Leslie-Fang / ai1901-connectfour

强化学习(Reinforcement Learning) 在conect4 棋上的应用

Introduction

conect4 游戏规则： 在6行7列的棋盘上轮流下棋，可以选择任意空的位置下棋，但是如果选择的位置下方是空的，下在该位置上的棋子会下落到底部。 最先在横向，竖向以及对角方向连出4个棋子的一方获胜。

Based on this repo, we add new stragety with RL(Reinforcement Learning)

git clone https://github.com/rmit-s3664000-cameron-sugiada/ai1901-connectfour.git

强化学习(Reinforcement Learning) 在conect4 棋上的应用实现参考了这个博客

Setup

The code was written in Python 3.6.

pip install pipenv
pipenv install --skip-lock --dev
pipenv install
pipenv shell

Running

The currently available player types are:

• HumanPlayer - Player is controlled by user via GUI [DEFAULT OPTION]

• RandomAgent - Player is controlled by computer and just chooses random valid columns to place token

• MonteCarloAgent - Player is controlled by computer and uses Monte Carlo Tree Search to find a good move 著名围棋Ai Alpha Go，核心的地方在于在突破传统的蒙特卡洛搜索树，采用预先训练好的【价值评估网络】和【走子网络】两个神经网络来评估，缩小了每层的搜索量，并且配以Google的TPU，实现超越人类棋手。

• StudentAgent - Currently, is a kind of stragety which we implemented

• RLAgent - Currently, is a the implementation of our RL algorithm(minmax). Alpha-beta剪枝算法是minmax算法的升级版本，减小运算量

Human VS Human

python -m connectfour.game --player-one HumanPlayer --player-two HumanPlayer

Robot VS Human

python -m connectfour.game --player-one HumanPlayer --player-two RLAgent

大部分情况下，我已经下不过它了

No GUI

python -m connectfour.game --player-one MonteCarloAgent --player-two StudentAgent --no-graphics --fast --auto-close

Calculate Win rate

python runtest.py

Raw RL method VS MonteCarloAgent: 99% (most times is 100%)