutation_factor(0.5)
+        .crossover_rate(0.7)
+        .population_size(10)
+        .seed(42)
+        .build();
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+
+    let x = FloatParam::new(-5.0, 5.0).name("x");
+    let y = FloatParam::new(-5.0, 5.0).name("y");
+
+    study
+        .optimize(100, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            let yv = y.suggest(trial)?;
+            Ok::<_, Error>(xv * xv + yv * yv)
+        })
+        .unwrap();
+
+    let best = study.best_trial().unwrap();
+    assert!(
+        best.value < 5.0,
+        "custom F/CR optimization should work, got {}",
+        best.value
+    );
+}

tests/sampler/gp.rs

@@ -0,0 +1,254 @@
+use optimizer::prelude::*;
+use optimizer::sampler::gp::GpSampler;
+
+#[test]
+fn sphere_function() {
+    let sampler = GpSampler::with_seed(42);
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+
+    let x = FloatParam::new(-5.0, 5.0).name("x");
+    let y = FloatParam::new(-5.0, 5.0).name("y");
+
+    study
+        .optimize(80, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            let yv = y.suggest(trial)?;
+            Ok::<_, Error>(xv * xv + yv * yv)
+        })
+        .unwrap();
+
+    let best = study.best_trial().unwrap();
+    assert!(
+        best.value < 2.0,
+        "sphere best value should be < 2.0, got {}",
+        best.value
+    );
+}
+
+#[test]
+fn rosenbrock_function() {
+    let sampler = GpSampler::builder().n_startup_trials(15).seed(42).build();
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+
+    let x = FloatParam::new(-5.0, 5.0).name("x");
+    let y = FloatParam::new(-5.0, 5.0).name("y");
+
+    study
+        .optimize(100, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            let yv = y.suggest(trial)?;
+            let val = (1.0 - xv).powi(2) + 100.0 * (yv - xv * xv).powi(2);
+            Ok::<_, Error>(val)
+        })
+        .unwrap();
+
+    let best = study.best_trial().unwrap();
+    assert!(
+        best.value < 100.0,
+        "rosenbrock best value should be < 100.0, got {}",
+        best.value
+    );
+}
+
+#[test]
+fn bounds_respected() {
+    let sampler = GpSampler::with_seed(123);
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+
+    let x = FloatParam::new(-2.0, 3.0).name("x");
+    let y = FloatParam::new(0.0, 10.0).name("y");
+
+    study
+        .optimize(100, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            let yv = y.suggest(trial)?;
+            Ok::<_, Error>(xv + yv)
+        })
+        .unwrap();
+
+    for trial in study.trials() {
+        let xv: f64 = trial.get(&x).unwrap();
+        let yv: f64 = trial.get(&y).unwrap();
+        assert!((-2.0..=3.0).contains(&xv), "x = {xv} out of bounds [-2, 3]");
+        assert!((0.0..=10.0).contains(&yv), "y = {yv} out of bounds [0, 10]");
+    }
+}
+
+#[test]
+fn mixed_params_float_and_categorical() {
+    let sampler = GpSampler::with_seed(42);
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+
+    let x = FloatParam::new(-5.0, 5.0).name("x");
+    let cat = CategoricalParam::new(vec!["a", "b", "c"]).name("cat");
+
+    study
+        .optimize(50, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            let cv = cat.suggest(trial)?;
+            let penalty = match cv {
+                "a" => 0.0,
+                "b" => 1.0,
+                _ => 2.0,
+            };
+            Ok::<_, Error>(xv * xv + penalty)
+        })
+        .unwrap();
+
+    let best = study.best_trial().unwrap();
+    assert!(
+        best.value < 10.0,
+        "best value should be < 10.0, got {}",
+        best.value
+    );
+}
+
+#[test]
+fn seeded_reproducibility() {
+    let x = FloatParam::new(-5.0, 5.0).name("x");
+    let y = FloatParam::new(-5.0, 5.0).name("y");
+
+    let run = |seed: u64| {
+        let sampler = GpSampler::with_seed(seed);
+        let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+        study
