We benchmarked PPO on several complex continuous control tasks coming from Brax suit (https://github.com/google/brax). We provide in this section results and hyperparameters to reproduce the experiments in GPU (both acquisition and loss computation).
The following curves are obtained by with the hyperparameters we provide in the yaml files (averaged over 5 different seeds). Run are bout 30 minutes using on GPU.
- Make sure your version of Brax is up to date (https://github.com/google/brax)
- If you want to change the env tested, choose the correct yaml file name in
ppo.py(line 117) - Run
OMP_NUM_THREADS=1 XLA_PYTHON_CLIENT_PREALLOCATE=false python ppo.py