Files
2026-05-18 19:02:23 +08:00

413 lines
17 KiB
Python

"""Module D: Three-phase HPI control flow (Algorithm 1) — NN edition.
Phase 1: Find admissible L0 with PD critic (NN-based).
Phase 2: Homotopy contraction L_i -> 0 with safety constraint.
Phase 3: Standard policy iteration on original system (L = 0).
"""
import copy
import numpy as np
import torch
from .data_collector import DataCollector
from .nn_models import check_lyapunov_decrease, check_positive_definite
from .nn_trainer import NNTrainer
class HPIController:
"""Implements the three-phase Homotopy-Based Policy Iteration algorithm
using neural network function approximators."""
def __init__(self, J=1.0, mgl=1.0, R=1.0, pe_amplitude=0.5,
lr=1e-3, epochs=500, lambda_weight=1e-6):
"""
Args:
J: Moment of inertia (default 1.0).
mgl: Mass x gravity x length product (default 1.0).
R: Control cost weight (default 1.0).
pe_amplitude: PE signal amplitude scale (default 0.5).
lr: Learning rate for Adam optimizer.
epochs: Training epochs per HPI iteration.
lambda_weight: L2 regularization coefficient.
"""
self.J = float(J)
self.mgl = float(mgl)
self.R = float(R)
self.lr = float(lr)
self.epochs = int(epochs)
self.lambda_weight = float(lambda_weight)
self.pe_amplitude = float(pe_amplitude)
self.collector = DataCollector(J=J, mgl=mgl)
self.trainer = NNTrainer(R=R, lr=lr, epochs=epochs,
lambda_weight=lambda_weight)
# Override PE amplitude
self.collector._scale = pe_amplitude / np.sum(
np.abs(self.collector._a) + np.abs(self.collector._b)
)
# ═══════════════════════════════════════════════════════════════
# Utility
# ═══════════════════════════════════════════════════════════════
def check_positive_definite(self, critic_nn=None, X_traj=None):
"""Check if critic NN corresponds to a PD value function.
Checks both Hessian at origin and V(x) positivity on trajectory.
"""
if critic_nn is None:
critic_nn = self.trainer.critic_nn
return check_positive_definite(critic_nn, X_traj=X_traj)
def check_lyapunov_decrease(self, X, f_X):
"""Check if current critic satisfies Lyapunov decrease on trajectory."""
return check_lyapunov_decrease(self.trainer.critic_nn, X, f_X)
def _compute_f_X(self, X):
"""Drift dynamics f(x) = [x2, sin(x1)] on trajectory."""
X = np.asarray(X)
f_X = np.zeros_like(X)
f_X[:, 0] = X[:, 1]
f_X[:, 1] = np.sin(X[:, 0])
return f_X
def _hessian_info(self):
"""Return (h11, h22, det) of the critic Hessian at origin for logging."""
x0 = torch.zeros(1, 2, requires_grad=True)
V0 = self.trainer.critic_nn(x0)
grad_V = torch.autograd.grad(V0.sum(), x0, create_graph=True)[0]
h11 = torch.autograd.grad(grad_V[0, 0], x0, retain_graph=True)[0][0, 0].item()
h22 = torch.autograd.grad(grad_V[0, 1], x0, retain_graph=True)[0][0, 1].item()
h12 = torch.autograd.grad(grad_V[0, 0], x0, retain_graph=True)[0][0, 1].item()
det = h11 * h22 - h12 ** 2
return h11, h22, det
# ═══════════════════════════════════════════════════════════════
# Phase 1: Find L0 — scan ALL L values, pick best loss
# ═══════════════════════════════════════════════════════════════
def phase_one_find_L0(self, x0, T, dt, L_start=0.1, L_step=0.5, L_max=20.0,
verbose=False):
"""Find admissible L0 by scanning all L values and picking the best.
Unlike the polynomial version which returns at the FIRST PD L,
this scans ALL L values and chooses the one with lowest Bellman
residual loss AND positive-definite critic. This ensures Phase 2
starts from the best possible critic.
