Dynamic difficulty adjustment by fuzzy rules using in a neural network controlled game

This paper describes a series of experiments using the offline trained artificial neural networks (ANN). The ANN acts as an embedded game agent in a shooting game to control the nonplayer character (NPC). The training datasets of ANN are constructed by three different levels of players (expert, medium and beginner players). And then the three different levels training datasets are used to train three different level´s ANN, respectively. Meanwhile, the optimal neurons of the hidden layer and the suitable period of training time is obtained by the method of three fold cross validation. In addition, a comparison between ANN and two traditional game AI - finite state machine (FSM) and computer random controlled method, is also implemented in this study. The simulated results show that ANN can get better winning rate than FSM and random method. Meanwhile, ANN obtains a pretty good human-like simulation results. Finally, a fuzzy rules-based approach is utilized to do the dynamic game difficulty adjustment. The experimental results show that the adaptive mechanism developed in this study could dynamic balance the equilibrium of game difficulty. All these, enhance the replayability of the game.