js-gym

JavaScript environment for training reinforcement learning agents.

Installation

To download the code and install the requirements, you can run the following shell commands:

$ git clone https://github.com/bobiblazeski/js-gym.git
$ cd js-gym
$ npm install

Getting started

This code is intended to be run locally by a single user. The server runs in node.js.

To start the server from the command line, run this:

$ node server/start.js

If you have pretrained weights you could pass them

$ node server/start.js --kano=t04051134 --subzero=t04051134

You can open your browser at http://localhost:3000/

Sample algorithms

1. Random Play
2. Random Search
3. HillClimbing
4. Augmented Random Search
5. Deep Deterministic Policy Gradient

Environments

MK

Adaptation of https://github.com/mgechev/mk.js

Action space

Action is an object containing two keys, subzero & kano. Each key contains an array of 18 probabilities which represent possible actions for the users. The sum of all actions should be ~1.

The environment is stochastic, and uses weighted random choice to select a move for your agent. Unless you pass one hot action.

State space

47 floating numbers between 0 & 1

TetNet

Adaptation of https://github.com/IdreesInc/TetNet

Action space

Integer in the range of [0, 11).

State space

Javascript object containing information about the game.