Custom checkpoints in RLLIB

How severe does this issue affect your experience of using Ray?

  • High: It blocks me to complete my task.

Hi all,
I am running multi-agent RL experiments with RLLib and would like to extend the Checkpoint logic. In particular, I am looking for two extra functionalities:

  1. Saving checkpoints based on multiple criteria. I would like to have one checkpoint for both the best episode_reward_mean and the episode_reward_max.

  2. Saving a checkpoint (potentially a third one besides the aforementioned) based on the individual agent’s metric. For example, I would like to save for each agent separately the policy that achieved the best individual episode reward.

Is there already a way to achieve this that I could not find yet? Thanks for the help!

Hi @pasztorb,

do I understand correctly: you run a multi-agent algorithm and then want to store in checkpoints from these multiple agents the ones that have the best policies following your criteria?

Otherwise Ray Tune might already give you something like this in the aftermath. Take a look at the ExperimentAnalysis.