rllib.js

Reinforcement learning library with JavaScript.
It provides some reinforcement learning algorithms and environments.
At this moment uses TensorFlow.js to implement functions and three.js for visualization.

Live example

Visual library: three.js

Motivation

JavaScript is a very popular programming language but JS developers have no instruments of RL.
This library was created to correct the situation. Web game makers can use it to train game bots in 2D and 3D environments.
Other can find it useful to make experiments.

Unity developers can use ML methods with ml-agents

Install and start example

Starting example:

Windows

On Windows you need preinstalled MS Visual Studio Community edition for build 'gl' library.

Linux, Windows

git clone https://github.com/polyzer/rllib.js.git  
npm i   

Hungry world 2D example using PPO algorithm in js

npm run ppo-hungry-js

Hungry world 2D example using PPO algorithm with Parcel

npm run ppo-hungry-jsm

Hungry world example using PPO algorithm with Parcel

npm run ppo-hungry3d-jsm

Hungry world example using PPO algorithm Node.js training

You can start Node.js example with:

npm run ppo_node

Algorithms:

Deep Q-learning:

• Base Deep Q-learning
• Dueling Q-learning
• Double Q-learning

Policy gradient:

• REINFORCE
• Actor-Critic
• PPO
• TRPO

Environments:

Without physics:
Flat world 3D. Discrete and continuous versions.

• Agent have 10 eye-detectors.
• Target: Agent trying to learn eating only green items if he see them.
• Actions: turns on specified angles.
• Space: Agent recieves signal from each eye that specifies type of saw item.

Known issues:

Node.js

symbol lookup error on nVidia

It could be helpful:

npm rebuild --build-from-source gl