with the advent of small, inexpensive, low-power a market ... · be a promising path for...

A Market-based Approach toSensor Management

VISWANATH AVASARALA

TRACY MULLEN

DAVID HALL

Given the explosion in number and types of sensor nodes, thenext generation of sensor management systems must focus on iden-tifying and acquiring valuable information from this potential floodof sensor data. Thus an emerging problem is deciding what to pro-duce, where, for whom, and when. Identifying and making trade-offs involved in information production is a difficult problem thatmarket-based systems can “solve” by allowing user values, or util-ities, to drive the selection process. Essentially this transforms thetraditional “data driven” approach (in which multiple sensors andinformation sources are used, with a focus on how to process thecollected data) to a user-centered approach in which one or moreusers treat the information collection and distribution system as amarket and vie to acquire goods and services (e.g., information col-lection, processing resources and network bandwidth). We describeour market-based approach to sensor management, and compareour prototype system to an information-theoretic system in a multi-sensor, multi-user simulation with promising results. This researchis motivated in part, by rapid technology advances in network tech-nology and in sensing. These advances allow near universal instru-mentation and sensing with worldwide distribution. However whileadvances in service-oriented architectures and web-based tools havecreated “the plumbing” for data distribution and access, improve-ments in optimization of these distributed resources for effectivedecision making have lagged behind the collection and distributionadvances.

Manuscript received August 13, 2007; revised January 6, 2009; re-leased for publication June 3, 2009.

Refereeing of this contribution was handled by Chee Chong andRobert Lynch.

This material is based upon work supported by the Army ResearchLaboratory under contract number W911NF-04-2-0049 and Ten-nessee State University under subcontract number 332.77-07.361.

Authors’ addresses: V. Avasarala, G. E. Research, Niskayuna, NY12309, E-mail: ([email protected]]; T. Mullen and D. Hall,College of Information Sciences and Technology, The Pennsylva-nia State University, University Park, PA 16802, E-mail: (ftmullen,[email protected]).

1557-6418/09/$17.00 c° 2009 JAIF

1. INTRODUCTION

With the advent of small, inexpensive, low-powersensor nodes that can provide sensing, data processing,and wireless communication capabilities, sensor net-works can potentially generate huge amounts of diversedata. However, just because the data can be produced,does not mean that it should be. Sensor networks areconstrained by limits on sensor “attention” that includelimitations on battery power, bandwidth, and the num-ber and type of measurements that the sensor can handleat any one time. In addition, sensor networks can in-clude data collection entities that operate on very differ-ent timescales, from human reports to high-speed video-frame collection of images. System users (who may behuman, software agents or data fusion processes) eachhave their own individual tasks and priorities, but sharea common sensor resource pool. The sensor manager’sjob is to efficiently allocate sensors to end-user tasksso as to maximize end-user utility while simultaneouslyminimizing the cost of collecting, storing, and process-ing the data. Sensor managers must also consider theinterplay between various network resources, weighingtradeoffs between resource constraints such as batterypower, bandwidth, and sensor accuracy.For our current work, we assume that end users be-

long to a common overarching non-commercial institu-tion. Example application areas for such networks in-clude: (1) network-centric warfare, in which multiplesensing platforms, sensor nets, and individual soldierswith sensors interact to allow rapid tactical situation as-sessment and threat assessment [11, 12], and (2) mon-itoring of the environment via ground-based, airborneand space-based sensing systems.In recent years, information-theoretic approaches

have emerged as a promising paradigm for the develop-ment of a comprehensive sensor management for multi-task, multi-sensor networks. These techniques rely onoptimization of a certain information-theoretic measurelike cross-entropy [5, 20] or information gain [6, 27].Kastella [18] used cross-entropy to determine the opti-mal search order for detection and classification prob-lem. Kolba et al. [20] extended this framework to per-mit operation with uncertain sensor probabilities. McIn-trye et al. [25, 26] used information gain (the entropychange in environment for a given sensor allocationas the predicate for their hierarchical sensor manage-ment architecture. A valuable advantage of information-theoretic approaches is that they are highly flexibleand can be easily adapted to new problems. However,information-theoretic sensor management is concernedprimarily with scheduling the data-collecting entities(sensors) and other network resources such as energyusage and communication bandwidth have to be consid-ered separately. Additionally, information-theoretic sen-sor management approaches are myopic in nature, sincethey optimize some measure of the “quantity of infor-mation” obtained during a particular round of schedul-

52 JOURNAL OF ADVANCES IN INFORMATION FUSION VOL. 4, NO. 1 JUNE 2009

ing and neglect the “value of information” to the missionobjectives.Non-myopic sensor managers need to solve a highly

complex multi-period scheduling problem, since mostnetwork tasks like target tracking occur over multipleperiods of scheduling. Techniques based on approxi-mate dynamic programming have been developed forthis problem. In [5], Castanon considered a multi-grid,single sensor detection problem. Under certain assump-tions about the target distributions and probability dis-tribution of sensor measurements, Castanon proved thatthe optimal allocation policy would be to search ei-ther of the two most likely target locations during eachround of scheduling. In [6], Castanon considered theproblem of dynamic scheduling of multi-mode sensorresources for the classification of multiple unknown ob-jects. To solve this problem, the author proposed a hier-archical algorithm based on a combination of approxi-mate dynamic programming and non-differentiable op-timization techniques. Washburn et al. [43] formulatea single-sensor, multi-target scheduling problem as astochastic scheduling problem and use the Gittin’s in-dex rule to develop approximate solutions. Williamset al. [44] consider a single-target, multi-sensor alloca-tion problem with communication constraints and useadaptive Lagrangian relaxation to solve the constraineddynamic programming problem. Schedier et al. [37]have used approximate dynamic programming to allo-cate gimbaled radars for detecting and tracking targetsover a multi-horizon time period. The authors use athree-phase rollout algorithm with the following stagesa. generation of candidate sensor allocations b. gener-ation of alternate sensor plans based on results fromthe first component c. evaluation of the alternate sensorplans to calculate an approximation of the reward func-tion. The details of the implementation of these compo-nents are not explained in the original paper. However,for the three-sensor simulation that was presented intheir paper, simple heuristics to generate feasible solu-tions and to evaluate solution performance would havesufficed. These approaches for non-myopic sensor man-agement are pioneering, but substantial further researchis required to adapt them to the generic multi-sensor,multi-task sensor management problems.The network-centric environments that we are in-

