Struct TrainingSession 

pub struct TrainingSession<E: Environment, A> { /* private fields */ }

A self-contained, loop-agnostic training coordinator.

TrainingSession wires together a LearningAgent, an optional TrainingRun, and a StatsTracker. It is driven purely by incoming data – it does not own a training loop. Feed it experiences and episode boundaries from wherever your loop lives: a plain for loop, Bevy’s ECS, or anything else.

Usage

let session = TrainingSession::new(agent)
    .with_run(TrainingRun::create("cartpole", "v1")?)
    .with_max_steps(200_000)
    .with_checkpoint_freq(10_000);

// Each environment step:
session.observe(experience);

// Each episode end:
session.on_episode(total_reward, steps, status, env_extras);

if session.is_done() { break; }

Implementations

impl<E, A> TrainingSession<E, A>

where E: Environment, E::Observation: Clone + Send + Sync + 'static, E::Action: Clone + Send + Sync + 'static, A: LearningAgent<E>,

pub fn new(agent: A) -> Self

Create a session with no run attached.

Stats are tracked in memory but nothing is persisted. Attach a run with with_run to enable checkpointing and JSONL logging.

pub fn with_run(self, run: TrainingRun) -> Self

Attach a TrainingRun for checkpointing and JSONL episode logging.

pub fn with_max_steps(self, n: usize) -> Self

Maximum number of steps before is_done() returns true. Default: no limit.

pub fn with_checkpoint_freq(self, freq: usize) -> Self

Checkpoint frequency in steps. Default: 10_000.

pub fn with_keep_checkpoints(self, keep: usize) -> Self

Number of numbered checkpoints to retain on disk. Default: 3.

pub fn with_stats(self, stats: StatsTracker) -> Self

Replace the default StatsTracker with a custom one.

pub fn act(&mut self, obs: &E::Observation, mode: ActMode) -> E::Action

Select an action for obs according to mode.

pub fn observe(&mut self, experience: Experience<E::Observation, E::Action>)

Record a transition. Checkpoints + prunes if a step milestone is hit.

pub fn on_episode( &mut self, total_reward: f64, steps: usize, status: EpisodeStatus, env_extras: HashMap<String, f64>, )

Record an episode boundary.

Merges agent and environment extras into the record, updates stats, and appends to the training JSONL log (if a run is attached).

env_extras should come from [crate::traits::EpisodeStats::episode_extras] if the environment implements it, or an empty map otherwise.

pub fn on_episode_start(&mut self)

Signal the start of a new episode (resets per-episode agent aggregators).

pub fn total_steps(&self) -> usize

Total environment steps observed so far.

pub fn steps_per_sec(&self) -> f64

Average environment steps per wall-clock second since the session was created.

pub fn is_done(&self) -> bool

Returns true when total_steps >= max_steps.

pub fn on_eval_episode(&self, record: &EpisodeRecord)

Log an eval episode to the run (if attached).

pub fn maybe_save_best(&mut self, mean_reward: f64)

Save best.mpk if mean_reward exceeds the best seen so far.

pub fn agent(&self) -> &A

Read-only access to the agent.

pub fn agent_mut(&mut self) -> &mut A

Mutable access to the agent.

pub fn stats_summary(&self) -> HashMap<String, f64>

Current stats summary.

pub fn run(&self) -> Option<&TrainingRun>

Read-only access to the run (if attached).

pub fn into_agent(self) -> A

Consume the session and return the inner agent.

pub fn eval_report(&self, n_episodes: usize) -> EvalReport

Snapshot the current stats as an EvalReport.