Args:
x0: Initial state (2,).
T: Simulation duration.
dt: Time step.
L_start: Starting L value.
L_step: Increment step.
L_max: Maximum L value before failure.
verbose: If True, print progress.
Returns:
tuple: (L0, critic_sd, actor_sd, results_dict).
Raises:
RuntimeError: If no admissible L is found below L_max.
"""
L_values = []
L = L_start
while L <= L_max:
L_values.append(L)
L += L_step
if not L_values:
raise RuntimeError("No L values to try (L_start > L_max)")
# Collect ONE shared trajectory with PE-only control (no actor)
# for training at all L values.
def u_func(t, x):
return self.collector.pe_signal(t)
t, X, U = self.collector.collect_trajectory(x0, T, dt, u_func)
f_X = self._compute_f_X(X)
best_L = None
best_loss = float('inf')
best_critic_sd = None
best_actor_sd = None
best_info = None
for L in L_values:
if verbose:
print(f" Phase 1: trying L={L:.2f}...", flush=True)
# Fresh trainer for each L
trainer = NNTrainer(R=self.R, lr=self.lr, epochs=self.epochs,
lambda_weight=self.lambda_weight)
result = trainer.train_one_iteration(t, X, U, L, prev_actor=None, dt=dt)
loss = result['loss_c']
pd = check_positive_definite(trainer.critic_nn, X_traj=X)
lyap = check_lyapunov_decrease(trainer.critic_nn, X, f_X)
if verbose:
h11, h22, det = self._hessian_info_via_trainer(trainer)
print(f" loss={loss:.4e} PD={pd} lyap={lyap:.3f} "
f"H=[{h11:.4f},{h22:.4f}] det={det:.4e}", flush=True)
# Only consider PD critics; pick lowest loss
if pd and loss < best_loss:
best_L = L
best_loss = loss
best_critic_sd = {k: v.clone() for k, v in trainer.critic_nn.state_dict().items()}
best_actor_sd = {k: v.clone() for k, v in trainer.actor_nn.state_dict().items()}
best_info = {'loss': loss, 'lyap': lyap, 'mse': result['mse_c']}
if best_L is None:
raise RuntimeError(
f"Phase 1 failed: no admissible L found up to L_max={L_max}. "
f"Try increasing L_max or adjusting training parameters."
)
if verbose:
print(f" Phase 1: selected L0={best_L:.2f} (loss={best_loss:.4e})",
flush=True)
# Load best weights into main trainer
self.trainer.critic_nn.load_state_dict(best_critic_sd)
self.trainer.actor_nn.load_state_dict(best_actor_sd)
return best_L, best_critic_sd, best_actor_sd, best_info
def _hessian_info_via_trainer(self, trainer):
"""Get Hessian info from a specific trainer instance."""
x0 = torch.zeros(1, 2, requires_grad=True)
V0 = trainer.critic_nn(x0)
grad_V = torch.autograd.grad(V0.sum(), x0, create_graph=True)[0]
h11 = torch.autograd.grad(grad_V[0, 0], x0, retain_graph=True)[0][0, 0].item()
h22 = torch.autograd.grad(grad_V[0, 1], x0, retain_graph=True)[0][0, 1].item()
h12 = torch.autograd.grad(grad_V[0, 0], x0, retain_graph=True)[0][0, 1].item()
det = h11 * h22 - h12 ** 2
return h11, h22, det
# ═══════════════════════════════════════════════════════════════
# Phase 2: Homotopy contraction
# ═══════════════════════════════════════════════════════════════
def _compute_safe_alpha(self, X, prev_actor, L, gamma=0.1):
"""Compute safe homotopy step alpha via inequality constraint."""
