Representations#

In GV, states and observations are represented by python objects. In order to process them using neural network models, we need a way to convert them into raw numeric form. In GV, this process is performed by the Representation classes. A representation converts state/observation spaces into numeric spaces (as dictionaries of Space), and State/Observation into numeric states and observations (as dictionaries of numpy arrays (ndarray)

State Representations#

Available state representations can be found in state_representations. State representations must implement the abstract interface StateRepresentation, which requires the implementation of a space() property and a convert() method:

class StateRepresentation(state_space)[source]

Converts a State into a dictionary of ndarray.

abstract property space: Dict[str, gym_gridverse.representations.spaces.Space]

returns representation space as as dictionary of numpy arrays

Return type: Dict[str, Space]

abstract convert(state)[source]

returns state representation as dictionary of numpy arrays

Return type: Dict[str, ndarray]

To create one of the representations defined in the GV library, you can use the respective function:

make_state_representation(name, state_space)[source]

Factory function for state representations

Parameters

name (str) – name of the representation
state_space (StateSpace) – inner-environment state space

Return type

StateRepresentation

Observation Representations#

Available observation representations can be found in observation_representations. Observation representations must implement the abstract interface ObservationRepresentation, which requires the implementation of a space() property and a convert()

class ObservationRepresentation(observation_space)[source]

Converts a Observation into a dictionary of ndarray.

abstract property space: Dict[str, gym_gridverse.representations.spaces.Space]