Predictive Models and Monte Carlo Tree Search: A Pipeline for Believable Agents

Abstract

Developing and assessing believable agents remains a sought out challenge. Recently, research has approached this problem by treating and assessing believability as a time-continuous phenomenon, learning from collected data to predict believability of games and game states. Our study will build on this work: by integrating this believability model with a game agent to affect its behaviour. In this short paper, we first describe our methodology and then the results obtained from our user study, which suggests that this methodology can help creating more believable agents, opening the possibility of integrating this type of models into game development. We also discuss the limitations of this approach, possible variants to tackle these, and ideas for future work to extend this preliminary work.