Ph.D Proposal by Brian Hrolenok
Title: Constructing and Evaluating Executable Models of Collective Animal Behavior Brian Hrolenok School of Interactive Computing College of Computing Georgia Institute of Technology Date: May 13, 2015 (Wednesday) Time: 10am-12pm EST Location: College of Computing Building room 345 Committee:-------------- Dr. Tucker R. Balch (Advisor, School of Interactive Computing, Georgia Tech) Dr. Aaron Bobick (School of Engineering and Applied Science, Washington University in St. Louis) Dr. Byron Boots (School of Interactive Computing, Georgia Tech) Dr. Magnus Egerstedt (School of Electrical and Computer Engineering, Georgia Tech) Dr. Greg Turk (School of Interactive Computing, Georgia Tech) Abstract:-----------Multiagent simulation (MAS) can be a valuable tool for biologists and ethologists studying collective animal behavior, as in the case of our recent work on recovering the social structure of a group of rhesus monkeys. However, constructing models for simulation is often a time-consuming manual task. Current state-of-the-art multitarget tracking algorithms can now provide high accuracy, high density tracking data of groups of research animals. Techniques from machine learning should be able to leverage the wealth of information such data provides in order to automatically find good models of animal behavior that can be executed in simulation. However, models trained using traditional single-step loss functions can lead to behaviors that are qualitatively dissimilar to the behavior of real animals. Expectation Maximization (EM) methods computed over full trajectories are subject to local suboptima. These problems are particularly compounded in the case of multiple interacting animals. In this work we propose a novel quantitative evaluation method that addresses these issues. Our method combines measures of statistical similarity among a number of high-level characteristics of behavior, which we call behavioral divergence. In this work, we propose to use this divergence as a cost function for globally optimizing the model parameters. By optimizing for our evaluation criteria directly we should achieve better performance than single-step loss, and by using global optimization techniques we should avoid local optima. This approach can be integrated with different model classes, and we propose to explore k-NN, linear regression, and Binary-switched Input Output Hidden Markov Models (BIOHMM) which we have previously developed. The proposed work differs significantly from methods that attempt to optimize predictive capability of learned models. We intend to compare the performance of our proposed approach with current state of the art in time-series modeling and prediction methods. Specifically, we will analyze their predictive performance as compared with behavioral divergence as described above. We will present results on the schooling behavior of fish and the foraging behavior of ants, both synthetic and from research animals.
- Workflow Status: Published
- Created By: Tatianna Richardson
- Created: 05/07/2015
- Modified By: Fletcher Moore
- Modified: 10/07/2016