RTS游戏中用户行为的神经网络预测模型

Neural network-based behavior prediction models for RTS game players

即时战略游戏(简称RTS游戏)中,用户的行为由于游戏自身庞大的决策空间而难以预测。针对这个问题,提出了通过对RTS游戏的对战记录进行分析,建立5种结构的神经网络模型来预测用户行为的方法。模型考虑了不同时间片的状态对于决策行为的影响,设计了单时间片输入和双时间片输入的神经网络,并与基于动态贝叶斯网络的模型进行了比较。实验结果表明,基于单时间片输入的神经网络模型能够更加快速地完成训练过程并达到满意的预测准确度。

Due to the large decision space of real time strategy（RTS） games,game players＇ behaviors are difficult to predict especially when facing different situations.In this case,five kinds of neural network models are constructed by analyzing RTS game data to predict players＇ behaviors in different states.Considering the influence of time slices on player behaviors,neural networks are designed respectively with single time slice input and double time slice input,and then compared their prediction performance with dynamic Bayesian network based model.Experimental results show that the model with single time slice input is the most efficient one and can achieve satisfactory prediction accuracy.