terested in also must consider issues related to privacyand communication costs. For example, in network-centric warfare applications, multiple distributed enti-ties accomplish different tasks by connecting decisionmakers, effectors, and information sources to a commonnetwork [27]. Therefore, task information may be local-ized across individual users. Allowing all required taskinformation to be accessed by an optimization routineis communication-intensive and may violate privacy is-sues in a distributed environment. Pricing mechanismscan be designed to address privacy issues and minimizecommunication requirements [42]. Also, market-basedapproaches offer an inherently distributed mechanism

that can compare “apples” and “oranges” using the com-mon numeraire of money, thus reducing communicationoverhead to the single dimension of price. Under certainassumptions, price systems have been proven to providethe minimum dimensionality of messages necessary todetermine Pareto-optimal allocations [16].For the above reasons, we believe markets based

on combinatorial auction mechanisms are a promis-ing paradigm for a comprehensive sensor management.A combinatorial auction is an auction based on ex-changing item bundles (e.g., sensor readings+channeltransmission) rather than single items. In earlier workon distributed multi-agent sensor management, Lesseret al. [22] surmised that combinatorial auctions couldbe a promising path for market-based sensor alloca-tion. Shortly after our initial work on combinatorialauctions for sensor management [2], Ostwald et al. [32]also published preliminary work on using combinatorialauctions to find optimal sensor settings in a distributedradar array. The authors optimize a domain-specific andmyopic utility function using a combinatorial auctionmechanism during each round of scheduling. Resourceconstraints other than the sensor schedules are not con-sidered.A generic market-oriented approach to sensor man-

agement that is customizable for different sensor net-work scenarios must address several key issues. Thefirst issue is the mismatch between what users wantto buy (e.g., tracking and identifying a target with aspecified accuracy) and what network resources are of-fering (e.g., cpu, battery power, bandwidth, and sensordirectivity and operation mode). The problem becomeseven more complicated when we consider that differentcombinations of sensors can be used to track a target,but each combination of sensors may give a differentquality of service (QoS). To accurately assess and bidfor different sensor combinations, users would need toknow the operating parameters of each sensor, and tocalculate the QoS for various combinations of sensors.This leads to the second issue of preference elicitation,or eliciting user valuations for all possible combina-tions of resources to different tasks/users. Clearly in thissetting, preference elicitation can be computationallyand/or communication intensive. For example, if thereare n sensors and m tasks, ((2n¡ 1)m+1) utility valua-tions must be acquired by the sensor manager from theuser to calculate an optimal allocation. The third issueis that winner determination, or determining an optimalallocation given all bids, is an NP-hard problem [35].Although fast algorithms have been developed, thanksin part to ecommerce-driven advances, these algorithmsmay not always meet real-time requirements.To address the above issues, we proposed a frame-

work for sensor management using a market-based ar-chitecture called MASM (Market-Architecture for Sen-sor Management) [15, 30, 33]. The sensor manager han-dles the mismatch between what providers (i.e., sen-sors) offer and consumers (i.e., end users) want by

A MARKET-BASED APPROACH TO SENSOR MANAGEMENT 53

Fig. 1. Single-platform market architecture for sensor management.

providing a mapping between user tasks and networkresources. Users bid on high-level tasks, while ser-vice mapping components convert the high-level usertasks to low-level sensor tasks and finally to actualbids. Once the necessary bids are created, an auc-tion winner determination algorithm computes the fi-nal resource allocation. Both the bid formulation andwinner determination steps are computationally ex-pensive. Traditionally, humans have been mainly re-sponsible for the bid formulation step, with compu-tational auctions focusing on the winner determina-tion step. One of our contributions has been to de-velop an approximate algorithm, called Seeded Ge-netic Algorithm (SGA) [29], that combines these twosteps and achieves polynomial run times with a mod-est loss of optimality. Our earlier work [2, 3, 28, 29]described a high-level framework for MASM, but didnot provide any implementation details. This paper de-scribes the implementation of MASM, including theauction protocol, pricing algorithms for network re-sources and heuristics for avoiding myopic schedulingbehavior.We currently focus on a single-platform design, al-

though we plan to extend this model to multiple plat-forms and sensor network environments. While wedraw from recent advances in ecommerce-based mar-ket research, we describe the significant challenges inadapting this approach to reflect typical sensor man-agement environments. We test our prototype sensormanagement system using a multi-sensor, multi-usersimulation framework that models bandwidth and bat-tery power constraints. Comparisons to a priority-basedinformation-theoretic system show that market-based al-gorithms hold promise for developing comprehensivesensor management systems.

It should be noted that the present approach has lim-ited applicability to smart dust environments, where thenumber of sensors could be on the order of few hundredthousands. In these environments, the communicationcosts of relaying sensor measurements to the sink arethe dominant costs of network operation. For these en-vironments, a centralized auctioneer cannot be used be-cause of the communication costs involved. Instead, taskutility information and price information should perco-late to the node level, where individual nodes decideon what actions to perform. Mainland et al. [24] haveproposed a price-based decision system for smart dustenvironments, and Padhy et al. [33] have proposed autility-based model.Our paper is organized as follows. In Section 2, we

describe the MASM architecture and provide an illus-trative scenario in Section 3. Section 4 talks about ourcontinuous combinatorial auction (CCA) protocol de-veloped to minimize communication involved in marketoperations. Section 5 describes the pricing mechanismsthat have been developed to enforce resource constraintsin the market. Section 6 introduces an agent learningscheme for market agents to assist users in formulatingoptimal bidding parameters for different tasks. Section 7describes our simulation environment, while Section 8describes our results. We summarize our findings anddiscuss future work in the last section.

