Disallowing actions for an agent in multi agent setting

RLlib
1

Hi everyone.
I am trying to setup a two agent setup with some special requirements. I expect agent1 to use only subset1 of action-set and agent2 to use only subset2 of same action-set. subset1 and subset2 may or may not overlap. Can anyone please help me know where to start and what to read for setting this up?

2

Hey @Saurabh_Arora , you can tell your “multiagent” setup to use different action spaces for the two policies:

config:
    multiagent:
        policies:
            pol1: (None, [obs-space], [action-space pol1], {})
            pol2: (None, [obs-space], [action-space pol2], {})

Does this make sense?

1 Like
3

I can definitely try that out. For now, this stage of code went on hold.