DARPA Integrated Sensing and Processing (ISP) Program. Approximation Methods for Markov Decision Problems in Sensor Management

摘要

This work addresses problems of sensor resource management (SRM) in which one or more sensors obtain measurements of the state of one or more targets. For example, an airborne radar may be attempting to track several ground targets, which are sometimes stationary (requiring a synthetic aperture radar mode) and sometimes moving (requiring a ground moving target indication radar mode). The challenge is to schedule the radar modes as the scenario evolves. Such problems can generally be formulated as partially observable Markov decision processes (POMDPs), which can express essential characteristics of the SRM problem such as uncertainty and dynamics. This work emphasizes a farsighted approach; the highest long-term payoff may not be generated by the action providing the highest immediate payoff. Accomplishments of this effort include the establishment of a boundary on optimal SRM performance, analysis of farsighted SRM strategies for controlling a multimode sensor, and the derivation of a novel set of sufficient conditions for optimality in Markov decision processes.