This is a project for COMP34111: AI and Games, the aim of this project is to develop a Hex AI agent that can play Hex game with human players.
Our group won all marking bots and achieved a very high reaction speed. We got the best performance on onaverage.
More details can be found below:
One unofficial test example is shown below:
You can also check the repo issue for more test results, we have uploaded several performance logs for test, test scripts are also provided.
Please check Hex Game Engine Documentation.pdf in the root directory for more details.
We achieved our goal by 3 Stages in different branches:
- We implemented the naive Monte Carlo Tree Search (MCTS) with UCT for Hex, along with a neural network to enhance the evaluation function(Check MCTS_Agent).
- Additionally, we developed a simple opening book and a parallelized version of MCTS to improve efficiency.(Check MCTS_Beta)
- To further improve our performance, we deployed a Aplha-Zero pre-trained model and fine-tuned it to guide the MCTS search. (Check AlphaZero)
by the way, one stupid training process is shown below:
:)
- KataGo: https://github.com/lightvector/KataGo
- Azalea, a pre trained model for Hex: https://github.com/jseppanen/azalea