DRAMA at the PettingZoo

Dynamically Restricted Action Spaces for
Multi-Agent Reinforcement Learning Frameworks

DRAMA is a framework to extend the PettingZoo agent-environment cycle with complex, dynamic, and potentially self-learning action space restrictions. DRAMA contains a wrapper implementing a new agent-restrictor-environment cycle, restrictions as Gymnasium Spaces, classes for restriction learning agents, and helpful utilities:

from drama.restrictors import Restrictor
from drama.wrapper import RestrictionWrapper

env = ...
restrictor = Restrictor(...)
wrapper = RestrictionWrapper(env, restrictor)
policies = {...}

wrapper.reset()
for agent in wrapper.agent_iter():
    observation, reward, termination, truncation, info = wrapper.last()
    action = policies[agent](observation)
    wrapper.step(action)

Note

Further documentation is in development …