Sensor Selection for Active Information Fusion

Active information fusion is to selectively choose the sensors so that the information gain can compensate the cost spent in information gathering. However, determining the most informative and cost-effective sensors requires an evaluation of all possible sensor combinations, which is computationally intractable, particularly, when information-theoretic criterion is used. This paper presents a methodology to actively select a sensor subset with the best tradeoff between information gain and sensor cost by exploiting the synergy among sensors. Our approach includes two aspects: a method for efficient mutual information computation and a graph-theoretic approach to reduce search space. The approach can reduce the time complexity significantly in searching for a near optimal sensor subset. Introduction There has been a great deal of interest in the development of systems capable of using many different sources of sensory information (Waltz & Llinas 1990). In many applications, e.g., battlefield situation assessment, a number of information sources can be generated, but they are often constrained by limited time and resources. The more sensors we use, the more information we can obtain. On the other hand, every act of information gathering incurs the cost of utilizing those sensors, e.g., computational cost, operation cost, etc. In order to efficiently provide information to a decision-maker, it is important to avoid unnecessary or unproductive sensor actions. Thus, we must actively select a subset of sensors that are the most informative yet cost-effective. An important issue is how to determine a subset of sensors that is worth to be instantiated at particular stage of information gathering. There are numerous applications of sensor selection including computer vision (Paletta & Pinz 2000; Denzler & Brown 2002), control systems (Miller & Runggaldier 1997; Logothetis & Isaksson 1999) and sensor networks (Zhao, Shin, & Reich 2002; Ertin, Fisher, & Potter 2003), etc. The strategies of sensor selection can be broadly classified into two categories: search-based approach and decisiontheoretic approach. The search-based approach regards sensor selection as a search problem to find the best soluCopyright c © 2005, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. tion among all possible sensor combinations. Kalandros et al. (Kalandros, Pao, & Ho 1999) proposed a superheuristics method, which begins with a base sensor combination and then generates alternative solution via random perturbations to the initial combinations. Fassinut-Mombot et al. (Fassinut-Mombot & Choquel 2004) proposed an entropy adaptive aggregation approach. After heuristically obtaining the initial subset, they iteratively aggregate and disaggregate the current subset until it converges. However, the search-based approach is computationally expensive due to the combinatorial search space. The decision-theoretic approach regards sensor selection as a decision-making problem. Kristensen (Kristensen 1997) treats the problem of choosing proper sensing actions as decision-making, and a decision tree is used to find the best sensor action at each step. Castanon (Castanon 1997) formulates the problem of dynamical scheduling of sensor for multiple object classification as a partially observed Markov decision process. However, it suffers from combinatorial explosion for a problem even in moderate size. Krishnamurthy (Krishnamurthy 2002) used dynamic programming to find an optimal sensor in a Hidden Markov model; while the approach is feasible only for the problems with a small number of sensors. This paper focuses on the sensor selection problem with information theoretic approach, where the selection criterion is defined as a mixture of both expected information gain and cost. However, there are two difficulties to use this criterion. First, the computation of higher order mutual information (information gain) generally requires time exponential in the number of sensors to compute information gain exactly. Second, selecting k sensors out of n sensors is also a NP-hard problem. These difficulties are the impediments for real time application. For a fusion system containing many sensors, it is practically infeasible to evaluate all sensor subsets. To avoid the computational intractability of exact computation of information gain, myopic approaches are often used (Oliver & Horvitz 2003). The myopic procedure assumes that the decision maker will act after observing only one sensor. However, we should consider the fact that at each time the decision maker may observe multiple sensors before acting. So, Hecherman et al. (Heckerman, Horvitz, & Middleton 1993) presented an approximate nonmyopic computation for value of information by exploiting the statistical properties of large samples; while the approach is