Generative Adversarial Imitation Learning with PyTorch

This repository is for a simple implementation of Generative Adversarial Imitation Learning (GAIL) with PyTorch. This implementation is based on the original GAIL paper (link), and my Reinforcement Learning Collection repository (link).

In this repository, OpenAI Gym environments such as CartPole-v0, Pendulum-v0, and BipedalWalker-v3 are used. You need to install them before running this repository.

Note: The environment's names could be different depending on the version of OpenAI Gym.

Install Dependencies

Install Python 3.
Install the Python packages in requirements.txt. If you are using a virtual environment for Python package management, you can install all python packages needed by using the following bash command:
```
$ pip install -r requirements.txt
```
Install other packages to run OpenAI Gym environments. These are dependent on the development setting of your machine.
Install PyTorch. The version of PyTorch should be greater or equal than 1.7.0.

Training and Running

Modify config.json as your machine setting.
Execute training process by train.py. An example of usage for train.py are following:
```
$ python train.py --env_name=BipedalWalker-v3
```
The following bash command will help you:
```
$ python train.py -h
```

The results of CartPole environment

The results of Pendulum environment

The results of BipedalWalker environment

This result suggests that the causal entropy has little effect on the performance.

Recent Works

The CUDA usage is provided now.
Modified some errors in GAE.
Modified some errors about horizon was corrected.

Future Works

Search other environments to running the algorithms

References

The original GAIL paper: link
Reinforcement Learning Collection with PyTorch: link

JulyAugust1231/gail-pytorch