[RLlib] Why some algorithms do not suppport multiagent or discrete/continuous action space?

What can I do about algorithms that do not support either action space or multiagent?
For example, if I want to use MAML in discrete multiagent environment or DQN agent with RNN, can I modify the MAML/DQN agent to support this? Maybe I could contribute some code to support some of the usecase I’m using. Thanks.

1 Like

Thanks for reporting this @51616 .

MAML should support discrete actions (in pytorch). There is now a test case with CartPole that proves that.
Please simply try multi-agent with MAML and report back here.
The reason it’s not mentioned in the documentation is simply that we haven’t tested this yet.

DQN + RNNs: Not supported for now, but we are planning on adding R2D2-style (off-policy + RNNs) algos this quarter.

1 Like