This repository has been archived by the owner. It is now read-only.
Deep Reinforcement Learning for the JVM (Deep-Q, A3C)
Clone or download
Latest commit 9b5238d May 21, 2018

README.md

IMPORTANT: THIS REPOSITORY HAS BEEN ARCHIVED AND RL4J HAS BEEN MOVED

RL4J has been migrated to a sub-module of the mono-repository here: https://github.com/deeplearning4j/deeplearning4j All future development will continue at that repository, which should be used for all issues and pull requests.

RL4J

RL4J is a reinforcement learning framework integrated with deeplearning4j and released under an Apache 2.0 open-source license. By contributing code to this repository, you agree to make your contribution available under an Apache 2.0 license.

  • DQN (Deep Q Learning with double DQN)
  • Async RL (A3C, Async NStepQlearning)

Both for Low-Dimensional (array of info) and high-dimensional (pixels) input.

DOOM

Cartpole

Here is a useful blog post I wrote to introduce you to reinforcement learning, DQN and Async RL:

Blog post

Examples

Cartpole example

Disclaimer

This is a tech preview and distributed as is. Comments are welcome on our gitter channel: gitter

Quickstart

** INSTALL rl4j-api before installing all (see below)!**

  • mvn install -pl rl4j-api
  • [if you want rl4j-gym too] Download and mvn install: gym-java-client
  • mvn install

Visualisation

webapp-rl4j

Quicktry cartpole:

Doom

Doom is not ready yet but you can make it work if you feel adventurous with some additional steps:

  • You will need vizdoom, compile the native lib and move it into the root of your project in a folder
  • export MAVEN_OPTS=-Djava.library.path=THEFOLDEROFTHELIB
  • mvn compile exec:java -Dexec.mainClass="YOURMAINCLASS"

Malmo (Minecraft)

Malmo

  • Download and unzip Malmo from here
  • export MALMO_HOME=YOURMALMO_FOLDER
  • export MALMO_XSD_PATH=$MALMO_HOME/Schemas
  • launch malmo per instructions
  • run with this main

WIP

  • Documentation
  • Serialization/Deserialization (load save)
  • Compression of pixels in order to store 1M state in a reasonnable amount of memory
  • Async learning: A3C and nstep learning (requires some missing features from dl4j (calc and apply gradients)).

Author

Ruben Fiszel

Proposed contribution area:

  • Continuous control
  • Policy Gradient
  • Update gym-java-client when gym-http-api gets compatible with pixels environments to play with Pong, Doom, etc ..