Abstract— Information-driven sensor management aims at making optimal decisions regarding the sensor type, mode and configuration in view of the sensing objectives. In this paper, an approach is developed for computing two information-theoretic functions, expected discrimination gain and expected entropy reduction, to optimize target classification accuracy based on multiple and heterogeneous sensors fusion. The measurement process is modeled by means of Bayesian networks (BNs). The two objective functions utilize the BN models to represent the expected effectiveness of the sensors search sequence. New theoretic solutions are presented and implemented for comput- ing the objective functions efficiently, based on the BN factorization of the underlying joint probability distributions. Dempster-Shafer fusion rule is embedded in the computations in order to account for the complementarity of multiple, heterogeneous sensor measurements. The efficiency of the two objective functions is demonstrated and compared using a landmine detection and classification application. I. INTRODUCTION HE problem of information-driven sensor planning and management for target classification consists of opti- mally deciding the sensor type and mode that maximize the expected information profit. The sensor information profit is defined as the expected value of the information obtained through the sensor measurements, minus the cost associated with the use of the sensor and related resources, such as, its platform. The main philosophy behind this approach is to base the decision for sensor planning and platform naviga- tion on dynamic sensor measurements that become available over time and whose outcome depends on the decision variables. For many sensor surveillance systems involving multiple and heterogeneous components or agents, the value of sensor measurements can be expressed as an information- theoretic objective function. Then, the measurements can be viewed as a feedback to the sensor manager (or controller), and can be used to make optimal decisions about measure- ment sequence and sensor parameters. Ultimately, the solutions must optimize classification accuracy, probability of detection, and minimize the probability of false alarms. The use of information-theoretic objective functions for sensor management has been proposed by several authors. Schmaedeke used a discrimination gain technique to solve a multisensor-multitarget assignment problem [1]. Kastella managed agile sensors to optimize detection and classifica- tion based on discrimination gain [2]. Zhao investigated Manuscript received September 24, 2006. C. Cai is a graduate student of Mechanical Engineering at Duke Univer- sity, Durham, NC, 27708, USA cc88@duke.edu S. Ferrari is with Faculty of Mechanical Engineering at Duke University, Durham, NC, 27708, USA sferrari@duke.edu information objective functions such as entropy and Maha- lanobis distance measure for sensor collaboration applications [3]. However, little work has been done to compare these objective functions and analyze their per- formance across distinct sensor applications, such as, feature estimation and target classification. In this paper, a BN framework is developed for computing the discrimination gain and entropy reduction in multiple and heterogeneous sensor systems. Two common applications of multiple sensor systems are the classification of the target features from fused sensor measurements, referred to as feature inference, and target classification. The problems of feature inference for a Gaussian target are provided in [2]. In this paper, the theo- retic solutions for non-Gaussian distributions are derived for both feature inference and target classification. When multiple heterogeneous sensors are employed, their complementarity and performance relative to the environ- mental conditions are exploited through fusion. Dempster- Shafer (D-S) fusion technique has been shown to be very effective for performing feature inference and target classifi- cation based on multiple and heterogeneous sensor measurements [4-7]. A novel contribution of this paper is that the D-S fusion rule is embedded in the computations of the information objective functions to evaluate the expected benefit of obtaining sensor information that will be fused a posteriori. Also, the Bayesian network (BN) sensor model- ing presented in [8] is used in order to obtain a methodology that can be generalized to any measurement process, regard- less of the form of the underlying probability distributions. The paper is organized as follows. In Section II, the dis- crimination gain and entropy reduction are introduced. In Section III, the computation of these objective functions is presented for the feature-inference and target-classification cases, and D-S fusion is incorporated in the BN classifica- tion frame. The demining application is presented and demonstrated in Section IV. II. BACKGROUND A. Bayesian Network Modeling of Sensor Measurements A Bayesian network (BN) model is a directed acyclic graph (DAG) [8] comprised of a set of nodes representing variables, and a set of directed arcs connecting the nodes. In this paper, capital letters denote sets of variables, lowercase letters denote variables, and subscripts in lowercase letters denote the possible states of the variables. In BN models, Bayes’ rule of inference is utilized together with graphical manipulations to compute the posterior probability distribu- Comparison of Information-Theoretic Objective Functions for Decision Support in Sensor Systems Chenghui Cai and Silvia Ferrari T Proceedings of the 2007 American Control Conference Marriott Marquis Hotel at Times Square New York City, USA, July 11-13, 2007 ThC02.1 1-4244-0989-6/07/$25.00 ©2007 IEEE. 3559