Poster Abstract: Entropy-based Sensor Selection for Localization

Hanbiao Wang
Department of Computer Science

Kung Yao
Department of Electrical Engineering

[email protected] Greg Pottie

Department of Electrical Engineering

[email protected] Deborah Estrin

Department of Computer Science

[email protected]

[email protected]

University of California, Los Angeles Los Angeles, CA 90095

ABSTRACT
We propose a novel entropy-based sensor selection heuristic for localization. Given 1) a prior probability distribution of the target location, and 2) the locations and the sensing characteristics of a set of additional sensors, we would like to select an optimal additional sensor such that fusion of its measurements with existing information would yield the greatest entropy reduction of the target location distribution. The heuristic can select a sub-optimal additional sensor without retrieving the measurements of candidate sensors. The heuristic is computationally much simpler than the mutual information based sensor selection approaches for localization and tracking [1, 2]. Just as those existing approaches do, the heuristic greedily selects one sensor in each step.

Categories and Subject Descriptors

H.1.1 [MODELS AND PRINCIPLES]: Systems and Information TheoryValue of information

General Terms
Algorithms

proposed in [3]. Sensor selection based on expected information gain has been introduced for decentralized sensing systems in [4]. The mutual information between the predicted sensor measurements and the target location distribution has been proposed to evaluate the expected information gain of incorporating a sensor in [1, 2]. The dependency of the localization uncertainty on the sensor locations relative to the target location has been identied during the development of localization algorithms without using information theory [6]. One example is the convex hull heuristic of localization using time-dierence-of-arrival (TDOA) sensors. Targets inside the convex hull of TDOA sensors can be much more accurately located than those outside the convex hull. We propose a novel entropy-based heuristic for sensor selection based on our experiences with target localization. It is computationaly more ecient than mutual information based methods proposed in [1, 2]. We use the following notations throughout this poster: (1) x is the random variable for the target location, xt is the is the maximum likelihood estimate true target location, x of the target location, (2) zi is the random variable for the observation of sensor i, v is the view of sensor i about the target location, zi (3) xi denotes the deterministic location of sensor i.

Keywords
sensor selection, localization, wireless sensor networks

2. SENSOR SELECTION HEURISTICS

The sensor selection problem discussed in this poster can be formally dened as follows. Given (1) the prior target location distribution: p(x), (2) the locations of additional sensors: xi , i S , (3) the sensing models of additional sensors: p(zi |x), i S , the objective is to nd the additional sensor i whose measurements z i minimizes the conditional entropy of the posterior target location distribution, i = arg min H (x|zi ).
iS

1.

INTRODUCTION

There have been many investigations of information-theoretic approaches to sensor fusion and management. The idea of using information theory in sensor management was rst

(1)

Equivalently, sensor i maximizes the expected target entropy reduction,

Copyright is held by the author/owner. SenSys03, November 57, 2003, Los Angeles, California, USA. ACM 1-58113-707-9/03/0011.

i = arg max(H (x) H (x|zi )).

iS

(2)

H (x) H (x|zi ) is one expression of I (x; zi ), the mutual

304

information between the target location x and the sensor measurements zi . Thus, using the denition of I (x; zi ), i = arg max
iS

p(x, zi ) log

p (x , z i ) dxdzi p (x )p (z i )

(3)

[1, 2] propose to select the sensor with maximum I (x; zi ). I (x; zi ) is computed using the predicted p(zi ) based on p(x) and p(zi |x) without obtaining the actual measurement of zi . However, I (x; zi ) could be computationally intensive to low-end micro sensor nodes because p(x, zi ) could be in a space of up to four dimensions. We propose an ecient sub-optimal solution to the above problem. During the development of wireless sensor networks for localization, we have noticed that the localization accuracy is greatly eected by two factors, namely, the entropy of the sensors view of the target and the sensors sensing uncertainty. From the sensors perspective, the target is randomly distributed in the measurement space. For example, to a direction-of-arrival (DOA) sensor, there is only a bearing distribution for the target. The probability distribution v of sensors view of the target, p(zi ), is the projection of the prior target location distribution p(x) onto the sensors observation perspective, which solely depends on the target location distribution p(x), the sensor measurement modality (i.e. TDOA, DOA, range), and the sensor location xi . v The entropy of p(zi ) is
v = Hi v v v p (z i ) log p(zi )dzi .

