syna.algo.dqn module

DQN agent implementation for discrete action spaces.

Contains: - QNet: simple 3-layer MLP producing Q-values for each action. - DQNAgent: lightweight DQN with replay buffer, target network and Adam optimizer.

class syna.algo.dqn.DQNAgent(action_size: int = 2, gamma: float = 0.98, lr: float = 0.0005, epsilon: float = 0.1, buffer_size: int = 10000, batch_size: int = 32)

Bases: object

Simplified DQN agent.

The agent uses an epsilon-greedy policy, a replay buffer, and a target network.

select_action(state: ndarray) → int

Return an action using epsilon-greedy policy.

sync_qnet()

Copy current Q network weights to the target network.

update(state, action, reward, next_state, done)

Store transition and update Q network from a sampled batch when available.

class syna.algo.dqn.QNet(action_size: int)

Bases: Model

Small feed-forward network that outputs Q-values for each action.

Args:

action_size: number of discrete actions.

forward(x)

Compute the forward pass. Must be implemented by subclasses.