2. MASM

Our current single platform design for MASM isshown in Fig. 1, and derives from the sensor manage-ment architecture proposed by Denton et al. [8]. Themission manager (MM) assesses mission-level decisions(e.g., assigning task priority to a mission goal), allo-cates tasks and budgets to end-users. Within the mis-


sion manager, approaches such as goal lattices (whichrelate high-level mission goals to lower-level actionabletasks) can be used to measure the criticality of vari-ous low-level goals to the overall mission goals, andthus help to determine their respective budgets. KennethHintz and Gregory McIntyre [14] used goal lattices tocompute the relative weights of actionable tasks (suchas tracking) on the basis of high-level mission goals. Ra-jani Muraleedharan and her colleagues [30] used goallattices to determine weights for combining various ob-jectives to optimize routing in a sensor network. Newerdevelopments include dynamic goal lattices [15] thatcan support more dynamic goal generation from a setof predefined goals.The sensor manager (SM) acts as a competitive mar-

ket for buyers and sellers of sensor resources. Sensorsand transmission channels are modeled as sellers. Sen-sors sell their sensor schedule (i.e., their “attention”) andtransmission channels sell raw bandwidth. End users,or consumers, of the sensor network are interested inhigher-end products such as target tracks, environmen-tal searches, and target identification. MASM maps be-tween these high-level tasks and actual resources avail-able in the market using its combined service chart/bidformulator functionality.MASM provides this functionality in two different

modes, either exact service mappings (E-MASM) orapproximate service mappings (A-MASM). When thenumber of sensors is small and the real-time constraintsare relaxed, E-MASM mode provides an exact servicemapping. In other words, given a task and a set of pos-sible resource combinations that can be used for thattask, E-MASM will explicitly calculate the utility ofassigning each combination to the task using domaininformation and task-specific utility functions providedby a service chart. Given n sensors in the network, andm tasks, then in the worst case, (2n¡ 1)m bids on re-source combinations might have to be formulated. Astandard combinatorial auction winner determination al-gorithm [1] then determines the optimal allocation. Oneapproach used to speed up the bid formulation auctionsand the winner determination optimization is to gener-ically restrict the type of bids considered for resourceallocation. For example, one could place a bound onthe maximum number of items in a bid. Polynomialalgorithms for bids with certain special structures [35]are available. However, imposing generic constraints onbid types can lead to market inefficiency. An alternateapproach is to use domain-specific knowledge to in-telligently restrict the number of resource bids formu-lated. For example, if the types or locations of sensorresources that can be used to accomplish a particulartask are limited, the combinations of resources that needto be considered can be reduced.When the number of sensors is large and real-time

constraints are strict, explicit mappings are no longerfeasible, and the A-MASMmode, with approximate ser-vice mappings, is used. Instead of the bid formulator

explicitly formulating combinatorial bids for each usertask, MASM searches the search space of useful sen-sor combinations directly using a polynomial, anytimeevolutionary algorithm [29].

3. SCENARIO EXAMPLE

To illustrate the MASM system, we describe a sim-ple scenario with two users and two sensors. Assumethat a particular user is interested in searching and iden-tifying reconnaissance drones approaching from a cer-tain region R1. Let the task that the user wants to ac-complish using the sensor network resources be the re-duction of entropy of the probability distribution of theexistence of reconnaissance drones in region R1 to lessthan a threshold "1 (task A).The user should submit a bid to MASM in the

following format:

(type: search/identifyentity: reconnaissance drone xregion: R1quality: (entropy< "1)price: PA)

where PA is the user’s bid price.

Assume that another user is interested in estimatingaccurately the position of an already identified slowermoving reconnaissance drone. Let the task that this useris trying to accomplish using the sensor network re-sources be the reduction of the track uncertainty (asmeasured by some reasonable metric such as state vec-tor covariance error) to less than "2. The user shouldsubmit a bid to MASM in the following format:

(type: trackentity: reconnaissance drone yquality: (covariance error< "2)price: PB)

where PB is the user’s bid price.

Assume that two sensors, a forward looking infrared(FLIR), or infrared camera, and a radar are available tothe SM for accomplishing these tasks. Note that anytwo sensors will generally have different abilities to lo-cate and identify targets depending on characteristicssuch as environment conditions, target characteristics,and target-sensor geometry. In other words, certain sen-sors, or combinations of sensors, can provide more orless value for task completion, and thus the bid valuesfor different sets of sensors may also vary. To expressthis, MASM generates combinatorial bids in exclusive-or format for each user’s task as shown in Table I duringeach of scheduling. The exclusive-or format ensures thattask A can win either bid 1 or bid 2, but not both. Foreach bid in Table I, the bid amount is represented usingthe format PU,S, t where U is the task identification, Sis the given sensor combination number and t indicatesthe scheduling round. This notation is used to indicatethat bid prices depend on sensor combination and user


TABLE IBids Generated by MASM for Sample Scenario During the First Round of Scheduling

Task A’s XOR bids Task B’s XOR bids

Bid 1 Bid 2 Bid 3 Bid 4

(type: search/id (type: search/id (type: track (type: track

entity: reconnaissance drone x entity: reconnaissance drone x entity: reconnaissance drone 3 entity: reconnaissance drone 3

region: R1 region: R1

sensors requested: FLIR sensors requested: FLIR and Radar sensors requested: FLIR sensors requested: FLIR and Radar

quality: (entropy< "1) quality: (entropy< "1) quality: (covariance error< "2) quality: (covariance error< "2)

price: PA,S1,1) price: PA,S2,1) price: PB,S1,1) price: PB,S2,1)

task. The value of PU,S, t is calculated by the SM usingthe bid prices of the original consumer bids (see Sec-tion 4 for details). Until the tasks are complete, the SMmonitors the progress of the tasks and adjusts the bidsaccordingly.We describe the methodology used by MASM to

generate the bids for resources during each round ofscheduling in the next section. The combinatorial auc-tion winner determination algorithm is then used to cal-culate the optimal resource allocation, given the MASMbids.

