https://github.com/ikostrikov/pytorch-a3c

PyTorch implementation of Asynchronous Advantage Actor Critic (A3C) from "Asynchronous Methods for Deep Reinforcement Learning".
https://github.com/ikostrikov/pytorch-a3c

a3c actor-critic asynch asynchronous-advantage-actor-critic asynchronous-methods deep-learning deep-reinforcement-learning python pytorch pytorch-a3c reinforcement-learning

Last synced: about 1 year ago
JSON representation

PyTorch implementation of Asynchronous Advantage Actor Critic (A3C) from "Asynchronous Methods for Deep Reinforcement Learning".

• Host: GitHub
• URL: https://github.com/ikostrikov/pytorch-a3c
• Owner: ikostrikov
• License: mit
• Created: 2017-02-13T03:57:55.000Z (over 9 years ago)
• Default Branch: master
• Last Pushed: 2019-09-25T18:08:56.000Z (over 6 years ago)
• Last Synced: 2025-04-19T13:58:48.317Z (about 1 year ago)
• Topics: a3c, actor-critic, asynch, asynchronous-advantage-actor-critic, asynchronous-methods, deep-learning, deep-reinforcement-learning, python, pytorch, pytorch-a3c, reinforcement-learning
• Language: Python
• Homepage:
• Size: 205 KB
• Stars: 1,265
• Watchers: 41
• Forks: 281
• Open Issues: 25
• Metadata Files:
  • Readme: README.md
  • License: LICENSE.md

Awesome Lists containing this project

README

          
# pytorch-a3c

This is a PyTorch implementation of Asynchronous Advantage Actor Critic (A3C) from ["Asynchronous Methods for Deep Reinforcement Learning"](https://arxiv.org/pdf/1602.01783v1.pdf).

This implementation is inspired by [Universe Starter Agent](https://github.com/openai/universe-starter-agent).

In contrast to the starter agent, it uses an optimizer with shared statistics as in the original paper.

Please use this bibtex if you want to cite this repository in your publications:

    @misc{pytorchaaac,

      author = {Kostrikov, Ilya},

      title = {PyTorch Implementations of Asynchronous Advantage Actor Critic},

      year = {2018},

      publisher = {GitHub},

      journal = {GitHub repository},

      howpublished = {\url{https://github.com/ikostrikov/pytorch-a3c}},

    }

## A2C

I **highly recommend** to check a sychronous version and other algorithms: [pytorch-a2c-ppo-acktr](https://github.com/ikostrikov/pytorch-a2c-ppo-acktr).

In my experience, A2C works better than A3C and ACKTR is better than both of them. Moreover, PPO is a great algorithm for continuous control. Thus, I recommend to try A2C/PPO/ACKTR first and use A3C only if you need it specifically for some reasons.

Also read [OpenAI blog](https://blog.openai.com/baselines-acktr-a2c/) for more information.

## Contributions

Contributions are very welcome. If you know how to make this code better, don't hesitate to send a pull request.

## Usage

```bash

# Works only wih Python 3.

python3 main.py --env-name "PongDeterministic-v4" --num-processes 16

```

This code runs evaluation in a separate thread in addition to 16 processes.

## Results

With 16 processes it converges for PongDeterministic-v4 in 15 minutes.

![PongDeterministic-v4](images/PongReward.png)

For BreakoutDeterministic-v4 it takes more than several hours.