Active Sensing: Theories and Algorithms

Thursday, April 19, 2012 - 11:00am - 12:00pm EDT

  • Location:Dibner Hall, LC400
    Two Metrotech Center, Brooklyn, New York, US

Speaker: Professor David Castañón 

Host Faculty: Professor Zhong-Ping Jiang

Abstract

Advances in embedded computing have introduced a new generation of sensors that have the capability of adapting their sensing dynamically in response to collected information. In this talk, we discuss approaches to sensor control problems for the purposes of locating and classifying multiple unknown objects. First, we pose the problem as a stochastic control problem using a simple model of sensor observations to evolve the underlying information state to desired conditions. The resulting optimization problem is a partially observed Markov decision process (POMDP) with combinatorially large action and state spaces. We develop a relaxation of this problem that yields a new POMDP that yields a lower bound to an optimal performance, and show that this POMDP can solved effectively using decomposition techniques. The solution of this approximate problem is used to develop approximate optimal solutions to the original problem.
We subsequently expand our sensor model to include extracted features with aspect-dependent visibility constraints, typical of imaging sensors. We present a random set model of observed features, and use this to develop information fusion algorithms from multiple sensors and multiple times. This framework is used as the foundation for development of algorithms that manage sensor networks to classify unknown objects from their features. In particular, we develop information theory bounds on classification performance between random sets that can be used as the basis for fast algorithms for real-time management.
Time permitting, I’ll also discuss some of our recent work on active learning for boosting algorithms, focusing on the problem of which training data is worth using to train a binary classifier.

About the Speaker

Dr. David A. Castañón is Professor and Chair of Electrical and Computer Engineering at Boston University. He received his B.S. degree in Electrical Engineering from Tulane University in 1971, and his Ph.D. degree in Applied Mathematics from the Massachusetts Institute of Technology in 1976. From 1976 to 1981, he was a research associate with the Laboratory for Information and Decision Systems at MIT. From 1982-1990, he was Chief Scientist at Alphatech, Inc. in Burlington, MA. He has been with Boston University since 1990. Prof. Castañón’s research interests include stochastic control, estimation, optimization and image understanding. He is co-director of Boston University’s Center for Information and Systems Engineering, Deputy Director of the NSF Research Center on Subsurface Censing and Imaging Systems and Associate Director of the DHS ALERT Center of Excellence on Explosives Detection and Mitigation. Prof. Castañón has served as associate editor for the Transactions on Automatic Control, and Computational Optimization and Applications. He served in numerous capacities for the IEEE Control Systems Society, including President, and serves on the IEEE Society Review Committee and the IEEE Prize Papers Awards Committee. He is a Senior member of IEEE, a member of SIAM and INFORMS, and served on the Air Force Scientific Advisory Board.