SJTU ACM Class Machine Learning 2021 Assignment ( CS420 Course Project )
A superhuman Board Game AI with the AlphaZero algorithm.
- Achieved super-human performance in 13x13 Gomoku, 9x9 Go and 9x9 Othello.
- Train only via self-play with the fairly limited computational power of a single home PC.
- Accelerate training by reducing the upper limit of performance that the AI can achieve, but still be able to beat top human players.
- Playout Cap Randomization, which is a new technique introduced by KataGo that accelerates training.
- Temperature Decay, which is a search technique that trades off randomness and quality of the play.
- Random Freeze, which plays a important role in unbalanced games like Gomoku.
- MobileNet + SENet with ConvNeXt’s tricks, which allows MCTS to do more playouts to achieve higher performance while ensuring a significantly fast training speed.
- Multi-threaded self-play agents.
- MCTS implemented in C++, Neural Network implemented in Python, working together by pybind11.
On the 13x13 Gomoku board (Swap after 1st rule), after a week of training with RTX 3080, it reached the same level as Rapfi, which is one of the strongest alpha-beta Gomoku AI in the world.