4. CCA PROTOCOL

In this section, we describe our continuous combi-natorial auction (CCA) protocol. The CCA protocol wasdesigned to increase the computational and communi-cation efficiency of our market-based scheduling algo-rithm. Since MASM uses discrete time slots to sched-ule resources, most user tasks, like tracking a target,require acquiring resources over multiple time slots.Each round of scheduling can either occur periodicallyat fixed times, or randomly. A simplistic allocation ofresources across multiple time slots can occur in twoways: i) Users send in a bid that covers resource needsacross multiple time slots. The SM updates the scheduleupon receiving each new user bid. ii) Users send in a bidfor the current time slot only. After each round, usersupdate their requirements based on what was receivedin the last scheduling round, and send in an updated bidfor the next time slot.The first approach is computationally expensive. De-

termining the optimal scheduling for n sensors over atime horizon T is exponentially complex in n and T.Clearly, the second approach is communication inten-sive. We designed the CCA protocol to avoid the com-munication and computation requirements of using mar-kets for sensor management. Below, we describe theCCA protocol in detail.CCA executes each of the following steps (except

initialization, which is executed once at the start ofoperations) during each round of scheduling.

4.1. Initialization

Auctioneer initializes the prices for all the resources.It informs the users about the set of tasks that it willaccept bids for.

4.2. Update Bids

At the beginning of each round, users can i) sendnew bids, ii) remove their current bids from the auction,iii) modify the parameters of their existing bids. Userbids are of type ht,pi where t is the task description,which includes the task type, and final task qualitydesired by the user and p is the price that the user iswilling to pay. For example, the task description fora bid to track a target x such that the trace of thecovariance matrix of the target estimate is less than0.001 is as follows:

(type: trackentity: target xquality: (trace of covariance matrix< 0:001))

The auctioneer predefines the set of tasks that theuser can bid for and the bid format. Here we make thestandard assumption that a scalar valued “quality” mea-sure can be calculated using the various task parameters.For example, for target tracking, the trace or the deter-minant of the covariance matrix can be used as one mea-sure of target track quality [13, 31, 36]. However, undercertain circumstances, it is advantageous to use moreelaborate task descriptions (see [17] for related discus-sion). For example, the requirement to identify a targetwith a given level of specificity and level of confidencemay require extensive models of multi-sensor perfor-mance in complex observation environments. These areapplication-specific, and would need to be developedfor the particular application being considered.

4.3. Update User Requests

Auctioneer accepts new bids or updates to existingbids during each round of scheduling. If the auctioneerreceives no message regarding a particular bid, the bid


stays active and competes for resources in the currentauction round.

4.4. Resource Bid Formulation

Since user bids are for high-level tasks, the auc-tioneer needs to compose bids for actual resourcesfrom them. This responsibility is handled by the bid-formulator module in E-MASM (explicit formulationof bids for resources is not required in A-MASM). Letthe user bid on a high level task T at time ti with pricePT. A high level task, such as tracking a target to a re-quired accuracy, might require resources over multiplerounds of scheduling. For each time slot t, the auction-eer constructs bids on each resource set, S, that can beallotted to task T. The auctioneer needs to calculate theprice associated with resource set S for task T duringeach round of scheduling, based on the user bid pricePT. To correctly align task priorities with bidding price,we have devised a novel mechanism for resource pricedetermination. For a resource set S, the auctioneer com-putes the bid price for a resource set as the percentageof the user task completed by the resource set given thecurrent task status. To determine the percentage of taskcompleted by a resource set S, we cast the problem interms of optimally scheduling sufficient readings from acanonical sensor A to meet the quality of service (QoS)task parameters. Let the task T require on average naconsecutive schedules of the standard sensor A to becompleted (task is considered complete, when the taskquality meets the QoS threshold in the task bid). Sup-pose a resource bundle S is used when the task quality isq, and the expected number of standard sensor readingsrequired is reduced to na Then fS,T,q or the percentage ofthe task completed by resource set S when the currenttask quality is q, is equal to the percentage savings inthe required number of canonical sensor readings.

fS,T,q = (na¡ na)=na:There is a possibility that fS,T,q is negative (na < na).For example, in spite of allocating sensing resources,the inaccuracy associated with a target estimate mightincrease with time. To avoid negative prices for bundlesof resources, the bid price for PS,T,q (bid price forallocating resource set S to task T during a particularround of scheduling when the current task quality is q)is calculated as

PS,T,q = PT¤fS,T,q¡PT¤fÁ,T,qwhere Á is the null set.The auctioneer uses this price to prioritize between

different tasks during a particular schedule. However,this price is not charged to the user. Users are chargedonly at the end of a successful task completion, or ifthey choose to withdraw a bid before the task could becompleted by the SM (see the round termination step).Calculation of na and na can be made faster, by storing

Fig. 2. Illustration of calculation of bid prices for resource bundlesusing QoS chart.

task specific performance data for the canonical sensoras QoS charts. For example, consider a user bid forsearching a particular grid for potential threats, whenQoS is measured in terms of entropy. A sample QoSchart is given in Fig. 2, and shows the expected fall inentropy with standard sensor readings. Let a user bidspecify a task for reduction of entropy of a particulargrid from ei to ef . If a resource bundle S is expected toreduce the entropy from ei to ej after the next reading,then

fS,T,ei = (nj ¡ ni)=(nf ¡ ni):Creating a QoS chart for a sample task is illustrated

in Section 7.Resource formulation is the slowest link in CCA

protocol, but heuristics can speed this step up. For ex-ample, for certain tasks, it may be feasible to use onlya few kinds of resources. Thus, if a task requires onlyacoustic data, then only the acoustic radar sensors needto be considered. However, in the worst-case scenario,the number of bids is exponential in the number ofsensors. This is clearly infeasible in case of large sys-tems and thus E-MASM is not scalable to large sys-tems, limiting its effectiveness. The A-MASM formu-lation avoids explicit bid formulation, and hence main-tains polynomial run-times, both in number of resourcesand users by using our SGA algorithm, an approximatepolynomial-time algorithm. For a detailed descriptionof this algorithm, and a comparison of A-MASM andE-MASM time performance, see [34].

