A Go A.I. based on MCTS WITHOUT Neural Networks
FoolGo is a Go A.I. based on MCTS (Monte Carlo Tree Search).
It shows the beginner level on the 9x9 board currently, for it does not use neural networks, but only the simple UCT (UCB for Tree) algorithm and Monte Carlo simulation.
The underlying data structure and algorithms are quite efficient. Specifically, it runs 40 thousand Monte Carlo games per second on my MacBook Air 2014.
The key ideas of the Go game's implementation are as follows:
For MCTS, we use multi-thread search on CPU. In detail, we simply prohibit a thread from exploring nodes that are being explored by peer threads.
This project is highly readable and extensible due to its well-designed OOP design. Some important classes are listed as follows:
No open issues yet, or sync has not completed.