H (zi ). However, the optimal sensor selection does not require an accurate evaluation of sensor information utility. Instead, a correct order of sensors in terms of information v utility is needed. Therefore, Hi can reasonably replace H (zi ) to sort sensors into the right order of their information utility. We have found a fast algorithm to compute v ) given p(x) and xi . If the sensing model p(zi |x) is very p (z i v complex, Hi will be much simpler to compute than H (zi ). H (zi |x) is actually the expected sensing uncertainty averaged for all possible x, H (zi |x) = p(x){ p(zi |x) log p(zi |x)dzi }dx. (9)

s Therefore, Hi dened in (5) and (6) are reasonable approxs is much simpler to compute than imations of H (zi |x). Hi s H (zi |x). Besides, Hi could be computed once and re-used multiple times if the sensing uncertainty does not change rapidly with time. On the other hand, H (zi |x) has to be re-computed each time even if the sensing uncertainty does not change with time. Therefore, the heuristic dened in (7) is much more ecient than the method using mutual information dened in (3).

3. RESULTS, CONCLUSION, AND FUTURE WORK

We have evaluated the above described sensor selection heuristic by using target localization simulations. The heuristic is much simpler to compute than mutual information. More details can be found in [5]. This poster presents the case study of sensor selection for localization using DOA sensors. We plan to implement the method on a real-time wireless sensor network testbed for localization.

(4)

The sensing uncertainty is the entropy of the sensors sensing model p(zi |xt ) at the true target location xt . For a singlemodal p(x), we can use the maximum likelihood estimate to approximate xt , and thus the sensing uncertainty is x approximated as
s Hi =

) log p(zi |x )dzi . p(zi |x

4. ACKNOWLEDGMENTS
(5) This material is based upon work partially supported by NSF under Cooperative Agreement #CCR-0121778, DARPA SensIT program under AFRL/IFG 315 330-1865 and ARODMURI PSU contract 50126.

(m) , m = 1, . . . , M ,the sensing For a p(x) with M modes x uncertainty is approximated as a weighted average of the sensing uncertainty for all modes,
s Hi = M m=1

p( x(m) )

(m) ) log p(zi |x (m) )dzi . p(zi |x

5. REFERENCES
(6) [1] J. Liu, J. Reich, and F. Zhao. Collaborative in-network processing for target tracking. EURASIP JASP: Special Issues on Sensor Networks, 2003(4):378391, March 2003. [2] E. Ertin, J. Fisher, and L. Potter. Maximum mutual information principle for dynamic sensor query problems. In IPSN, Palo Alto, April 2003. [3] K. Hintz and E. McVey. A measure of the information gain attributable to cueing. IEEE Transactions on Systems, Man, and Cybernetics, 21(2):434442, 1991. [4] J. Manyika and H. Durrant-Whyte. Data fusion and sensor management: a decentralized information-theoretic approach. Ellis Horwood, New York, 1994. [5] H. Wang, K. Yao, G. Pottie, and D. Estrin. Entropy-based sensor selection for localization. Technical Report 0011, NSF CENS, 2003. [6] K. Yao, R. Hudson, C. Reed, D. Chen, and F. Lorenzelli. Blind beamforming source localization on a sensor array system. In AWAIRS project presentation at UCLA, Los Angeles, CA, USA, December 1997.

The sensing uncertainty depends on the signal-to-noise ratio (SNR), the sensor hardware, and the sensing algorithms. We have repeatedly observed that the incorporation of the v s and a smaller Hi measurements of sensor i with a larger Hi will yield larger uncertainty reduction of the posterior target v location distribution. Therefore, we propose to use Hi s as a suboptimal but ecient alternative to I (x; zi ) for Hi evaluating the information utility of sensor i. The heuristic is to select sensor i such that
v s i = arg max(Hi Hi ). iS

(7)

v s Hi can be closely related to The entropy dierence Hi I (x; zi ) under several assumptions. Another expression of I (x; zi ) is H (zi ) H (zi |x). H (zi ) is the entropy of p(zi ),

p (z i ) =

p(zi |x)p(x)dx.

(8)

v ) when the sensing model p(zi |x) is dep(zi ) becomes p(zi v terministic without uncertainty. Thus Hi is smaller than

305