Rubix

Rubix is a deep reinforcement learning Rubik's Cube solver written in Jax and Haiku.

Environment

The custom environment is developed in the style of the environments in Jumanji.

Disclaimer: Since starting this repo, InstaDeep has brought out Jumanji 0.2.0 which contains a RubiksCube environment.

Agents

This repo currently supports DQN, QR-DQN and a discretized PPO agent. The implementations of the DQN-based agents are inspired from the DQN Zoo implementations.

Code Structure

TBA

How to use

Requirements

All dependencies can be installed using:

pip install -r requirements/requirements.txt

Training

The specific model can then be trained by running the train file.

Below is an example with the DQN agent on a 5x5 Rubik's Cube;

python train.py --agent=DQN --cube_dim=5

TO DO

This repo is still in the early stages - below are some things I am currently working on

1. Add more agents
2. Set up a predict capacity (user input capacity)
3. Report best results for each agent/cubesize

Reference