Struct ReinforcementLearningAgent 

pub struct ReinforcementLearningAgent { /* private fields */ }

Production-quality tabular RL agent supporting multiple algorithms and policies.

Register states and their valid actions, then call update after each environment step to learn Q-values. Call run_episode to process a complete episode at once.

Implementations

impl ReinforcementLearningAgent

pub fn new(config: AgentConfig) -> Self

Create a new agent with the given configuration.

Returns RlAgentError::InvalidConfig if hyperparameters are out of range.

pub fn register_state( &mut self, state: RlState, actions: Vec<RlAction>, ) -> Result<(), RlAgentError>

Register state as a valid environment state with actions as its legal action set. Can be called multiple times to add more actions; duplicate actions are ignored.

Errors

Returns RlAgentError::InvalidConfig when actions is empty.

pub fn select_action( &self, state: &RlState, rng_seed: &mut u64, ) -> Result<RlAction, RlAgentError>

Select an action for state according to the configured policy.

Errors

• RlAgentError::StateNotFound if the state is not registered.

pub fn best_action(&self, state: &RlState) -> Result<RlAction, RlAgentError>

Return the greedy (argmax Q) action for state.

Errors

• RlAgentError::StateNotFound if the state is not registered.

pub fn value(&self, state: &RlState) -> f64

V(s) = max_a Q(s, a). Returns 0.0 if the state is unregistered.

pub fn update(&mut self, transition: &Transition) -> Result<f64, RlAgentError>

Apply a single TD update from transition.

Returns the absolute TD error |δ| so callers can track convergence.

Errors

• RlAgentError::StateNotFound — state or next_state not registered.
• RlAgentError::ActionNotFound — action not valid for state.

pub fn run_episode( &mut self, transitions: Vec<Transition>, _rng_seed: u64, ) -> Result<EpisodeStats, RlAgentError>

Process a complete sequence of transitions as one episode.

Epsilon is decayed once after all transitions are processed. Episode statistics are accumulated into AgentStats.

Errors

Propagates any error from update.

pub fn decay_epsilon(&mut self)

Decay epsilon for EpsilonGreedy policies: ε ← max(min_ε, ε × decay). No-op for other policies.

pub fn add_experience(&mut self, t: Transition)

Push t into the experience replay buffer.

pub fn sample_experience( &self, n: usize, rng_seed: u64, ) -> Result<Vec<Transition>, RlAgentError>

Draw n transitions uniformly at random from the replay buffer.

Errors

• RlAgentError::InsufficientExperience when the buffer has fewer than n entries.

pub fn stats(&self) -> AgentStats

Return a snapshot of aggregate statistics.

Trait Implementations

impl Debug for ReinforcementLearningAgent

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.