Skip to content

A collection of Reinforcement Learning implementations with PyTorch

License

Notifications You must be signed in to change notification settings

hcnoh/rl-collection-pytorch

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

61 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Reinforcement Learning Collection with PyTorch

This repository is a collection of the following reinforcement learning algorithms:

  • Policy-Gradient
  • Actor-Critic
  • Trust Region Policy Optimization
  • Generalized Advantage Estimation
  • Proximal Policy Optimization

More algorithms will be added on this repository.

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

  1. Install Python 3.

  2. 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
  3. Install other packages to run OpenAI Gym environments. These are dependent on the development setting of your machine.

  4. Install PyTorch. The version of PyTorch should be greater or equal than 1.7.0.

Training and Running

  1. Modify config.json as your machine setting.

  2. Execute training process by train.py. An example of usage for train.py are following:

    $ python train.py --model_name=trpo --env_name=BipedalWalker-v3

    The following bash command will help you:

    $ python train.py -h
  3. You can run your pre-trained agents by executing run.py. The usage for running run.py is similar to that of train.py. You can also check the help message by the following bash bash command:

    $ python run.py -h

The results of CartPole environment

The results of Pendulum environment

The results of BipedalWalker environment

Recent Works

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

Future Works

  • Find the errors of the Actor-Critic
  • Implement ACER
  • Search other environments to running the algorithms

References

  • An explaination of TRPO line search: link
  • Additional stability method for PPO value function: link