187 lines
6.5 KiB
Python
187 lines
6.5 KiB
Python
|
|
"""
|
|||
|
|
optimizer/lhs_sampler.py
|
|||
|
|
Latin Hypercube Sampling — generates diverse parameter sets that
|
|||
|
|
span the FULL optimization range with minimum samples.
|
|||
|
|
|
|||
|
|
Why LHS instead of random:
|
|||
|
|
With 20 samples and 8 parameters:
|
|||
|
|
- Pure random: clusters near the center, misses extremes
|
|||
|
|
- LHS: divides each parameter into 20 equal bands, samples exactly
|
|||
|
|
one value per band — GUARANTEED coverage of the full space.
|
|||
|
|
|
|||
|
|
Result: 20 LHS samples cover the space better than 200 random samples.
|
|||
|
|
"""
|
|||
|
|
from __future__ import annotations
|
|||
|
|
|
|||
|
|
import random
|
|||
|
|
import math
|
|||
|
|
from typing import Any
|
|||
|
|
|
|||
|
|
from loguru import logger
|
|||
|
|
|
|||
|
|
from ea.schema import ParameterSchema, ParameterDef
|
|||
|
|
|
|||
|
|
|
|||
|
|
class LatinHypercubeSampler:
|
|||
|
|
"""
|
|||
|
|
Generates n_samples diverse parameter dicts from a ParameterSchema.
|
|||
|
|
|
|||
|
|
Usage:
|
|||
|
|
sampler = LatinHypercubeSampler(seed=42)
|
|||
|
|
samples = sampler.sample(schema, n_samples=20)
|
|||
|
|
# → list of 20 dicts, each covering different regions
|
|||
|
|
|
|||
|
|
Each sample is a complete param dict (all params, optimizable + fixed).
|
|||
|
|
Fixed params always take their automation-safe default value.
|
|||
|
|
"""
|
|||
|
|
|
|||
|
|
def __init__(self, seed: int = None):
|
|||
|
|
self.rng = random.Random(seed)
|
|||
|
|
|
|||
|
|
def sample(self, schema: ParameterSchema, n_samples: int) -> list[dict[str, Any]]:
|
|||
|
|
"""
|
|||
|
|
Generate n_samples diverse parameter sets.
|
|||
|
|
Returns list of complete param dicts (ready for IniBuilder).
|
|||
|
|
"""
|
|||
|
|
opts = schema.optimizable()
|
|||
|
|
if not opts:
|
|||
|
|
logger.warning("LHSampler: no optimizable params — returning n_samples copies of defaults")
|
|||
|
|
return [schema.defaults() for _ in range(n_samples)]
|
|||
|
|
|
|||
|
|
logger.info(
|
|||
|
|
f"LHSampler: generating {n_samples} samples over "
|
|||
|
|
f"{len(opts)} optimizable params: {[p.name for p in opts]}"
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
# Build LHS columns: one per optimizable param
|
|||
|
|
# Each column is a shuffled list of n_samples values — each from a different band
|
|||
|
|
columns: dict[str, list[Any]] = {}
|
|||
|
|
for param in opts:
|
|||
|
|
columns[param.name] = self._lhs_column(param, n_samples)
|
|||
|
|
|
|||
|
|
# Assemble rows: combine column values
|
|||
|
|
base = schema.defaults()
|
|||
|
|
samples = []
|
|||
|
|
for i in range(n_samples):
|
|||
|
|
row = dict(base) # start with all defaults (includes fixed params)
|
|||
|
|
for param in opts:
|
|||
|
|
row[param.name] = columns[param.name][i]
|
|||
|
|
samples.append(row)
|
|||
|
|
|
|||
|
|
return samples
|
|||
|
|
|
|||
|
|
# ── Internal ─────────────────────────────────────────────────────────────
|
|||
|
|
|
|||
|
|
def _lhs_column(self, param: ParameterDef, n: int) -> list[Any]:
|
|||
|
|
"""
|
|||
|
|
Generate n values for one parameter using Latin Hypercube spacing.
|
|||
|
|
Each value comes from a different equally-sized band of the range.
|
|||
|
|
The list is shuffled so rows don't correlate across parameters.
