Efficient data collection strategies for rapid learning in physical environments

Abstract

With the ubiquity of intelligent systems capable of sensing, inferring and acting upon their surroundings, it becomes critical to learn rapidly about unknown systems or environments. However, obtaining empirical data is often costly and involves setting up time consuming experiments or deploying specialized sensors. We are interested in deriving scalable algorithms and system architectures that facilitate efficient data collection, maximizing inference quality under limited resource budget. In this work, we consider efficient data collection strategies in several applications involving physical environments. We study the problem of learning dynamical systems with initial approximated models, where we prescribe methods for choosing near optimal experimental parameters to collect empirical data. We study the problem of antenna array topology design where we prescribe configurations allowing efficient scene inference under various measurement schemes and budget constraints. We introduce a novel nonlinear radar modality and discuss efficient design techniques for this setting. Finally, we introduce a novel methodology for optical imaging of non line of sight hidden scenes by utilizing occlusions and investigate how to achieve efficient illumination of the scene for fast hidden target interrogation.