Files
optimiz-rs/src/optimal_control/hjb_solver.rs
T

350 lines
10 KiB
Rust
Raw Normal View History

//! HJB PDE Solver
//! ==============
//!
//! Generic Hamilton-Jacobi-Bellman equation solver using finite differences.
use crate::optimal_control::{OptimalControlError, Result};
use ndarray::Array1;
/// Configuration for HJB solver
#[derive(Debug, Clone)]
pub struct HJBConfig {
/// Mean-reversion speed (κ in OU process)
pub kappa: f64,
/// Long-term mean (θ in OU process)
pub theta: f64,
/// Volatility (σ in OU process)
pub sigma: f64,
/// Discount rate
pub rho: f64,
/// Transaction cost per trade
pub transaction_cost: f64,
/// Number of grid points
pub n_points: usize,
/// Maximum iterations
pub max_iter: usize,
/// Convergence tolerance
pub tolerance: f64,
/// Number of standard deviations for domain
pub n_std: f64,
}
impl Default for HJBConfig {
fn default() -> Self {
Self {
kappa: 0.5,
theta: 0.0,
sigma: 0.1,
rho: 0.04,
transaction_cost: 0.001,
n_points: 200,
max_iter: 2000,
tolerance: 1e-6,
n_std: 4.0,
}
}
}
/// Result from HJB solver
#[derive(Debug, Clone)]
pub struct HJBResult {
/// State space grid
pub x: Array1<f64>,
/// Value function V(x)
pub value: Array1<f64>,
/// First derivative V'(x)
pub gradient: Array1<f64>,
/// Second derivative V''(x)
pub hessian: Array1<f64>,
/// Lower boundary (buy signal)
pub lower_boundary: f64,
/// Upper boundary (sell signal)
pub upper_boundary: f64,
/// Number of iterations until convergence
pub iterations: usize,
/// Final residual
pub residual: f64,
}
/// Generic HJB PDE Solver
pub struct HJBSolver {
config: HJBConfig,
}
impl HJBSolver {
/// Create new HJB solver with configuration
pub fn new(config: HJBConfig) -> Result<Self> {
// Validate parameters
if config.kappa <= 0.0 {
return Err(OptimalControlError::InvalidParameters(
"kappa must be positive".to_string(),
));
}
if config.sigma <= 0.0 {
return Err(OptimalControlError::InvalidParameters(
"sigma must be positive".to_string(),
));
}
if config.rho <= 0.0 {
return Err(OptimalControlError::InvalidParameters(
"rho must be positive".to_string(),
));
}
if config.n_points < 50 {
return Err(OptimalControlError::InvalidParameters(
"n_points must be at least 50".to_string(),
));
}
Ok(Self { config })
}
/// Solve HJB equation using finite differences
pub fn solve(&self) -> Result<HJBResult> {
let cfg = &self.config;
// Compute stationary standard deviation
let sigma_inf = cfg.sigma / (2.0 * cfg.kappa).sqrt();
// State space: θ ± n_std * σ_∞
let x_min = cfg.theta - cfg.n_std * sigma_inf;
let x_max = cfg.theta + cfg.n_std * sigma_inf;
let dx = (x_max - x_min) / (cfg.n_points - 1) as f64;
// Create grid
let x = Array1::from_iter((0..cfg.n_points).map(|i| x_min + i as f64 * dx));
// Initialize value function
let mut v = Array1::<f64>::zeros(cfg.n_points);
let mut v_old = Array1::<f64>::zeros(cfg.n_points);
// Running reward: quadratic tracking penalty around θ (maximisation of
// -(x-θ)²). Without a source term the stationary equation ρV = LV has
// only the trivial solution V ≡ 0, which made gradients — and hence
// the V' = ±1 switching boundaries — meaningless.
let f: Vec<f64> = x.iter().map(|&xi| -(xi - cfg.theta).powi(2)).collect();
let sig2 = cfg.sigma * cfg.sigma;
let dx2 = dx * dx;
// Iterative solver: implicit (Thomas) solve of the linear part
// ρV = κ(θ-x)V' + ½σ²V'' + f (KushnerDupuis upwind rates), followed
// by projection on the singular-control obstacles
// V(x) ≥ V(x±dx) - dx (unit proportional control cost),
// repeated until the fixed point. A pointwise Jacobi update diverges
// here (σ²/dx² ≫ ρ) and plain value iteration contracts too slowly.
