https://github.com/lgvaz/rlbox

RLbox: Solving OpenAI Gym with TensorFlow
https://github.com/lgvaz/rlbox

atari continuous-control deep-reinforcement-learning deep-rl dqn mujoco openai-gym ppo proximal-policy-optimization tensorflow

Last synced: 2 months ago
JSON representation

RLbox: Solving OpenAI Gym with TensorFlow

• Host: GitHub
• URL: https://github.com/lgvaz/rlbox
• Owner: lgvaz
• License: mit
• Created: 2017-07-15T19:52:02.000Z (over 7 years ago)
• Default Branch: master
• Last Pushed: 2018-04-19T18:43:10.000Z (over 6 years ago)
• Last Synced: 2024-08-15T00:20:36.004Z (5 months ago)
• Topics: atari, continuous-control, deep-reinforcement-learning, deep-rl, dqn, mujoco, openai-gym, ppo, proximal-policy-optimization, tensorflow
• Language: Python
• Homepage:
• Size: 2.62 MB
• Stars: 7
• Watchers: 4
• Forks: 6
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# Reinforcement Learning Box

*Discountinued I'm now working on a [Pytorch RL library](https://github.com/lgvaz/torchrl)*  

RLbox provides a framework for rapid experimentation with popular Deep Reinforcement Learning algorithms, it focus on making very easy to implement new ideias, which can be rapidly evaluated using [OpenAI Gym](https://github.com/openai/gym).  

## Installation

```bash

git clone https://github.com/apparatusbox/rlbox.git

cd rlbox  

pip install -e .  

```

## How to use  

Examples on how to run different agents can be found on the [examples](https://github.com/lgvaz/rlbox/tree/master/examples) folder.  

## Implemented algorithms  

### State of the art

* [DQN](https://storage.googleapis.com/deepmind-media/dqn/DQNNaturePaper.pdf) and extensions

  * [Double Q-learning](https://arxiv.org/pdf/1509.06461.pdf)  

  * [Dueling networks](https://arxiv.org/pdf/1511.06581.pdf)

  * N-step learning

  * Soft target update

* [PPO](https://arxiv.org/pdf/1707.06347.pdf)

  * Clipped Surrogate Objective  

  * Adaptive KL Penalty Coefficient  

  

### Classical

* Vanilla Policy Gradient

* REINFORCE

* Actor-Critic

## Results  

* __DQN on BreakoutNoFrameskip-v4__  

Episode 0 ---------------- Episode 3500 ----------- Episode 6000 ----------- Episode 7500 ----------- Episode 21500  

![episode 0](assets/ep0_nolegend.gif)

![episode 3500](assets/ep3500_nolegend.gif)

![episode 6000](assets/ep6000_nolegend.gif)

![episode 7500](assets/ep7500_nolegend.gif)

![episode 21500](assets/ep21500_nolegend.gif)   

Mean reward after training: 421 (Averaged over 100 episodes)  

Dark blue: Standard DQN  

Light blue: Double DQN  

![Breakout reward](assets/breakout_plots.png)  

* __PPO on Hopper-v1__ [Video](https://www.youtube.com/watch?v=QHAu8EWRJJ0&feature=youtu.be)  

![Hopper reward](assets/ppo_reward.png)