4.5. Resource Allocation

Resource bids, obtained from step 3 are exclusive-OR bids in the form hS1,P1ixorhS2,P2i : : :xorhSn,Pni.This bid indicates that during the current round of


scheduling, the user is willing to pay a price P1 for theresource bundle S1 and a price P2 for S2, but only themaximum of P1 and P2 for S1 [ S2. The auctioneer needsto translate these bids to OR bids, so that standard inte-ger programming formulations [1] can be used. An ORbid of the form hS1,P1iorhS2,P2i : : :orhSn,Pni indicatesthat the user is willing to pay P1 +P2 for the bundle S1 [S2. This can be done by the addition of phantom items[9]. The idea is to translate an exclusive-OR bid B-xorof the form hS1,P1ixorhS2,P2i : : :xorhSn,Pni into a B-ORbid of the form hS1b [ b,P1iorhS2 [b,P2i : : :orhSn [b,Pni,where b is a phantom item. The phantom item b ensuresthat a maximum of only one bid from the OR bids canbe labeled as winner (since each item can be allocatedto maximum of one bids). Once the bids are translatedinto the “OR” format, the winner determination problembecomes a standard integer programming (IP) problem,that can be characterized as

maxnXj=1

pjxj s:t:Xjji2Sj

xj · 1 8 i 2 f1 : : :mg

where xj is 1 if the bid is accepted in the final allo-cation and 0 otherwise. The IP problem can be solvedusing a commercial software package like CPLEX. Thewinner determination problem is NP-hard [18] and intheory, this resource allocation step could prove compu-tationally expensive for E-MASM. Performance of theIP formulation depends greatly on the characteristics ofthe probability distribution from which bids are gener-ated. For example, the time taken by a problem withone thousand bids and hundred items on a 2.8 GHzPentium IV processor varied between 0.001 seconds to5000 seconds, depending on the bid distribution. How-ever, we found that the problems generated by the sen-sor network simulation are relatively easy for CPLEX(see Section 4).

4.6. Round Termination

The auctioneer updates the costs of resources ex-pended on a particular bid by adding the price of itsallocated bundle. For each user bid b, the cost of re-sources allocated to the bid is updated as

Cb = Cb+#tsi

where Si is the bundle allocated to b during the tth roundof scheduling and #tsi is the price of the bundle Si duringthe tth round of scheduling. This is calculated as the sumof the prices of the individual resources comprising Si(see Section 5 for explanation of resource pricing).Also, the auctioneer verifies if the task quality re-

quired by each user bid was achieved. When a task iscomplete, the bids for that task are removed from theauction and the corresponding user is sent the completedtask details. The bidding user is charged the minimumof his bid price Pb or the cost of resources spent on thetask by SM, Cb.

paymentb =min(Cb,Pb)

where paymentb is the fee charged to the user, Pb isthe bid price, and Cb is the total cost of the resourcesallocated to the bid. This fee structure ensures thatno user is charged more than their bid price for anytask. When Cb < Pb, the user has a positive surplusof Pb ¡Cb. An alternate fee structure that divides thesurplus between the SM and user is as follows:

paymentb =min(Cb,Pb)+H(Pb¡Cb):¤° ¤ (Pb¡Cb)where H(x) is the Heaviside step function and ° is thepercentage of surplus given to SM. If the user withdrawsa bid before the task demanded in the bid is completed,the SM charges the user

paymentb =min(Cb,f,b¤Pb)+H(f,b,¤Pb ¡Cb):¤° ¤ (f,b ¤Pb¡Cb)

where f,b, is the percentage of the task that is alreadycompleted by the SM. This is calculated using QoScharts (as described in the resource bid formulationstep). This fee structure has been designed to mitigatethe impact of dishonest user behavior (see Section 8 fordetails).Finally, the auctioneer updates the prices of the

resources based on the demand in the current round.We describe how prices are updated in the next section.The auctioneer then updates the resources about theirschedules during the current round and sleeps until thenext round of scheduling begins.

5. PRICING MECHANISMS

To set prices for individual resources, we use a pric-ing protocol similar to the tatonnement process. Taton-nement is an iterative procedure for finding equilibriumprices based on the search parameter (e.g., price orquantity) [21, 34, 40]. The price adjustment processstarts with an auctioneer communicating an arbitraryprice set to the users. The users compute their demandfor the first good at the given prices and communicate itto the auctioneer. Depending on whether the aggregatedemand for the first good is positive or negative, theauctioneer either increases or decreases its price. Thisprocess continues until a price at which aggregate de-mand for the first good equals zero is reached. Thisprocess is then repeated for the second good and soon. At the end of the first cycle, only the last good isguaranteed to have a zero demand, but assuming grosssubstitutability (i.e., when the price of good j goes up,there is a positive increase in the demand for every othergood by each user) the price set arrived at after eachcycle is closer to equilibrium than the previous one.More refined algorithms using partial derivatives of thedemand functions have been developed to search forequilibrium in parallel [38, 45]. Though the gross sub-stitutability assumption is often violated (as in sensornetworks), the tatonnement process has been found togive satisfactory results [7].


To use the tatonnement process in MASM, we modelthe supply and demand functions for a resource at aparticular price. MASM estimates these functions usingthe current resource usage rate. Prices for individual re-sources are initialized to zero during the sensor networkinitialization. After each round of scheduling, the prices(#t+1S ) for the resource S for the next round of schedul-ing are calculated based on the current usage rate of theresource (rtS) and the available usage rate of a

tS .

#t+1S =max(0,#tS + ¿ ¤ (rtS ¡ atS))#0S = 0

where ¿ is the constant which determines the rate atwhich prices are updated.The definition of rtS and a

tS is dependant on the

resource being modeled. For example, for sensors, wehave used the available battery power. Let sensor A beendowed with initial battery power bi and assume thatSensor A needs to be available for a total operating timeof T. At time t, if the available battery power is bt, then

rtA = (bi¡ bt)=t if t > 0, 0 otherwise;

atA = (bt)=(T¡ t) if t < T, 0 otherwise:

Ideally, the tatonnement process would update the priceof one resource, run the winner determination algorithmto find the new demand for resources, then conductprice updates for the second resource, and so on. How-ever, because of time and communication constraints,all the price updates are conducted simultaneously us-ing the current rate of utilization, during every round.We expect that the results between the two approacheswill not be very different, since the usage rates are mov-ing averages and do not vary significantly based on theusage during the current time slot.