+            .optimize(50, |trial: &mut optimizer::Trial| {
+                let xv = x.suggest(trial)?;
+                let yv = y.suggest(trial)?;
+                Ok::<_, Error>(xv * xv + yv * yv)
+            })
+            .unwrap();
+        study.trials().iter().map(|t| t.value).collect::<Vec<_>>()
+    };
+
+    let results1 = run(42);
+    let results2 = run(42);
+    assert_eq!(results1, results2, "same seed should produce same results");
+}
+
+#[test]
+fn different_seeds_different_results() {
+    let x = FloatParam::new(-5.0, 5.0).name("x");
+    let y = FloatParam::new(-5.0, 5.0).name("y");
+
+    let run = |seed: u64| {
+        let sampler = GpSampler::with_seed(seed);
+        let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+        study
+            .optimize(20, |trial: &mut optimizer::Trial| {
+                let xv = x.suggest(trial)?;
+                let yv = y.suggest(trial)?;
+                Ok::<_, Error>(xv * xv + yv * yv)
+            })
+            .unwrap();
+        study.trials().iter().map(|t| t.value).collect::<Vec<_>>()
+    };
+
+    let results1 = run(42);
+    let results2 = run(99);
+    assert_ne!(
+        results1, results2,
+        "different seeds should produce different results"
+    );
+}
+
+#[test]
+fn single_dimension() {
+    let sampler = GpSampler::with_seed(42);
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+
+    let x = FloatParam::new(-10.0, 10.0).name("x");
+
+    study
+        .optimize(100, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            Ok::<_, Error>((xv - 3.0).powi(2))
+        })
+        .unwrap();
+
+    let best = study.best_trial().unwrap();
+    assert!(
+        best.value < 1.0,
+        "1-D optimization should converge, got {}",
+        best.value
+    );
+}
+
+#[test]
+fn integer_params() {
+    let sampler = GpSampler::with_seed(42);
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+
+    let n = IntParam::new(1, 20).name("n");
+
+    study
+        .optimize(100, |trial: &mut optimizer::Trial| {
+            let nv = n.suggest(trial)?;
+            Ok::<_, Error>(((nv - 10) * (nv - 10)) as f64)
+        })
+        .unwrap();
+
+    let best = study.best_trial().unwrap();
+    let best_n: i64 = best.get(&n).unwrap();
+    assert!(
+        (1..=20).contains(&best_n),
+        "integer value {best_n} out of bounds"
+    );
+    assert!(
+        best.value < 10.0,
+        "integer optimization should converge, got {}",
+        best.value
+    );
+}
+
+#[test]
+fn log_scale_params() {
+    let sampler = GpSampler::with_seed(42);
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+
+    let lr = FloatParam::new(1e-5, 1.0).log_scale().name("lr");
+
+    study
+        .optimize(100, |trial: &mut optimizer::Trial| {
+            let lrv = lr.suggest(trial)?;
+            Ok::<_, Error>((lrv.ln() - 0.01_f64.ln()).powi(2))
+        })
+        .unwrap();
+
+    for trial in study.trials() {
+        let lrv: f64 = trial.get(&lr).unwrap();
+        assert!(
+            (1e-5..=1.0).contains(&lrv),
+            "log-scale value {lrv} out of bounds"
+        );
+    }
+}
+
+#[test]
+fn builder_configuration() {
+    let sampler = GpSampler::builder()
+        .n_startup_trials(5)
+        .n_candidates(500)
+        .noise_variance(1e-4)
+        .seed(42)
+        .build();
+    let study: Study<f64> = Study::with_sampler(Direction::Minimize, sampler);
+
+    let x = FloatParam::new(-5.0, 5.0).name("x");
+    let y = FloatParam::new(-5.0, 5.0).name("y");
+
+    study
+        .optimize(100, |trial: &mut optimizer::Trial| {
+            let xv = x.suggest(trial)?;
+            let yv = y.suggest(trial)?;
+            Ok::<_, Error>(xv * xv + yv * yv)
+        })
+        .unwrap();
+
+    let best = study.best_trial().unwrap();
+    assert!(
+        best.value < 5.0,
+        "custom config optimization should work, got {}",
+        best.value
+    );
+}