Closed alex-petrenko closed 1 year ago
Base: 79.86% // Head: 80.10% // Increases project coverage by +0.24%
:tada:
Coverage data is based on head (
4fa05ff
) compared to base (0b1d3fc
). Patch coverage: 83.33% of modified lines in pull request are covered.
:umbrella: View full report at Codecov.
:loudspeaker: Do you have feedback about the report comment? Let us know in this issue.
@edbeeching I updated some parameters to be more inline with https://github.com/Denys88/rl_games/blob/master/docs/MUJOCO_ENVPOOL.md and I think the performance is substantially better (both sample efficiency and throughput). Please let me know what you think!
Both rl_games and our config use only 64 agents which is really quite a small number (in IGE we use 4K-8K). I tried playing with params a bit more (more agents, larger batch, shorter rollout), and I can easily get same wall-time performance (so better FPS but worse sample efficiency), but with configs I tried I could not get better wall-time reward than this baseline config.
I also noticed that one some seeds the result is a bit worse than rl_games (on Ant I got ~5000 reward at 5M and rl_games is ~6000 reward). This can be addressed later. Should be a good task for @andrewzhang505 to figure this out in the future.