6. AGENT LEARNING

In MASM, the SM accepts bids only on a set ofpre-defined tasks. The user agent is responsible for de-composing the high level tasks or goals that it has autility for into a sequence of SM acceptable subtaskson which it can bid. Also, the user agent has to assignappropriate priorities or bid prices to these sub-tasks, sothat its overall performance is optimized. Appropriateassignment of priorities to these sub-tasks has a signif-icant impact on agent performance in the market. Asan initial exercise, we experimented with agent learningthat uses a simple, greedy Widrow-Hoff based learningto optimize bid parameters based on current market data.For reasons of brevity, the details of this approach arenot provided here, but readers are directed to [42]. Amore rigorous learning method will be the subject offuture research.

7. SIMULATION ENVIRONMENT

A simulation environment consisting of a two-dimensional search area involving multiple targets,

multi-user and multiple sensors was developed for test-ing MASM and comparing its performance with othersensor management approaches. The design of the sen-sor network, including the communication channel, isinspired by the DARPA sensor network implemented tocarry out research in sensor management domain [22].These sensor networks are more platform-based and dif-fer from the emerging field of “smart dust,” where thesensor network could consist of millions of sensors.Our simulation environment is representative of the

types of sensors, communications resources, and mis-sion objectives for a tactical military environment. Thevarious sensor parameters we used are based on real-istic sensor models and are obtained from [26]. Whilethis is a basic scenario, with a limited number of sen-sors and targets, it is representative of the types of non-commensurate sensors that would be available for otherapplications such as environmental surveillance and cri-sis management systems (e.g., for homeland security).We make a few simplifying assumptions about sensormodels since our main purpose is to test SM perfor-mance rather than absolute fidelity to field conditions.Below we describe our simulation model.

7.1. Users

Users consist of a set of software market agents thatsearch for and destroy targets. These agents have theability to attack any position within a range of r metersand any target that falls within ° meters of an attackedposition is destroyed. The agents are not provided withany sensing resources and they depend on the sensornetwork for obtaining information about the environ-ment. They bid for sensor resources during each roundof scheduling and update their status based on informa-tion provided by the sensor manager. Initially, agentsmove along the simulation area with constant velocityvc, searching for targets. They use the sensor network’sresource to search for potential targets and if the prob-ability of target existence within their range exceeds athreshold pthreshold, initialize target tracks. Once a tar-get track is initialized, agents can attack a target if the99% confidence interval of the target’s position is lessthan ° meters. Hence, they are required to track the tar-get to the required accuracy before attacking it. This isagain accomplished by buying sensing resources fromthe sensor network. Agents are assumed to have a utilityut for destroying a target. To divide the overall utilityinto utilities for search and track tasks, agents initiallyuse equal priorities. During the simulation run, agentsupdate the search to track budget ratio using the learningmethod, mentioned in Section 6.

7.2. Targets

Targets are randomly distributed throughout thesearch area. They move randomly along the city roadswith constant velocity vt, corrupted by a Gaussian white


TABLE IISensor Characteristics used in Simulation

Sensor No. Type Range Bearing Axis PD PFA

1 Doppler 90 m§ 10% 1± = 6¾ x 0.95 0.0012 Radar 30 m§ 10% 0:1± = 6¾ x 0.95 0.0013 FLIR NA 0:1± = 6¾ y 0.99 0.0014 ESM NA 1± = 6¾ x 0.5 0.015 IR NA 100 ¹rad = 6¾ x 0.99 0.016 Radar 30 m§ 10% 0:1± = 6¾ y 0.95 0.0017 Doppler 90 m§ 10% 1± = 6¾ y 0.95 0.0018 Radar 30 m§ 10% 0:1± = 6¾ y 0.95 0.001

noise with variance Q. Two different types of targetsare modeled (T1 and T2). Users have greater utility fordestroying T2 targets. Only T1 targets were used in thesimulation experiments, unless otherwise specified.

7.3. Sensors

The simulation models several different kinds ofsensors, including sensors that provide range and bear-ing, bearings-only sensors, and Electronic Support Mea-sure (ESM) sensors. Measurements of two bearings onlysensors, which are not located at the same position,can be combined to create both range and bearing es-timates and can be used as a pseudo-sensor. A for-mal way of modeling sensors is to model their proper-ties, such as bandwidth, wavelength, duration of wave-form, signal power per pulse, receiver noise strengthdiameter of radar aperture. A much simpler model-ing technique, in which a sensor’s characteristics arecharacterized by three parameters, its probability of de-tection PD, probability of false alarm PFA and bearing[6], is used in this simulation. The simulation envi-ronment has eight different sensors of five differenttypes that are located on two different platforms or-thogonal to each other. The operating characters of thevarious sensors are given in Table II. Both the plat-forms are 100 km away from the search area. Sincethe distance of the sensors from the simulation areais large, small angle approximation s= r¤dμ, wheres is the length of area that falls under the sensor’sbeamwidth, dμ is a beamwidth of the sensor in radi-ans and r is the distance of the sensor platform to thecity (100 km). For a detailed description of the sensormodeling techniques adopted in the simulation, refer to[25, 26]. Each sensor has a battery with einitial unitsof energy. For the purpose of brevity, all the sensortasks are assumed to cost zero energy, except the taskof transmitting messages. The energy spent in trans-mitting a message of m bytes over a distance of dmeters is calculated as _®d2m where _® is a constant(see Table III).

7.4. Communication Channel

For communication purposes, a RF communicationchannel with capacity C is used. All the messages are

assumed to be of uniform size M bytes. The commu-nication protocol used is a contention-based protocol(like CSMA/CD) where each agent with a message tocommunicate senses to see if the channel is busy andtransmits if it is not. If two entities start transmittingat the same time, they back off and wait for a randomamount of time. The time taken for communication, viathis channel, for a fixed number of messages is stochas-tic. SM enforces the bandwidth constraint by restrictingthe probability that the time taken for communication isgreater than tcom to less than ¯%.