let n = cfg.n_points;
let mut iterations = 0;
let mut residual = f64::INFINITY;
let mut sub = vec![0.0_f64; n];
let mut diag = vec![0.0_f64; n];
let mut sup = vec![0.0_f64; n];
let mut rhs = vec![0.0_f64; n];
for iter in 0..cfg.max_iter {
v_old.assign(&v);
// Assemble tridiagonal system (upwind, unconditionally stable)
for i in 1..n - 1 {
let mu = cfg.kappa * (cfg.theta - x[i]);
let p_up = 0.5 * sig2 / dx2 + mu.max(0.0) / dx;
let p_dn = 0.5 * sig2 / dx2 + (-mu).max(0.0) / dx;
sub[i] = -p_dn;
diag[i] = cfg.rho + p_up + p_dn;
sup[i] = -p_up;
rhs[i] = f[i];
}
// Neumann boundaries: V'(x_min) = V'(x_max) = 0
diag[0] = 1.0;
sup[0] = -1.0;
rhs[0] = 0.0;
sub[n - 1] = -1.0;
diag[n - 1] = 1.0;
rhs[n - 1] = 0.0;
// Thomas algorithm
let mut d = diag.clone();
let mut r = rhs.clone();
for i in 1..n {
let w = sub[i] / d[i - 1];
d[i] -= w * sup[i - 1];
r[i] -= w * r[i - 1];
}
v[n - 1] = r[n - 1] / d[n - 1];
for i in (0..n - 1).rev() {
v[i] = (r[i] - sup[i] * v[i + 1]) / d[i];
}
// Obstacle projection: acting costs 1 per unit of displacement
for i in 1..n {
let candidate = v[i - 1] - dx;
if candidate > v[i] {
v[i] = candidate;
}
}
for i in (0..n - 1).rev() {
let candidate = v[i + 1] - dx;
if candidate > v[i] {
v[i] = candidate;
}
}
// Check convergence
residual = (&v - &v_old)
.mapv(|x| x.abs())
.iter()
.fold(0.0f64, |acc, &x| acc.max(x));
iterations = iter + 1;
if residual < cfg.tolerance {
break;
}
}
// `!(a < b)` also catches NaN residuals
if !(residual < cfg.tolerance) {
return Err(OptimalControlError::ConvergenceError(format!(
"Failed to converge after {} iterations (residual: {:.2e})",
iterations, residual
)));
}
// Compute gradient (first derivative)
let gradient = self.compute_gradient(&v, dx);
// Compute hessian (second derivative)
let hessian = self.compute_hessian(&v, dx);
// Find optimal boundaries
let (lower_boundary, upper_boundary) = self.find_boundaries(&x, &gradient, cfg.theta);
Ok(HJBResult {
x,
value: v,
gradient,
hessian,
lower_boundary,
upper_boundary,
iterations,
residual,
})
}
/// Compute first derivative using central differences
fn compute_gradient(&self, v: &Array1<f64>, dx: f64) -> Array1<f64> {
let n = v.len();
let mut gradient = Array1::<f64>::zeros(n);
// Interior points (central difference)
for i in 1..n - 1 {
gradient[i] = (v[i + 1] - v[i - 1]) / (2.0 * dx);
}
// Boundaries (forward/backward difference)
gradient[0] = (v[1] - v[0]) / dx;
gradient[n - 1] = (v[n - 1] - v[n - 2]) / dx;
gradient
}
/// Compute second derivative using finite differences
fn compute_hessian(&self, v: &Array1<f64>, dx: f64) -> Array1<f64> {
let n = v.len();
let mut hessian = Array1::<f64>::zeros(n);
// Interior points
for i in 1..n - 1 {
hessian[i] = (v[i + 1] - 2.0 * v[i] + v[i - 1]) / dx.powi(2);
}
// Boundaries (one-sided)
hessian[0] = hessian[1];
hessian[n - 1] = hessian[n - 2];
hessian
}
/// Find optimal switching boundaries
#[allow(unused_variables)] // theta parameter reserved for future use
fn find_boundaries(&self, x: &Array1<f64>, gradient: &Array1<f64>, theta: f64) -> (f64, f64) {
let n = x.len();
let mid_idx = n / 2;
// Lower boundary: V' ≈ 1 (below mean)
let mut lower_idx = 0;
let mut min_dist = f64::INFINITY;
for i in 0..mid_idx {
let dist = (gradient[i] - 1.0).abs();
if dist < min_dist {
min_dist = dist;
lower_idx = i;
}
}
// Upper boundary: V' ≈ -1 (above mean)
let mut upper_idx = n - 1;
min_dist = f64::INFINITY;
for i in mid_idx..n {
let dist = (gradient[i] + 1.0).abs();
if dist < min_dist {
min_dist = dist;
upper_idx = i;
}
}
(x[lower_idx], x[upper_idx])
}
}
#[cfg(test)]
mod tests {
use super::*;
use approx::assert_relative_eq;
#[test]
fn test_hjb_solver_convergence() {
let config = HJBConfig {
kappa: 0.5,
theta: 0.0,
sigma: 0.1,
rho: 0.04,
transaction_cost: 0.001,
n_points: 100,
max_iter: 1000,
tolerance: 1e-5,
n_std: 3.0,
};
let solver = HJBSolver::new(config).unwrap();
let result = solver.solve().unwrap();
assert!(result.iterations < 1000);
assert!(result.residual < 1e-5);
assert!(result.lower_boundary < result.upper_boundary);
assert!(result.lower_boundary < 0.0);
assert!(result.upper_boundary > 0.0);
}
#[test]
fn test_hjb_solver_symmetry() {
let config = HJBConfig {
kappa: 1.0,
theta: 0.0,
sigma: 0.2,
..Default::default()
};
let solver = HJBSolver::new(config).unwrap();
let result = solver.solve().unwrap();
// For symmetric OU process, boundaries should be symmetric
assert_relative_eq!(
result.lower_boundary.abs(),
result.upper_boundary.abs(),
epsilon = 0.1
);
}
}