QMIX returns empty action

RLlib Configure Algorithm, Training, Evaluation, Scaling
1

I have a multi-agent environment with:

  • Observation space: Nested dict space like {agent_id: {'obs': observation, 'state': env_state}}
  • Action space: Dict action space with discrete actions like {agent_id: action}
  • Reward and done: Dicts {agent_id: reward} and {agent_id: done}

I want to use QMIX, so in my env creator function I do the following:

def env_creator(config):
    env = MyEnv(**config)
    env = env.with_agent_groups(
        groups= {'group_1': env.agent_ids},  # All agents in same group
        observation_space= Tuple([
            Dict({
                'obs': env.observation_space[agent_id]['obs']
                'state': env.observation_space[agent_id]['state']
            })
            for agent_id in env.agent_ids
        ]),
        action_space= Tuple([
            env.action_space[agent_id] for agent_id in env.agent_ids
        ])
    )
    return env
  1. If I run, I receive from QMIX empty actions, i.e. action = {}.
  2. Is the above correct? Both the env observation and action space and the env creator function?