Research paper

Report

hive-Alphazero

About

Play with ranking Humman (AI play as white):

Dear Professor,

We have achieved a good performance model using search to play with humans. The current model has been trained on 3 datasets.

1. 42,000 games replay of human vs bot and human vs human
2. 32, 000 games self-play using MCTS, with 50 simulations per move
3. 3, 000 games self-play, with 250 simulations per move.

The latest model with 500 simulations per move, play as White wins the 254-ranking human player as Black.

Hive reinforcement learning by using Chess AlphaZero methods.

This project is based on these main resources:

1. The development of Chess-Alpha-zero by @Zeta36: https://github.com/Zeta36/chess-alpha-zero
2. The development of Chess-Alpha-zero using pytorch by @geochri : https://github.com/geochri/AlphaZero_Chess
3. Hive board game - python verion by @dboures: https://github.com/dboures/Hive
4. DeepMind's Oct 19th publication: Mastering the Game of Go without Human Knowledge.
5. DeepMind Chess AlphaZero: https://arxiv.org/pdf/1712.01815.pdf

Note

Environment

• Python 3.8.3
• pytorch

Modules

Supervised Learning

I trained the SL model with datasets from boardspace.net: http://www.boardspace.net/hive/hivegames/

Reinforcement Learning

This AlphaGo Zero implementation consists of three workers: self, opt and eval.

• self is Self-Play to generate training data by self-play using BestModel.
• opt is Trainer to train model, and generate next-generation models.
• eval is Evaluator to evaluate whether the next-generation model is better than BestModel. If better, replace BestModel.