|
|||
|
|
"""
|
|||
|
|
if param.type == "bool":
|
|||
|
|
# Alternate True/False, shuffled
|
|||
|
|
vals = [True if i < n // 2 else False for i in range(n)]
|
|||
|
|
# Ensure roughly 50/50
|
|||
|
|
if n % 2 == 1:
|
|||
|
|
vals.append(self.rng.choice([True, False]))
|
|||
|
|
vals = vals[:n]
|
|||
|
|
self.rng.shuffle(vals)
|
|||
|
|
return vals
|
|||
|
|
|
|||
|
|
if param.type == "enum":
|
|||
|
|
# Cycle through enum values with shuffled repetition
|
|||
|
|
enum_vals = param.enum_values
|
|||
|
|
vals = []
|
|||
|
|
while len(vals) < n:
|
|||
|
|
cycle = list(enum_vals)
|
|||
|
|
self.rng.shuffle(cycle)
|
|||
|
|
vals.extend(cycle)
|
|||
|
|
self.rng.shuffle(vals)
|
|||
|
|
return vals[:n]
|
|||
|
|
|
|||
|
|
if param.type in ("int", "float"):
|
|||
|
|
return self._lhs_continuous(param, n)
|
|||
|
|
|
|||
|
|
# fixed: return default repeated
|
|||
|
|
return [param.default] * n
|
|||
|
|
|
|||
|
|
def _lhs_continuous(self, param: ParameterDef, n: int) -> list[Any]:
|
|||
|
|
"""LHS for continuous (float) or discrete (int) parameters."""
|
|||
|
|
lo = float(param.min)
|
|||
|
|
hi = float(param.max)
|
|||
|
|
step = float(param.step) if param.step else (hi - lo) / n
|
|||
|
|
|
|||
|
|
band_size = (hi - lo) / n
|
|||
|
|
|
|||
|
|
values = []
|
|||
|
|
for i in range(n):
|
|||
|
|
band_lo = lo + i * band_size
|
|||
|
|
band_hi = band_lo + band_size
|
|||
|
|
|
|||
|
|
# Random point within the band
|
|||
|
|
raw = self.rng.uniform(band_lo, band_hi)
|
|||
|
|
|
|||
|
|
# Snap to valid step grid
|
|||
|
|
if step > 0:
|
|||
|
|
steps_from_lo = round((raw - lo) / step)
|
|||
|
|
snapped = lo + steps_from_lo * step
|
|||
|
|
# Clamp to [lo, hi]
|
|||
|
|
snapped = max(lo, min(hi, snapped))
|
|||
|
|
else:
|
|||
|
|
snapped = raw
|
|||
|
|
|
|||
|
|
values.append(param.clamp(snapped))
|
|||
|
|
|
|||
|
|
# Shuffle so the ordering isn't correlated with other params
|
|||
|
|
self.rng.shuffle(values)
|
|||
|
|
return values
|
|||
|
|
|
|||
|
|
def sample_neighbors(
|
|||
|
|
self,
|
|||
|
|
base_params: dict[str, Any],
|
|||
|
|
schema: ParameterSchema,
|
|||
|
|
n_neighbors: int,
|
|||
|
|
step_pct: float = 0.20,
|
|||
|
|
) -> list[dict[str, Any]]:
|
|||
|
|
"""
|
|||
|
|
Generate n_neighbors variants of base_params by nudging
|
|||
|
|
the most impactful optimizable params by ±step_pct of their range.
|
|||
|
|
|
|||
|
|
Used in Phase 2 refinement.
|
|||
|
|
|
|||
|
|
step_pct=0.20 means ±20% of (max - min). Much larger
|
|||
|
|
than the old ±0.5 step — actually explores the space.
|
|||
|
|
"""
|
|||
|
|
opts = schema.optimizable()
|
|||
|
|
if not opts:
|
|||
|
|
return [dict(base_params)]
|
|||
|
|
|
|||
|
|
neighbors = []
|
|||
|
|
for _ in range(n_neighbors):
|
|||
|
|
candidate = dict(base_params)
|
|||
|
|
|
|||
|
|
# Perturb 2–4 random optimizable params
|
|||
|
|
n_perturb = min(len(opts), self.rng.randint(2, 4))
|
|||
|
|
to_perturb = self.rng.sample(opts, n_perturb)
|
|||
|
|
|
|||
|
|
for param in to_perturb:
|
|||
|
|
current = candidate[param.name]
|
|||
|
|
if param.type == "bool":
|
|||
|
|
# 50% chance to flip
|
|||
|
|
if self.rng.random() < 0.5:
|
|||
|
|
candidate[param.name] = not current
|
|||
|
|
continue
|
|||
|
|
if param.type == "enum":
|
|||
|
|
candidate[param.name] = self.rng.choice(param.enum_values)
|
|||
|
|
continue
|
|||
|
|
|
|||
|
|
# float / int: nudge by ±step_pct of range
|
|||
|
|
span = float(param.max) - float(param.min)
|
|||
|
|
delta = span * step_pct * self.rng.choice([-1, 1])
|
|||
|
|
raw = float(current) + delta
|
|||
|
|
candidate[param.name] = param.clamp(raw)
|
|||
|
|
|
|||
|
|
neighbors.append(candidate)
|
|||
|
|
|
|||
|
|
return neighbors
|