ESTrainer

class maze.train.trainers.es.es_trainer.ESTrainer(algorithm_config: maze.train.trainers.es.es_algorithm_config.ESAlgorithmConfig, policy: maze.core.agent.torch_policy.TorchPolicy, shared_noise: maze.train.trainers.es.es_shared_noise_table.SharedNoiseTable, normalization_stats: Optional[Dict[str, Tuple[numpy.ndarray, numpy.ndarray]]])

Trainer class for OpenAI Evolution Strategies.

Parameters

  • algorithm_config – Algorithm parameters.

  • policy – Multi-step policy encapsulating the policy networks

  • shared_noise – The noise table, with the same content for every worker and the master.

  • normalization_stats – Normalization statistics as calculated by the NormalizeObservationWrapper.

load_state(file_path: Union[str, BinaryIO]) → None

implementation of Trainer

load_state_dict(state_dict: Dict) → None

Set the model and optimizer state. :param state_dict: The state dict.

train(distributed_rollouts: maze.train.trainers.es.distributed.es_distributed_rollouts.ESDistributedRollouts, model_selection: Optional[maze.train.trainers.common.model_selection.model_selection_base.ModelSelectionBase]) → None

Run the ES training loop.

Parameters

  • distributed_rollouts – The distribution interface for experience collection.

  • model_selection – Optional model selection class, receives model evaluation results.