摘要： 计算机博弈程序难以处理局面估值问题。为此，结合时间差分算法和反向传播神经网络，设计一种局面估值算法BP-TD(?)，实现估值函数参数的自动调整。为提高博弈训练的性能，针对开局和中局，提出分阶段设置参数值的策略。以五子棋为应用背景，实现博弈系统RenjuTD。实验结果表明，该算法可使程序的博弈水平得到较大提高。

关键词: 计算机博弈, 差分学习, 反向传播神经网络, 估值算法, 增强学习, 博弈训练

Abstract: Situation valuation is the most difficult issue in all kinds of computer game programs. A valuation method named BP-TD(?) is presented combining temporal difference algorithm and back propagation neural network, which can solve the problem of adjusting the parameter values of valuation function. On this basis, in order to enhance the performance of game training, the strategy of setting different parameter values is proposed for opening and middle game phases. The game system RenjuTD is implemented using Renju as application background. Experimental results show the game level of program is significantly improved.

Key words: computer game, difference learning, back propagation neural network, valuation algorithm, reinforcement learning, game training

中图分类号: 

• TP18