7.5. Sensor Manager (SM)

Since the number of sensors is not large, E-MASMformulation is used. Bids on two types of tasks, searchand track, are accepted by SM. The QoS for search tasksis in terms of entropy and for the tracking tasks, thenorm of the estimate covariance is used. To create theQoS mapping shown in Fig. 2 for the detection task, thefollowing procedure is used. Let the initial probabilityof target presence in a particular cell be ¼0 = 0:5. (with¼0 = 0:5, the cell has the highest possible entropy). Theinitial entropy of the grid Ho is calculated as

Ho = g(¼0) where g(¼) is defined as

¡¼ ¤ log(¼)¡ (1¡¼) ¤ log(1¡¼):Assume that the canonical sensor A with probability ofdetection ¸D and ¸FA is used for verifying the presenceof the target in this cell. Assume that a target is presentin the cell. Then, the estimated probability ¼tn of targetpresence in the cell after n consecutive readings of A,can be calculated using Bayesian analysis. Similarly, letthe estimated probability after n consecutive readingsby A, if the target is not present in the cell be ¼ntn . Theexpected entropy of the cell after n consecutive readingsof A is

Hn = ¼0¤g(¼tn) + (1¡¼0)¤g(¼ntn ):The plot of Hn vs. n is used as the QoS mapping for thedetection task.

7.6. Information-Theoretic Sensor Manager (ITSM)

To compare the performance of MASM, we neededan alternate sensor manager that can handle multiple


TABLE IIIParameter Values used in Simulation

Parameter Value Description

Total no of time slots 500 Total number of resource allocation schedulesnc 5 No of consumersnt 10 No of targetsnt2

2 No of targets with offensive capabilities

vc 50 mps Velocity of consumerspthreshold 0.99 Detection thresholdvt 50 mps Velocity of targetsQ 0.01 Variance of Gaussian white noise of target motionr 50 m Maximum distance that consumers can attack° 1.5 m Radius of destruction around attacked position½ 99% Required confidence interval length of target's position estimateut 1.0 m Utility for destroying a target¿ 0.005 Tatonement factorC 2 Mbps Bandwidth of communication channelM 1 Kb Size of communication messagetcom 2 millisec Maximum time allowed for communication¯ 0.01 Required probability that time taken for communication is greater than tcom® 1 pJ/bit/m2 Energy required to send messages per unit distance per unit message sizeei 2.5 KJ Initial energy of sensor batteries

heterogeneous tasks and multiple heterogeneous sen-sors. As explained in Section 1, the currently availableapproximate dynamic programming based approacheswere not directly applicable to this problem withoutsubstantial additional work. For this purpose, we im-plemented an information theoretic sensor manager(ITSM). Hint and McIntrye [25] used information gain(the entropy change in the environment for a given sen-sor allocation) as the predicate for their hierarchical sen-sor management architecture. The amount of informa-tion gained can be measured by the change in entropyprior to and preceding a sensor measurement. ITSMcalculates the information gain, associated with eachpossible allocation and schedules the resources as perthe allocation with the highest information gain. To en-sure that ITSM considers the “value of information,”we optimized a weighted measure of information gain,instead of relying on the raw information gain. Weused the formulation in Kalandros et al. [17] for pri-ority based information-theoretic based sensor manage-ment. Instead of multiplying the information gain by thecorresponding task weight, the authors use the formulaI0s,t = Is,t+ log(μ

tt) where I

0s,t is the weighted information

gain, Is,t is the information gain obtained from allocatingsensor suite s to task t and μtt is the priority of task t. Akey issue in the use of ITSM is the priorities that needto be assigned to the various tasks. We exhaustivelytested the performance of ITSM by varying the trackand search budget ratios and found that the optimal userperformance was obtained when a track to search bud-get ratio of 0:9 : 1 was used. To enforce the bandwidthconstraint, ITSM does not consider allocations that re-quire bandwidth, which has more than 0.01% chanceof crossing the tcom limit. The expected time taken fora particular bandwidth consumption was determined by

using monte-carlo simulations. ITSM does not modelenergy constraints and these are handled in the experi-mental setup as explained in the results section.

8. RESULTS

8.1. ITSM vs. MASM Experiments

Information-theoretic sensor managers schedule sen-sors to minimize the entropy of the environment. How-ever, incorporating battery power constraints into ITSMis not straightforward since most systems either use ad-hoc metrics or else do not address power constraintsexplicitly. Instead of initially testing against multiple ad-hoc solutions, we compare the ITSM system using twosets of experiments: 1) all the energy requirements ofthe sensor network are assumed to be zero, and 2) ITSMand MASM consume the same amount of energy fordifferent tasks (as shown in Table III), but ITSM doesnot use any explicit policy for allocating battery poweracross the mission. Once a sensor has exhausted its bat-tery, it is not considered in future allocations. In thesimulation, the user’s primary goal is to destroy as manytargets as possible. Therefore, we evaluate sensor man-agement performance by calculating the average num-ber of targets destroyed by ITSM and MASM, as shownin Fig. 3. The left bar graph shows experiments whereenergy constraints are zero, while the right bar graphshows experimental results when energy constraints areenforced. In both cases, MASM was more successful inmeeting user objectives (i.e., in destroying the targets)than ITSM. However, in the second set of experiments,some of MASM’s success can be attributed to a betterenergy enforcement policy and it is not clear from theseexperiments whether MASM will outperform ITSM for


Fig. 3. Comparison of MASM with ITSM (averaged over 100 runs). Left bar graph shows experiments where energy constraints areneglected. Right bar graph shows experiment results when energy constraints are enforced.

any given energy usage policy. We note however, giventhat MASM does achieve a higher number of targetskilled even when the sensor network battery power is“free” for both systems, that this would appear to indi-cate that MASM’s superior performance is not entirelydue to a better energy enforcement policy.Two reasons that MASM outperforms ITSM in

meeting user objectives may be that MASM 1) oper-ates to maximize user utility rather than to maximizethe information content, and 2) uses prices to prioritizetasks. We discuss these two reasons below.1) MASM acts to allocate resources to maximize user

utility (as indicated by their bid prices). Since a user’sutility depends on how well the allocated resource setcontributes to the user’s goals, market-based resourceallocation automatically takes goal-related parametersdirectly into consideration. On the other hand, ITSMconcentrates on maximizing information content, ne-glecting the value of the information to the goals. Thepriority-based ITSM does a better job than the stan-dard ITSM at incorporating user goals (as priorities) butthe system itself has no means of considering a task’sprogress toward the goal. As an example of why track-ing progress toward a goal can be useful, consider thefollowing simple scenario. A single sensor is used totrack two targets T1 and T2 with equal priority simulta-neously. For the first reading, ITSM gets the most in-formation content from tracking T1, then for the secondreading, ITSM gets the most information from track-ing T2. When the confidence interval necessary to attackthese two targets is tight, ITSM will never get enoughsequential readings to lower the uncertainty sufficientlyand will oscillate between the two targets. On the otherhand, MASM has equal likelihood for choosing eitherof the two targets, in any round of scheduling. This hap-pens because the fraction of task completed per readingfor either of the target tracks remains constant. There-fore, MASM finishes the tasks in a finite time.

