Human behaviour modelling in games using agents

Abstract

The incorporation of real human behaviour for nonplayer characters (NPCs) in games is a significant research challenge.The nature of events and actions within a game scenario is associated with temporal and spatial redundancies. Such data cannot be considered in isolation, and should be processed in sequence. Multi Agent Systems are proposed as a potential technology to model natural human behaviour into non player characters in game environments. A three phase solution is proposed for this purpose. During the first phase real data in a game are collected by letting ahuman player play the game unconditionally and naturally as possible. All objects around the player and each action executed by the player are recorded against the time. The second phase involves the cleaning and normalization of the noise present within datagathered during phase 1. Then for each sequence, data is modelled as HMM along with the player’s current action. Rules are generated with these HMM and player actions. The NPC decision making strategy, consequently are implemented by agents using these rules. HMMs are mapped as resource agents and the NPC in the game is mapped as a request agent. Based on the observation sequence that NPC receives, it broadcasts this information to the resource agents and bids for the best deal and tries to find the most suitable state sequence (that is action sequence). Jade is used as the agent framework and Jahmm is used for HMM modelling. As the test bed (simulation platform) of this experiment, Virtual Battle Space2 (VBS2) along with the VBS2Fusion library is used to prototype the approach. The validity of the proposed method is verified by experiments and analysis by statistical methods. Odds theory is used as one of the statistical methods. Human control missions as well as NPC control missions are recorded as video files. Then arbitrary observers who haven’t played this game are used to observe the videos and let them to recognize what is the mission that real player plays and what is the mission that NPC is controlling. According to the results of the observers, it is statistically proven by showing that there is no significant discrimination between real player video and NPC video.