koulakis / reacher-deep-reinforcement-learning Goto Github PK

Hello,
nice project.

I was wondering if you had considered trying generalized state-dependent exploration (gSDE) and the tuned hyperparameters from the zoo:

A2C: https://github.com/DLR-RM/rl-baselines3-zoo/blob/master/hyperparams/a2c.yml#L126
PPO: https://github.com/DLR-RM/rl-baselines3-zoo/blob/master/hyperparams/ppo.yml#L137
SAC: https://github.com/DLR-RM/rl-baselines3-zoo/blob/master/hyperparams/sac.yml#L141

In a nutshell, gSDE is a different exploration strategy that is made for the continuous action case. It allows to train RL agent directly on real robot for instance.
Cf paper: https://arxiv.org/abs/2005.05719

If you do so, please upgrade your SB3 version, as we pushed a fix recently for PPO+gSDE
and keep me up to date, I'm interested by the results ;)
See https://github.com/DLR-RM/stable-baselines3/releases for the last release