felipemartins96 / rl-jax Goto Github PK

List of things to try on this repo:

• Profile agents using jax profiler
• Run 0.0.3 on rsoccer_gym VSS-v0 environment
• SAC learnable temperature
• Batch Normalization
• Hindsight experience replay
• Wandb episode statistics wrapper
• Use jax rng everywhere to force reproducibility

DDPG:

• Change network to Double Q (togglable?)
• Add OU exploration noise (togglable?)
• Hyperparameter for number of observations per update step and number of gradients per update step

TD3:

• Implement TD3

• Unique agent test experiment file for all agents
• Make training information uniform
• Base agent

The DDPG paper describes using batch normalization, this is not currently implemented on this agent