Time-Evolution Characterization of Behavior Class Prototypes

Abstract

Neuroscientists and roboticists alike are interested in how both the external cues and internal states contribute to determining behavioral sequences. As the motivation drivers change through time, so does the animal switch between alternating activities and outward exhibition of patterned behavior. It is accepted that underlying neural integration of internal states - as well as incoming information of the external environment - give rise to the adaptive abilities of animal behavior in different contexts. Here, the sequences of hunting-motivated behavior of praying mantises were modeled as Markov chains, with each sequence giving rise to a corresponding transition probability matrix. From these transition matrices, three methods of prototype generation were used - cumulative, centroid, and medoid - to produce categorical representatives of the time series data of all five feeding states used in the experiments. Novel to this paper, is the use of Markovian chain metrics to compare the efficacy of these prototypes at capturing the time-evolution behavior unique to each feed state. Results show that the cumulative prototypes best exhibited temporal behaviors most consistent with the real data.