Good prediction is necessary for autonomous robots to make informed decisions in dynamic environments. These predictions can be generated by models whose parameters are estimated from sampled sensory data. The sampling strategy for collecting this data strongly impacts realized model performance. Methods for curiosity incentivize exploration based on mathematical proxies for expected information gain which can be used to perform targeted sampling. These curious approaches to choosing actions allow a framework for learning predictive models which is not tied to accomplishing specific tasks enabling higher model generalization performance. In this presentation, I will present our recent research on curiosity starting with simulation results in video games. Then, I will present a robotic domain transfer problem in which we improve prediction performance with the use of curiosity-driven decision making. I will also present our preliminary work analyzing the effect of curiosity-driven decision making on multiagent system dynamics in a competitive resource allocation problem.
Bernadette Bucher is a second year computer science PhD student in the GRASP lab at University of Pennsylvania working with Dr. Kostas Daniilidis. Prior to joining GRASP, Bernadette held various positions at Lockheed Martin Corporation from 2014 to 2019, most recently working as a Senior Software Engineer. Her professional experience encompasses image processing, geospatial data integration, signal processing, and cyber/electronic warfare projects for multiple government customers. During that time, she completed graduate courses in computer science from Georgia Institute of Technology. She received her M.A. in Mathematics, M.A. in Economics, and B.S. in Mathematics and Economics from the University of Alabama in 2014. Her current research interests focus on the development of neuromorphic approaches to autonomous decision making with the goal of improving prediction performance through targeted sampling of the environment.