Skip to content

Instantly share code, notes, and snippets.

@kengz
Last active August 11, 2019 18:22
Show Gist options
  • Save kengz/5bc352e18c24ce1fdb4a8679bbb1624a to your computer and use it in GitHub Desktop.
Save kengz/5bc352e18c24ce1fdb4a8679bbb1624a to your computer and use it in GitHub Desktop.
SAC benchmark

Roboschool (continuous control) Benchmark

Note that the Roboschool reward scales are different from MuJoCo's. All results are ran with 4 sessions with distinct random seeds. mean_returns_ma is the returns moving-average over 100 checkpoints from the sessions averaged.

Env. \ SAC mean_returns_ma graph
RoboschoolAnt 2451.55 sac
RoboschoolHalfCheetah 2004.27 sac
RoboschoolHopper 2090.52 sac
RoboschoolWalker2d 1711.92 sac

LunarLander (discrete control) Benchmark

sac_lunar_t0_trial_graph_mean_returns_vs_frames sac_lunar_t0_trial_graph_mean_returns_ma_vs_frames
Trial graph Moving average
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment