agent/base_agent.py

"""
BaseAgent — abstract interface that all RL agents must implement.

Every concrete agent (Q-Learning, DQN, …) must override:
  • select_action(state, training) → int
  • train_step(state, action, reward, next_state, done) → float | None
  • save(filepath)
  • load(filepath)

Optional hooks:
  • update_target_network()  – used by DQN
  • reset()                  – called between episodes if needed
"""

from abc import ABC, abstractmethod

try:
    import torch
    _TORCH_AVAILABLE = True
except ImportError:
    _TORCH_AVAILABLE = False


class BaseAgent(ABC):
    """Abstract base class for RL agents."""

    def __init__(self, state_size: int, action_size: int, config: dict):
        self.state_size = state_size
        self.action_size = action_size
        self.config = config

        if _TORCH_AVAILABLE:
            import torch
            self.device = torch.device(
                "cuda" if torch.cuda.is_available() else "cpu"
            )
        else:
            self.device = None

    @abstractmethod
    def select_action(self, state, training: bool = True) -> int:
        """Return an action integer for the given state."""

    @abstractmethod
    def train_step(self, state, action, reward, next_state, done):
        """Perform one update step; return loss (or None if not applicable)."""

    @abstractmethod
    def save(self, filepath: str):
        """Persist the agent to *filepath*."""

    @abstractmethod
    def load(self, filepath: str):
        """Restore the agent from *filepath*."""

    # Optional hooks
    def update_target_network(self):
        """Sync target network (DQN only)."""

    def reset(self):
        """Reset any per-episode internal state."""