Does MARLlib support a mixed scenario where each agent has a different policy?

[terranovafr]

I was wondering whether the library supports currently the mixed scenario (e.g. two competitive teams with cooperative agents inside) but allowing the possibility of having no policy sharing across teams and inside teams?

If in the policy_mapping_dict I set :
"all_agents_one_policy": False, "one_agent_one_policy": False,
while keeping:
mappo.fit(env, .... share_policy='individual', ....)
I receive the following error:
agent number too large, we disable no sharing function

[Morphlng]

If I'm not misunderstanding, you are trying to train two sets of policy, one for each side of the game. This should be done via share_policy='group' instead of "individual". You should declare how the grouping is done through env_config, see their example for SMAC.

MARLlib/examples/customize_policy_sharing.py

Lines 93 to 94 in 80e9973

	mappo.fit(env, model, stop={'episode_reward_mean': 2000, 'timesteps_total': 10000000}, local_mode=True, num_gpus=1,
	num_workers=2, share_policy='group', checkpoint_freq=50)

There is also an example of scenario settings in their document:

For instance, consider a mixed mode scenario from MAgent, which can be represented using the following policy mapping:

"adversarial_pursuit": {
    "description": "one team attack, one team survive",
    "team_prefix": ("predator_", "prey_"),
    "all_agents_one_policy": False,
    "one_agent_one_policy": False,
},