snu-mllab / edac Goto Github PK

Hello, I have a question about the performance of SAC-N or EDAC on AntMaze tasks. Have you ever tested it?

In my experiments based on this official implementation, I found that average returns in evaluation are always 0, which is worse than behavior cloning. Then I try to run with a modified reward (r=4*(r-0.5)) and max Q backup. However, they didn't help.

I'll appreciate it a lot if you give me some related advices. Thanks a lot.

Hi! Really puzzled by the REM performance as it is a bit counterintuitive that it does not actually work at all but very similar to SAC-N in theory! The appendix doesn't say how the ensemble was implemented and I don't see it in the code. So, I decided to ask if separate nets were used for the ensemble or just different heads with shared first layers as in original REM paper?

Thanks!