It could be possible that MASAC in RLlib is designed for standard MDPs, but it can also be applied to POMDPs if you use memory-based models (like LSTMs) or augment observations appropriately. RLlib itself does not restrict MASAC to only MDPs or POMDPs—the distinction depends on your environment and model choice, not the algorithm implementation itself.
Though I’m not entirely certain, it might be that for POMDPs, you need to use recurrent models or custom observation preprocessing, as discussed in RLlib forums and examples. Would you like more detail on handling POMDPs with MASAC?
Sources:
- https://discuss.ray.io/t/approaching-a-pomdp-problem-with-rllib/9508
- https://discuss.ray.io/t/global-optima-with-centralized-critic-basic-understanding/1478
Note: I wasn’t able to find highly relevant data sources, but above are a few potentially relevant links.
Hint: Mention @RunLLM in the post for followups.