Add RAdam optimizer - Githubissues

kengz / SLM-Lab

Modular Deep Reinforcement Learning framework in PyTorch. Companion library of the book "Foundations of Deep Reinforcement Learning".

https://slm-lab.gitbook.io/slm-lab/

MIT License

1.25k stars 264 forks source link

Add RAdam optimizer #403

Closed kengz closed 5 years ago

kengz commented 5 years ago

Feature / Fix

Add RAdam optimizer. https://arxiv.org/abs/1908.03265
initial Roboschool benchmark shows the same results as Adam, but this opens up the option to more stably experiment with learning rate.
see the full results for A2C and PPO on Dropbox under PR403

ppo_roboschool_RoboschoolAnt-v1_t0_trial_graph_mean_returns_ma_vs_frames ppo_roboschool_RoboschoolHalfCheetah-v1_t0_trial_graph_mean_returns_ma_vs_frames ppo_roboschool_RoboschoolHopper-v1_t0_trial_graph_mean_returns_ma_vs_frames ppo_roboschool_RoboschoolWalker2d-v1_t0_trial_graph_mean_returns_ma_vs_frames