IMPALA

class maze.train.trainers.impala.impala_trainer.IMPALA(algorithm_config: ImpalaAlgorithmConfig, rollout_generator: DistributedActors, evaluator: RolloutEvaluator | None, model: TorchActorCritic, model_selection: BestModelSelection | None)

Multi step advantage actor critic.

train(n_epochs: int | None = None) → None

(overrides Trainer)

Train function that wraps normal train function in order to close all processes properly

param n_epochs:

number of epochs to train.