An Interpretable Measurement for Playing Archetypes of Driving Agents in Video Games

Abstract

Analysis and evaluating the behavior of different none-player characters (NPCs) plays a significant role in improving the playing experience of a video game. Furthermore, different playing archetypes can be acquired by changing the configuration of NPCs. Hence, by having an interpretable measurement for playing archetypes of NPCs, a wide range of behavior can be attained. Additionally, the aforementioned measurement gives insight of how the changes in the configuration of the NPC changes its behavior. The proposed method can significantly increase the interpretability of the NPCs and ease the designing of desired NPCs with appropriate behavior. In this paper, by comparing the behavior of different NPCs with that of the previously gathered data, cross-correlation as a similarity measurement is used to interpret the changes in the behavior of the NPCs, given their configurations. The proposed method can help game designers evaluate the performance of their NPCs. It is shown that by comparing the gathered data from the state of NPCs with that of users, useful information for determining the driving style of a NPC can be acquired. Afterwards, by setting references for behavior patterns, new NPCs can be evaluated and classified into different playing archetypes. Moreover, in this paper the importance of the cross-correlation values and also the fitted Gaussian model on this signal, in the behavior of different NPCs is discussed in detail. The test bed for the proposed method is a driving game in Unity game engine. © 2020 IEEE.