x_tensor = torch.tensor(X, dtype=torch.float32, requires_grad=True)
V = self.trainer.critic_nn(x_tensor)
grad_V = torch.autograd.grad(V.sum(), x_tensor, create_graph=False)[0]
dot_V = (grad_V * x_tensor).sum(dim=1).detach().numpy() # (M,)
Q = X[:, 0] ** 2 + X[:, 1] ** 2 # (M,)
if prev_actor is not None:
with torch.no_grad():
u_old = prev_actor(torch.tensor(X, dtype=torch.float32)).numpy()
else:
u_old = np.zeros(len(X))
with torch.no_grad():
u_new = self.trainer.actor_nn(torch.tensor(X, dtype=torch.float32)).numpy()
gamma_c = 1.0 - gamma
rhs = gamma_c * (Q + self.R * u_new ** 2) + self.R * (u_old - u_new) ** 2
lhs = gamma_c * Q
eps = 1e-10
lower_candidates = []
upper_candidates = []
for k in range(len(dot_V)):
d = dot_V[k]
if d > eps:
lower_candidates.append(lhs[k] / d)
upper_candidates.append(rhs[k] / d)
if not lower_candidates or not upper_candidates:
return None
alpha_lo = max(lower_candidates)
alpha_hi = min(upper_candidates)
if alpha_lo <= alpha_hi and alpha_hi > 0:
return min(alpha_hi, L)
return None
def phase_two_contraction(self, x0, T, dt, L0, actor_sd_init, critic_sd_init,
gamma=0.1, max_iter=50, verbose=False):
"""Contract homotopy parameter L to zero via safety-constrained steps.
Returns:
tuple: (actor_sd, critic_sd, history).
"""
L = L0
self.trainer.critic_nn.load_state_dict(critic_sd_init)
self.trainer.actor_nn.load_state_dict(actor_sd_init)
self.trainer.reset_optimizer()
best_actor_sd = {k: v.clone() for k, v in actor_sd_init.items()}
best_critic_sd = {k: v.clone() for k, v in critic_sd_init.items()}
reject_streak = 0
history = [(L, 0.0, True)] # (L, loss, PD)
for iteration in range(max_iter):
if verbose:
print(f" Phase 2 iter {iteration}: L={L:.4f}", flush=True)
prev_actor = self.trainer.copy_actor()
def u_func(t, x):
pe = self.collector.pe_signal(t)
u_policy = self.trainer.compute_u_hat(x)
return pe + u_policy
t, X, U = self.collector.collect_trajectory(x0, T, dt, u_func)
f_X = self._compute_f_X(X)
# Fresh optimizer per iteration
self.trainer.reset_optimizer()
result = self.trainer.train_one_iteration(t, X, U, L, prev_actor, dt)
pd_new = check_positive_definite(self.trainer.critic_nn, X_traj=X)
alpha = self._compute_safe_alpha(X, prev_actor, L, gamma)
if alpha is None:
alpha = L * 0.2
if pd_new:
best_actor_sd = {k: v.clone() for k, v in self.trainer.actor_nn.state_dict().items()}
best_critic_sd = {k: v.clone() for k, v in self.trainer.critic_nn.state_dict().items()}
reject_streak = 0
if verbose:
lyap = check_lyapunov_decrease(self.trainer.critic_nn, X, f_X)
print(f" alpha={alpha:.4f} loss={result['loss_c']:.4e} "
f"lyap={lyap:.3f} PD", flush=True)
else:
self.trainer.actor_nn.load_state_dict(best_actor_sd)
self.trainer.critic_nn.load_state_dict(best_critic_sd)
reject_streak += 1
if verbose:
print(f" alpha={alpha:.4f} loss={result['loss_c']:.4e} "
f"not PD (reject streak={reject_streak})", flush=True)
L = max(0.0, L - alpha)
history.append((L, result['loss_c'], pd_new))
if L < 1e-14:
if verbose:
print(" L reached 0, Phase 2 complete", flush=True)
break
if reject_streak > 15:
if verbose:
print(" >15 consecutive rejections, stopping Phase 2", flush=True)
break
return best_actor_sd, best_critic_sd, history
# ═══════════════════════════════════════════════════════════════
# Phase 3: Standard policy iteration
# ═══════════════════════════════════════════════════════════════
def phase_three_pi(self, x0, T, dt, actor_sd_init, critic_sd_init,
epsilon=1e-6, max_iter=30, verbose=False):
"""Standard policy iteration on the original system (L = 0).
