Behavioral cloning for simulator validation

Abstract

Behavioral cloning is an established technique for creating agent behaviors by replicating patterns of behavior observed in humans or other agents. For pragmatic reasons, behavioral cloning has usually been implemented and tested in simulation environments using a single nonexpert subject. In this paper, we capture behaviors for a team of subject matter experts engaged in real competition (a soccer tournament) rather than participating in a study. From this data set, we create software agents that clone the observed human tactics. We place the agents in a simulation to determine whether increased behavioral realism results in higher performance within the simulation and argue that the transferability of real-world tactics is an important metric for simulator validation. Other applications for validated agents include automated agent behavior, factor analysis for team performance, and evaluation of real team tactics in hypothetical scenarios such as fantasy tournaments. © 2008 Springer-Verlag Berlin Heidelberg.