This directory includes examples on running the PETS agent on Mujoco benchmarks described in the original paper and a classical cartpole environment.
Under the hood, it uses ml_collections to manage configurations. Optionally
you can log results to wandb.
Run
python run_cartpole.py --config=configs/cartpole_continuous.py
Note that you need should use the cartpole_continuous.py config instead of
cartpole.py which is used by the Mujoco benchmark.
Run
python run_cartpole.py --helpshort
For a list of additional configuration besides those that can be overriden with ml_collections.config_flags
For the classical cartpole, you should consistently get an episode return of 200 (the maximum) within 10 episodes of training. On a machine with NVIDIA GeForce RTX 3080, 10 episodes should take less than 2 minutes. If your running time significantly exceeds this value, make sure that you are installing the GPU version of JAX.
Run
python run_mujoco.py --config=configs/halfcheetah.py
You can change the configuration file to point to any file in configs/