Artificial Intelligence on Board Game with Reinforcement Learning

Abstract

In the most resent years, the Arti cial Intelligence (AI) technology has been one of the most attractive elds in computer science and information technology, and various AIs are developed by numerous developers, researchers and institutions. Among all methodologies and implementations that are introduced in AI technology, reinforcement learning is also one of the most famous path to develop high performance AI. This project mainly aims on implementing and optimizing an AI algorithm using reinforcement learning based on the idea of MuZero algorithm to play board games like Reversi and Connect4 with as minimal computing resource as possible. For the algorithm part, this project will implement Convolutional Neural Networks (CNN) in conjunction with Monte Carlo Tree Search (MCTS). Apart from the work based on previous research regarding the AI algorithm, this project also tries to implement a self-playing program which can help with checking the ability of the AI by automatically playing board games against human players on online platforms using various techniques regarding computer vision and date transformation. The self-playing program is to be developed individually using computer vision technologies together with mouse simulations. This project uses the result of the AI's competing against human players as the benchmark to assess the ability and performance level of the algorithm.