How to compare usability of techniques for the specification of virtual agents' behavior? An experimental pilot study with human subjects
Abstrakt
Reactive or dynamic planning is currently the dominant paradigm for controlling virtual agents in 3D videogames. Various reactive planning techniques are employed in the videogame industry while many reactive planning systems and languages are being developed in the academia.
Claims about benefits of different approaches are supported by the experience of videogame programmers and the arguments of researchers, but rigorous empirical data corroborating alleged advantages of different methods are lacking. Here, we present results of a pilot study in which we compare the usability of an academic technique designed for programming intelligent agents' behavior with the usability of an unaltered classical programming language.
Our study seeks to replicate the situation of professional game programmers considering using an unfamiliar academic system for programming in-game agents. We engaged 30 computer science students attending a university course on virtual agents in two programming assignments.
For each, the students had to code high-level behavior of a 3D virtual agent solving a game-like task in the Unreal Tournament 2004 environment. Each student had to use Java for one task and the POSH reactive planner with a graphical editor for the other.
We collected quantitative and qualitative usability data. The results indicate that POSH outperforms Java in terms of usability for one of the assigned tasks but not the other.
This implies that the suitability of an AI systems-engineering approach is task sensitive. We also discuss lessons learnt about the evaluation process itself, proposing possible improvements in the experimental design.
We conclude that comparative studies are a useful method for analyzing benefits of different approaches to controlling virtual agents.