OK! i continued digging around. as pointed out here, the queue.Empty error surfaces when the observation space is not correct.
in particular, my observation space was a Tuple([Tuple([Box, Box]), Tuple([Box, Box])])
, which covers both groups, but a multiagent observation’s Space
should only describe the righthand side of each multiagent observation in the multiagent observation dictionary. in my case, this should have been Tuple[Box, Box]
.
same problem with my action space, it should have been Tuple([Discrete, Discrete])
.
this made me confused as to how we can have heterogeneous agent groupings, but, i’m guessing that’s where the multiagent mapping section of the trainer config comes in to play.