基于MCTS和卷积神经网络的五子棋策略研究

Research on Gobang Strategy Based on MCTS and Convolutional Neural Networks

随着AlphaGo的诞生,人机对弈和人工智能再次成为研究热点。传统的MCTS(蒙特卡洛树搜索)虽然在迭代搜索方面已有良好的成效,但由于五子棋搜索空间较大,算法极易陷入局部最优化问题,且耗时严重。我们用MCTS和卷积神经网络上设计的策略系统,让其与MCTS进行训练(self-play),使五子棋的策略系统能在一定时间内对自身进行升级,然后又回来继续训练自身,这样得到的五子棋策略系统不仅比传统的MCTS更具有即时性,棋力也更强。

With the birth of AlphaGo,man-machine game and artificial intelligence have once again become research hotspots.Although the traditional MCTS(Monte Carlo Tree Search)has a good effect in iterative search,because the search space of Gomoku is large,the algorithm is easy to fall into the local optimization problem,and it takes time.We use MCTS and the convolutional neural network to design a strategy system to let it train with the MCTS(self-play),so that the Gomoku’s strategy system can upgrade itself within a certain period of time,and then come back to continue training itself.The Gobang strategy system is not only more immediacy than the traditional MCTS,but also much stronger.