Help with questions about custom environments about marllib

[Theohhhu]

It appears that the issue may be related to differences between versions of Ray/RLlib, which are not backward and forward compatible. I recommend transitioning to at least version 1.8, which is the foundation of MARLlib. Additionally, transferring an algorithm from our framework to yours might be a challenging task.

[huigeopencv]

The version I am using is ray1.8.0. I guess it is because of my training environment problem and it should be incompatible. But I really don’t have a clue to DEBUG

[Aequatio-Space]

Since you want to use the algorithm of MARLlib, I guess you may need to override the abstract class MultiAgentEnv provided by ray, or write a wrapper for the algorithm to work. Also, try to adhere implementation example from add_new_env.py, which will ensure interfaces are correctly matched.

MARLlib/examples/add_new_env.py

Lines 66 to 83 in 368c617

	class RLlibMAGym(MultiAgentEnv):
	def __init__(self, env_config):
	map = env_config["map_name"]
	env_config.pop("map_name", None)
	self.env = REGISTRY[map](**env_config)
	# assume all agent same action/obs space
	self.action_space = self.env.action_space[0]
	self.observation_space = GymDict({"obs": Box(
	low=0.0,
	high=1.0,
	shape=(self.env.observation_space[0].shape[0],),
	dtype=np.dtype("float64"))})
	self.agents = ["red_0", "blue_0"]
	self.num_agents = len(self.agents)
	env_config["map_name"] = map
	self.env_config = env_config