ctm-dqn/agents/d3pg_agent.py

"""
D3PG: a pragmatic DDPG/TD3 hybrid baseline for motorway VSL.

Design:
- keep deterministic actor and simple continuous-action proxy from DDPG
- keep twin critics and delayed actor updates from TD3
- disable target policy smoothing because actions are ultimately snapped to
  discrete speed-limit levels, so smoothed target actions add mismatch
  without much benefit
"""
from __future__ import annotations

from typing import List, Sequence

import gymnasium as gym
import numpy as np
import torch.nn as nn
from gymnasium import spaces
from stable_baselines3 import TD3
from stable_baselines3.common.noise import NormalActionNoise
from stable_baselines3.common.torch_layers import FlattenExtractor

from utils.sb3_manual import ensure_manual_logger, sync_manual_timesteps


class MultiDiscreteWrapper(gym.Env):
    def __init__(self, state_dim: int, action_dims: Sequence[int]):
        super().__init__()
        self.state_dim = state_dim
        self.action_dims = action_dims
        self.num_zones = len(action_dims)
        self.observation_space = spaces.Box(
            low=-np.inf,
            high=np.inf,
            shape=(state_dim,),
            dtype=np.float32,
        )
        self.action_space = spaces.Box(
            low=0.0,
            high=1.0,
            shape=(self.num_zones,),
            dtype=np.float32,
        )

    def reset(self, seed=None, options=None):
        return np.zeros(self.state_dim, dtype=np.float32), {}

    def step(self, action):
        return np.zeros(self.state_dim, dtype=np.float32), 0.0, False, False, {}


def _resolve_activation_fn(name: str):
    key = (name or "relu").strip().lower()
    if key == "relu":
        return nn.ReLU
    if key == "silu":
        return nn.SiLU
    if key == "elu":
        return nn.ELU
    raise ValueError(f"Unsupported D3PG activation: {name}")


def _as_arch_list(value, default: List[int]) -> List[int]:
    if value is None:
        return list(default)
    return [int(v) for v in value]


class D3PGAgent:
    """Twin-critic delayed deterministic policy gradient without target smoothing."""

    def __init__(
        self,
        state_dim: int,
        action_dims: list,
        learning_rate: float = 3e-4,
        buffer_size: int = 100000,
        learning_starts: int = 100,
        batch_size: int = 64,
        tau: float = 0.005,
        gamma: float = 0.99,
        policy_delay: int = 2,
        exploration_sigma: float = 0.1,
        device: str = "cuda",
        actor_hidden_dims: Sequence[int] | None = None,
        critic_hidden_dims: Sequence[int] | None = None,
        activation_fn: str = "relu",
    ):
        self.state_dim = state_dim
        self.action_dims = action_dims
        self.num_zones = len(action_dims)
        self.device = device
        self.learning_starts = learning_starts
        self.total_steps = 0
        self.exploration_sigma = exploration_sigma

        dummy_env = MultiDiscreteWrapper(state_dim, action_dims)
        action_noise = NormalActionNoise(
            mean=np.zeros(self.num_zones),
            sigma=float(exploration_sigma) * np.ones(self.num_zones),
        )
        policy_kwargs = {
            "net_arch": {
                "pi": _as_arch_list(actor_hidden_dims, [256, 256]),
                "qf": _as_arch_list(critic_hidden_dims, [256, 256]),
            },
            "activation_fn": _resolve_activation_fn(activation_fn),
            "features_extractor_class": FlattenExtractor,
        }

        self.model = TD3(
            "MlpPolicy",
            env=dummy_env,
            learning_rate=learning_rate,
            buffer_size=buffer_size,
            learning_starts=learning_starts,
            batch_size=batch_size,
            tau=tau,
            gamma=gamma,
            policy_delay=policy_delay,
            target_policy_noise=0.0,
            target_noise_clip=0.0,
            action_noise=action_noise,
            device=device,
            verbose=0,
            policy_kwargs=policy_kwargs,
        )
        ensure_manual_logger(self.model)

    def select_action(self, state: np.ndarray, deterministic: bool = False):
        if not deterministic and self.total_steps < self.learning_starts:
            discrete_action = np.array(
                [np.random.randint(self.action_dims[i]) for i in range(self.num_zones)],
                dtype=np.int64,
            )
            return discrete_action, 0.0, 0.0

        continuous_action, _ = self.model.predict(state, deterministic=deterministic)
        if not deterministic:
            noise = np.random.normal(0.0, self.exploration_sigma, size=self.num_zones)
            continuous_action = np.clip(continuous_action + noise, 0.0, 1.0)

        discrete_action = np.array(
            [
                int(cont * (self.action_dims[i] - 1) + 0.5)
                for i, cont in enumerate(continuous_action)
            ]
        )
        discrete_action = np.clip(discrete_action, 0, [d - 1 for d in self.action_dims])
        return discrete_action, 0.0, 0.0

    def store_transition(self, state, action, reward, next_state, done):
        self.total_steps += 1
        sync_manual_timesteps(self.model, self.total_steps)
        continuous_action = np.array(
            [action[i] / (self.action_dims[i] - 1) for i in range(self.num_zones)],
            dtype=np.float32,
        )
        self.model.replay_buffer.add(
            state, next_state, continuous_action, reward, done, [{}]
        )

    def update(self):
        if self.model.replay_buffer.size() < self.model.batch_size:
            return {}
        self.model.train(gradient_steps=1, batch_size=self.model.batch_size)
        return {"updates": float(self.model._n_updates)}

    def save(self, path: str):
        self.model.save(path)

    def load(self, path: str):
        self.model = TD3.load(path, device=self.device)
        ensure_manual_logger(self.model)
        self.total_steps = int(getattr(self.model, "num_timesteps", 0))