Question about the normalization of results and the benchmarked CQL performance in Table 2 (ICLR submission)

[zhaoyi11]

Hi,

Thanks so much for your work. I have a question about the normalization of results. Specifically, e.g., in the Gym domain, each result is normalized according to the expert-policy (sac) and random-policy. But which number should we refer to? On the Wiki/"Off policy evaluation" page, there is a form that includes the expert-policy and random-policy, should we refer these? Also, the results of the expert-policy are different from the SAC results in Table3 (ICLR), so which one should we use?

And I noticed that in Table 2 and 3 (ICLR), the result of CQL-'hopper-medium' seems not aligned, could you please confirm this (maybe also the CQL-'walker2d-medium')?

Thanks.

[zhihanyang2022]

We can find the random and expert scores here: https://github.com/rail-berkeley/d4rl/blob/master/d4rl/infos.py.

[zhaoyi11]

Thanks a lot! I will close this issue.