Action space compatibility with algorithms

RLlib
1

Hello community,

I had this exception with DQN and DDPG but not with PPO.

UnsupportedSpaceException: Action space (Dict('agent_1': Discrete(5), 'agent_2': Discrete(5))) of DDPGTorchPolicy is not supported for DDPG.

Do we have somewhere in the documentation (or elsewhere) more information on which action spaces are supported or not by which algorithms?

2

Here’s the doc on discrete/continuous action support: Algorithms — Ray 2.4.0

DDPG does not support Discrete action spaces, so it cannot support Dict[…, Discrete]. MADDPG: Algorithms — Ray 2.4.0 may be what you’re looking for?

DQN should support it though; could you post your full algorithm config and multi-agent policy_mapping_fn?