train.coach

train.coach

Training orchestrator: self-play -> replay buffer -> SGD -> checkpoint.

Classes

Name	Description
Coach	Runs the AlphaZero training loop.

Coach

train.coach.Coach(
    network,
    optimizer,
    replay_buffer,
    num_simulations=100,
    batch_size=256,
    games_per_iteration=50,
    training_steps_per_iteration=100,
    checkpoint_dir='checkpoints',
    log_dir='logs',
    experiment_name='',
    checkpoint_every=10,
    num_workers=8,
    virtual_loss_count=1,
    dirichlet_alpha=0.0,
    noise_eps=0.25,
    value_bootstrap_alpha=1.0,
    scheduler=None,
    notes='',
)

Runs the AlphaZero training loop.

Methods

Name	Description
run_iteration	One full iteration: self-play -> train -> checkpoint -> log.
self_play_phase	Generate training data through parallel self-play.
training_phase	Train the network on sampled positions from the replay buffer.

run_iteration

train.coach.Coach.run_iteration(iteration, last_iteration=None)

One full iteration: self-play -> train -> checkpoint -> log.

self_play_phase

train.coach.Coach.self_play_phase()

Generate training data through parallel self-play.

training_phase

train.coach.Coach.training_phase()

Train the network on sampled positions from the replay buffer.

target_value is already normalised to [-1, 1] by self_play (raw backgammon returns ±1/±2/±3 divided by 3), so it can be matched directly against the tanh-bounded value head output.