In-silico Behavior Discovery System

It is now widely accepted that animal behavior optimizes certain performance criteria. The question is, which criteria are being optimized? To answer this question, today, ethologists start by conducting some exploratory experiments. After the preliminary data is gathered, they infer a set of potential hypotheses. In order to find out which hypothesis is more likely to be correct, they have to carry out real animal experiments for each and every one of the hypotheses. This process could be repetitive, tedious, time consuming and costly. To alleviate all these difficulties, in this thesis, we propose an in-silico system – termed “in-silico behavior discovery” – to explore the feasibility of performing an in silico study on the underlying performance criteria. This study is distinct from animal dynamics and trajectory simulation as the focus is on the decision making strategy behind the observed motions, rather than the specific motions in space. Key to the system is the use of Partially Observable Markov Decision Processes (POMDPs) to generate an optimal strategy under a given hypothesis. POMDPs enable the system take into account imperfect information about the animals’ dynamics and the operating environment. Given multiple hypotheses and a set of preliminary observational data, the system will compute the optimal strategy under each hypothesis, generate a set of synthesized data for each optimal strategy, and then rank the hypotheses based on the similarity between the set of synthesized data generated under each hypothesis and the provided observational data. Using this system, the ethologists can focus on the most promising hypotheses first, thus reduce the number of animal experiments.