Nonparametric and Information Theoretic Approaches for Sensor Resource Management

• Activity Number: 299
• Type: Invited
• Date/Time: Tuesday, August 9, 2005 : 2:00 PM to 3:50 PM
• Sponsor: Section on Nonparametric Statistics
• Author(s): John W. Fisher*+ and Jason L. Williams and Alan S. Wilsky
• Companies: Massachusetts Institute of Technology and Massachusetts Institute of Technology and Massachusetts Institute of Technology
• Keywords: Information Theory ; Sensor Networks ; Resource Management

Abstract

Estimates of statistical information measures play an increasingly important role in sensor resource management. This is particularly true for intelligent sensor networks in which the energy costs of intersensor communications are significant relative to the energy costs of local computation. Additionally, as a consequence of nonlinear measurement models, uncertainty is naturally characterized using nonparametric models and Markov chain Monte Carlo methods for state estimation. This naturally leads to a tradeoff between the information content of distributed measurements, characterized by conditional entropy, versus the cost of communicating those measurements. We formulate this tradeoff within an approximate dynamic programming approach using finite time horizons called roll-out methods. A critical issue in any of these approaches is the need to estimate statistical information measures (or bounds) from high-dimensional measurements. We present information bottleneck bounds that provide guidance for setting optimization time horizons and methods for estimating information measures from high-dimensional measurements.