SplitActionsWrapper

class maze.core.wrappers.split_actions_wrapper.SplitActionsWrapper(env: StructuredEnvSpacesMixin | StructuredEnv, split_config: Dict[str, Dict[str, Dict[str, Tuple | List]]])

Splits an actions into separate ones.

An example is given by the LunarLanderContinuous-v3 env. Here we have a box action spaces with shape (2,) such that dimension 0 is the up/down action and dimension 1 is the left/right action. Now if we would like to split this action correspondingly we can wrap the env with the following config:

split_config:

action:

action_up:

indices: [0]

action_side:

indices: [1]

Now the actions as well as the action space is consists of two actions (action_up/action_side).

Parameters:

• env – Environment/wrapper to wrap.

• split_config – The action splitting configuration.

action(action: Dict[str, numpy.ndarray]) → Dict[str, numpy.ndarray]

(overrides ActionWrapper)

Implementation of ActionWrapper interface.

property action_space: gymnasium.spaces.Dict

(overrides StructuredEnvSpacesMixin)

The currently active gym action space.

property action_spaces_dict: Dict[int | str, gymnasium.spaces.Dict]

(overrides StructuredEnvSpacesMixin)

A dictionary of gym action spaces, with policy IDs as keys.

clone_from(env: SplitActionsWrapper) → None

(overrides SimulatedEnvMixin)

implementation of SimulatedEnvMixin.

reverse_action(action: Dict[str, numpy.ndarray]) → Dict[str, numpy.ndarray]

(overrides ActionWrapper)

Implementation of ActionWrapper interface.