Assessing learned models of fish schooling behavior

Agent-based simulation is a valuable tool for validating the theoretical models biologists and ethologists use to explain animal behavior. By automating the process of constructing agent-based models (ABM) directly from observation data we can enable researchers to focus more of their time on the analysis of the behaviors and animals in question. We present experimental results using a modified version of k Nearest Neighbor to learn an executable model of fish schooling behavior from both synthetic data and tracking data of live juvenile Notemigonus Crysoleucas, and quantitatively asses the performance of the learned behavior. Our experiments illustrate that our method can successfully learn fish schooling, and provide an objective criteria for comparing competing biological theories of behavior.