https://github.com/ondrejbiza/drl_gym

Deep Reinforcement Learning gym.
https://github.com/ondrejbiza/drl_gym

deep-neural-networks deep-reinforcement-learning openai-gym reinforcement-agents reinforcement-learning tensorflow tensorflow-experiments

Last synced: 26 days ago
JSON representation

Deep Reinforcement Learning gym.

• Host: GitHub
• URL: https://github.com/ondrejbiza/drl_gym
• Owner: ondrejbiza
• Created: 2017-07-02T21:16:36.000Z (over 7 years ago)
• Default Branch: master
• Last Pushed: 2017-08-27T20:15:05.000Z (over 7 years ago)
• Last Synced: 2024-11-10T23:12:32.414Z (3 months ago)
• Topics: deep-neural-networks, deep-reinforcement-learning, openai-gym, reinforcement-agents, reinforcement-learning, tensorflow, tensorflow-experiments
• Language: Python
• Homepage:
• Size: 1.02 MB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
## Deep Reinforcement Learning gym

This repository contains a modular framework for testing Deep Reinforcement Learning algorithms in 

[OpenAI's Gym](http://gym.openai.com/). All the agents are implemented in [Tensorflow](https://www.tensorflow.org/).

### Requirements ###

* Python >= 3 - if you want to use [Mujoco](https://gym.openai.com/envs#mujoco), you will need to install Python 3.5.2 (as of 23.7.2017)

* OpenAI gym

* Tensorflow and other packages listed in **requirements.txt**

#### Mujoco ####

Mujoco is required only if you want to test continuous control tasks like Reacher-v1. Installing Mujoco for Python

is quite difficult as of 23.7.2017. Your best bet is to follow this [thread](https://github.com/openai/mujoco-py/issues/47).

Be advised that installing *xserver-xorg-dev* might stop Ubuntu from receiving keyboard and mouse input.

### Usage ###

You can run continuous control task using **train_continuous_task.py** and discrete control tasks with **train_discrete_task.py**.

Both of the scripts take the following four arguments:

1. environment name

2. agent name

3. state preprocessing

4. exploration policy

#### State preprocessing ####

Most of the environments don't need any specific state preprocessing. Pass the value **simple** in this case.

On the other hand, environments like **Atari** are easier to solve if several frames are stacked on top of each other

and preprocessed. In that case, pass the value **atari**.

### Contribute ###

You can contribute by create a new agent.

Each agent should have the following public methods:

* **learn**: perform a single training step

* **act**: generate an action given a state

* **perceive**: save a transition into the replay buffer (only Temporal Difference methods are supported for now)

* **log_scalar**: log scalar to a tensorboard summary; you can leave this method empty

* **save**: save the agent

* **close**: close the agent's session