Question about the reported results

[jxy33zrhd]

Thanks for your great work!
I want to konw, is the reported scores, along with the dreamerv3, "are the average episode returns within the last 10k steps, that is, all episodes that finished between 390k and 400k environment frames"? Or the last step model evaluation? Or other calculation methods?

[Shengjiewang-Jason]

Hi, jxy33zrhd
Thank you very much for your interest in our work.
We do not use the environment scores during training. Instead, we have a separate process that evaluates the checkpoint. It runs 10 test environments in parallel to obtain the average return. From a single experiment, we take the highest average score, and then we run multiple experiments with different random seeds and report the average reward across them.