Wave-based signal processing

This paper addresses the problem of adaptive sensing based on multiple sensor modalities. Assuming a multi-modality sensing problem is considered, that physical insight is available, and that a certain quantity of data has been collected thus far from the scene under test. The objective is to optimally choose what new multi-sensor data should be collected, with the objective of maximizing classification performance while minimizing sensing costs (e.g., battery use, time, etc.). It is desirable that the algorithm be non-myopic, in the sense that it accounts for the immediate utility of a given sensor, as well as the properties of sensing over a discounted infinite horizon. For example, a sequence of inexpensive measurements may have the same utility as a single expensive measurement, and the inexpensive measurements are only preferred if the algorithm operates non-myopically (i.e., the algorithm addresses long-term performance, looking ahead, rather than myopically and greedily choosing the sensor that yields the immediate best performance, independent of cost). The problem is solved via a partially observable Markov decision process (POMDP), and this paper explains how the underlying wave physics is employed to improve POMDP sensing performance. The basic framework has been applied successfully to several sensing scenarios, and this paper focuses on the specific problem of multi-modality sensing of buried land mines