chloechsu / revisiting-ppo

Revisiting Design Choices in Proximal Policy Optimization

The PPO implementation is based on the open-source code for ICLR 2020 paper "Implementation Matters in Deep RL: A Case Study on PPO and TRPO": https://github.com/implementation-matters/code-for-paper.

All our analysis and plots are produced via Jupyter notebooks in the analysis folder.

Failure mode examples

The failure mode example environments are defined in analysis/envs.py, and the experiments are analyzed in the corresponding Jupyter notebooks in the analysis folder.

Beta policy implementation

The Beta policy is implemented as a CtsBetaPolicy class in src/policy_gradients/models.py.

MuJoCo experiments

We assume that the user has a machine with MuJoCo and mujoco_py properly set up and installed. To see if MuJoCo is properly installed, try running the following:

import gym
gym.make_env("Humanoid-v2")

The dependencies are listed in the src/requirements.txt file, can be installed via pip install -r requirements.txt.

As an example, to reproduce our MuJoCo Gaussian vs beta policy comparison figures: run the following commands:

1. cd src/gaussian_vs_beta/
2. python setup_agents.py: the setup_agents.py script contains detailed experiments settings and sets up configuration files for each agent.
3. cd ../
4. Edit the NUM_THREADS variables in the run_agents.py file according to your local machine.
5. Train the agents: python run_agents.py gaussian_vs_beta/agent_configs
6. Plot results in the corresponding Jupyter notebook in the analysis folder.

For other MuJoCo comparisons, similarly see the agent setup files in src/kl_direction and src/base_exp, or create your own custom agent setup file with the desired configurations.

For more details about the code, see the README file in the original GitHub repo: https://github.com/implementation-matters/code-for-paper.