Implementing environments

Subclassing BaseEnvironment#

To create custom environments, you shold inherit from BaseEnvironment and implement the _step() method, which is called by the step() method wrapper:

class MyEnv(BaseEnvironment):
    def _step(self, any_env_sepcific_arguments) -> Any:
        # Implement the logic for a single update of the environment.

Note that the semantics of the step depend on the environment and is up to you in your implementation. A step could be one step in an Markov Decisioen Process, one turn or round in a game, one second of simulated time, as well as simulating the environemnt for the given duration. The main consideration here is composing or checpointing environments, or allowing any other external interaction with the environment.

Note that _step may also return any value that may repersent the result of the step. You can also define any attributes to expose the state of the environment to the outside (the agents inside observe the environment only via actor.observe(...)), e.g. a result attribute, a scores array, etc.