reversi-game-AI-MCTS: A Go repository from eamonn-zh

Reversi Game AI Based on Concurrent Monte Carlo Tree Search (MCTS)

1. Introduction

This project implements an AI for Reversi game, it is based on Monte Carlo tree search (MCTS) algorithm, which is used in AlphaGo AI. Algorithm runs on the Python backend server, and also there is a web frontend, where users can play the Reversi games against the AI. Frontend and backend interact through REST APIs.

The secret to the success of the MCTS algorithm is to simulate as many games as possible in a limited time. So the algorithm is implemented in Go and is called by Python through the foreign function interface. In addition, I made this this algorithm concurrent.

2. Languages & Communication

Languages:
1. Go
2. Python
3. JavaScript
4. Bash
Communication methods:
1. Foreign function interface (between Python and Go through ctypes)
2. REST APIs (between JavaScript and Python)

3. Project Structure

cookbooks/ --- Chef configurations
FastAPI_Server/ --- high-performance Python API backend server (similar to Django)
Go_MCTS/ --- Golang program (implemented concurrent Monte Carlo tree search algorithm)
Web_Reversi/ --- frontend web page (implemented via Vue.js, JavaScript code is in Web_Reversi/src/App.vue)
Vagrantfile --- vagrant configuration

4. How to Run

# start the VM
vagrant up

# run the provisioner
vagrant provision

Then open your web browser and access http://localhost:8080/

Notes: Due to unknown reasons, Chef cannot run the Node.js web server in background with command npm run serve &, so I use npm run serve instead, which will stall the bash when execute vagrant provision, it does not affect the operation of the program. When you see:

==> default:                 App running at:
==> default:                 - Local:   http://localhost:8080/ 
==> default:                 - Network: http://10.0.2.15:8080/
==> default:               
==> default:                 Note that the development build is not optimized.
==> default:                 To create a production build, run npm run build.
==> default:

Now the project are running, just visit http://localhost:8080/.

5. Concurrent Limitation

MCTS requires a lot of calculations, but the virtual machines has limited performance, so running MCTS concurrently on poorly performing hardware is not always good. To solve this problem, I implemented two versions of MCTS, with or without concurrency, I defined a flag in my code (at Go_MCTS/src/main/MonteCarloTreeSearch.go, line 14) and the default version is without concurrency. You can manually switch the flag to turn on / off concurrency:

// open concurrent mode
var concurrent = false