Returns:
tuple: (actor_sd, critic_sd, converged, history).
"""
self.trainer.actor_nn.load_state_dict(actor_sd_init)
self.trainer.critic_nn.load_state_dict(critic_sd_init)
self.trainer.reset_optimizer()
prev_loss = float('inf')
converged = False
history = []
for iteration in range(max_iter):
if verbose:
print(f" Phase 3 iter {iteration}", flush=True)
prev_actor = self.trainer.copy_actor()
def u_func(t, x):
pe = self.collector.pe_signal(t)
u_policy = self.trainer.compute_u_hat(x)
u_policy = np.clip(u_policy, -5.0, 5.0)
return pe + u_policy
t, X, U = self.collector.collect_trajectory(x0, T, dt, u_func)
self.trainer.reset_optimizer()
result = self.trainer.train_one_iteration(t, X, U, 0.0, prev_actor, dt)
loss = result['loss_c']
history.append(loss)
pd = check_positive_definite(self.trainer.critic_nn, X_traj=X)
if verbose:
f_X = self._compute_f_X(X)
lyap = check_lyapunov_decrease(self.trainer.critic_nn, X, f_X)
print(f" loss={loss:.4e} PD={pd} lyap={lyap:.3f}", flush=True)
if abs(prev_loss - loss) < epsilon:
converged = True
break
prev_loss = loss
if loss > 1e6 or not np.isfinite(loss):
break
actor_sd = {k: v.clone() for k, v in self.trainer.actor_nn.state_dict().items()}
critic_sd = {k: v.clone() for k, v in self.trainer.critic_nn.state_dict().items()}
return actor_sd, critic_sd, converged, history
# ═══════════════════════════════════════════════════════════════
# Full pipeline
# ═══════════════════════════════════════════════════════════════
def run(self, x0=None, T=10.0, dt=0.01,
L_start=0.1, L_step=0.5, L_max=20.0,
gamma=0.1, epsilon=1e-6, verbose=True):
"""Run the full three-phase HPI algorithm.
Returns:
dict: Results containing:
- phase1_L0, phase1_critic_sd, phase1_actor_sd, phase1_info
- phase2_history, phase2_actor_sd, phase2_critic_sd
- phase3_actor_sd, phase3_critic_sd, phase3_converged, phase3_history
"""
if x0 is None:
x0 = np.array([0.5, 0.0])
results = {}
# ── Phase 1: Find L0 ──
L0, critic_sd1, actor_sd1, info1 = self.phase_one_find_L0(
x0, T, dt, L_start, L_step, L_max, verbose=verbose)
results["phase1_L0"] = L0
results["phase1_critic_sd"] = critic_sd1
results["phase1_actor_sd"] = actor_sd1
results["phase1_info"] = info1
# ── Phase 2: Homotopy contraction ──
actor_sd2, critic_sd2, hist2 = self.phase_two_contraction(
x0, T, dt, L0, actor_sd1, critic_sd1, gamma=gamma, verbose=verbose)
results["phase2_actor_sd"] = actor_sd2
results["phase2_critic_sd"] = critic_sd2
results["phase2_history"] = hist2
# ── Phase 3: Policy iteration ──
actor_sd3, critic_sd3, converged, hist3 = self.phase_three_pi(
x0, T, dt, actor_sd2, critic_sd2, epsilon, verbose=verbose)
results["phase3_actor_sd"] = actor_sd3
results["phase3_critic_sd"] = critic_sd3
results["phase3_converged"] = converged
results["phase3_history"] = hist3
# Load final weights
self.trainer.actor_nn.load_state_dict(actor_sd3)
self.trainer.critic_nn.load_state_dict(critic_sd3)
return results