Does high-level behavior specification tool make production of virtual agent behaviors better?

• Authors: Jakub Gemrot;Zdeněk Hlávka;Cyril Brom
• Venue: CAVE'12 Proceedings of the First international conference on Cognitive Agents for Virtual Environments
• Year: 2012

Part of Academia's motivation for improving agent-based languages and architectures is the hope that some of its solutions might be, one day, adopted by the videogame industry for specification of agents' behavior. However, the games industry already employs its own techniques for that purpose. Thus we need rigorous empirical data to provide an insight to the expected utility of academic solutions and pinpoint their most promising features. Game pro-grammers often face situations when they have to understand and modify the work of current or former colleagues, or to extend their own work from months or even years ago. Thus, one way an academia's solution could provide value would be that it outperforms the industry's typical approach under these circumstances --- offering better code readability and maintainability. Here we present results of an empirical study modeling this problem. We adopt Java programming as the industry approach (modeling scripting) and choose the POSH reactive planner as an academic approach. We engaged 22 computer science students attending a university course on virtual agents on two programming assignments, in which they had to produce specific high-level behaviors of 3D virtual agents solving different game-like tasks. First, students had to produce the behavior of a particular task from scratch. Second, these behaviors were used in an assignment where students had to extend the behavior coded by someone else. Finally, three months later, 8 of these students were told to extend their own behavior they coded in the first assignment. The quantitative results indicate that Java is as good as POSH in terms of subjective programmers' preference and that there is no objective difference between qualities of created behaviors. The qualitative results suggest several useful but also troublesome features of POSH, some of which are shared by other languages, suggesting possible improvements.