Algorithm trainer specification

An algorithm trainer class exposes at least the following API

  • __init__(self, config: Config, algorithm: Algorithm)

  • train(self, env: Env, **options) -> TrainInfo

The constructor Config parameter specifies the configuration options for the trainer e.g. number of episodes or number of workers. The Algorithm parameter specifies the algorithm to train. Training occurs when the train method is called with the environment instance that should be used for training. An instance of the TrainInfo is returned to the calling site providing basic information about the training

Currently, there are two classes that implement the above API:

  • The RLSerialAlgorithmTrainer class

  • The PyTorchParallelTrainer class

The RLSerialAlgorithmTrainer` class

The PyTorchParallelTrainer class