To ensure that our intuition about ITSM vs. MASMwas accurate, we implemented the following simpleexperiment, based on the above scenario, where a singlesensor tracks the two targets T1 and T2 simultaneously.Target motion is simulated by the equation:

xT(t+1) = xT(t) +wT

where xT(t) is the target position at time t and wTis white Gaussian noise with constant covariance Q =0:005. Targets can be attacked and destroyed if the99% confidence interval of their position is less than¯threshold = 0:5 unit. The sensor makes one measurementduring each time period, and the measurement equationis:

z(t) = x(t) + v(t),

where x(t) is the state vector, and v(t) is zero mean whitenoise with constant variance, R = 0:03. Let the initialuncertainties in the position of T1 and T2, ¯1 and ¯2 areequal to 1 unit.Two sensor-scheduling approaches were imple-

mented. The first approach schedules the sensor to max-imize the information gain from the sensor measure-ment. The second approach schedules the sensor tomaximize the utility of measurements, which is definedas the inverse of the total number of sensor measure-ments required for bringing the targets to threshold un-certainty. The optimal measurement is determined usingexhaustive enumeration techniques. The change in un-certainty of the two approaches is shown in Fig. 4 andFig. 5. ITSM oscillates between the two targets withoutcollecting enough information on any one target longenough to successfully destroy either. The above exper-iments give an unfair advantage to the utility-based ap-proach since the optimization routine considers multiplesensor schedules simultaneously. In spite of this bias,these experiments offer an insight into the handicap suf-fered by ITSM due to its inability to take goal relatedparameters, like ¯threshold, and utility-based calculations


Fig. 4. Change in uncertainty of target tracks, while using an information-theoretic approach.

Fig. 5. Change in uncertainty of target tracks, while using a utility-based approach.

directly into consideration. On the other hand, marketsprovide a principled way to take the utility of a givenallocation to high-level goals directly into considera-tion during scheduling. These results are analogous tothose obtained by Castanon [5]. Operating under someassumptions, Castanon considered the problem of de-termining the optimal sequence of measurements of asingle sensor such that the probability that at-least onetarget is successfully located in a multi-cell grid is max-imized. Results demonstrated that the greedy approxi-mation to the optimal solution performed vastly betterthan an algorithm based on entropy minimization.2) MASM prioritizes using prices. Another advan-

tage of MASM may be due to its use of prices to pri-oritize tasks while ITSM schedules sensors so as to op-

timize the information gain from the environment. Al-though both ITSM and MASM used the same weightsto prioritize between tasks in the environment, price-based task prioritization has some inherent advantages.This is because a price-oriented approach has the abilityto implicitly reserve resources for future use by highpriority tasks, even if no high priority tasks are cur-rently in progress. For example, consider a situationwhere the first user is tracking a target and the rest ofthe users are in search mode. Both MASM and ITSMgive highest priority to the track task. The first userhas a high-budget for a track-bid and bids accordingly.However, during the tracking task, the prices associatedwith the sensing resources increases since the rate oftheir battery power usage during tracking is high (refer


Fig. 6. A comparison of the number of sensors used for measurements, based on whether target tracks are currently in progress or not.

to Section 5 for a description of how prices vary withrate of utilization). After the tracking task is completedand only when detection tasks are in progress, pricesof the sensor schedules would have increased. Conse-quently, sensors will be used at a slower rate during thedetection phase, effectively reserving sensors for futurehigher-priority tasks. However, ITSM has no method ofprioritizing between two tasks, except when both thetasks are currently in progress. Fig. 6 shows the num-ber of sensors used during different rounds of schedul-ing using MASM, where the number of sensors usedwhen tracking tasks are in progress is higher than thenumber of sensors used when only detection tasks arein progress. When only detection tasks are present, asignificant percent of sensors are resting, thereby pre-serving their battery power for future use.

8.2. MASM-Specific Performance Measures

In addition to our comparison with ITSM, wediscuss other MASM-specific performance measures,namely managing resource constraints, task deadlines,scalability, and surplus sharing. The purpose of theseexperiments is to show how the various “knobs” of amarket-based approach can be adjusted to affect perfor-mance.

8.2.1. Resource UsageAs explained in Section 5, MASM uses a taton-

nement process for enforcing resource constraints suchas battery power constraints’ using current and avail-able rates of utilization. Fig. 7 shows the price varia-tions of the first three batteries using a tatonnement rate¿ = 0:005. Figs. 8 and 9 show the fall in the energy of

the first three sensors’ battery with successive scheduleswith ¿ = 0:005 and ¿ = 0 respectively. It is clear thattatonnement process is successful in ensuring uniformusage of sensors and in keeping them functional till theend of network operation. For bandwidth also, a similarprocedure is adopted. The supply of capacity is con-stant and is proportional to tcom. During round t+1, ifthe price of channel is pt units/sec and bandwidth con-sumption is w, then the bound of the 1¡ _ one-sidedupper confidence interval of the expected time takento communicate, ´, is calculated based on monte-carlosimulations. The demand at pt is proportional to ´. Val-ues of ´ for different bandwidth usages are calculatedat the start of network operations and stored. The priceof a channel during the current round is

pt+1 = pt+ ¿(´¡ tcom):The price updates for process works as a soft constrainton channel capacity. That is, if the channel is too con-gested, then prices of the channel will increase till de-mand for channel capacity decreases. However, it is pos-sible that during certain schedules, the actual time takenfor communication is more than the prescribed limit.Figs. 10 and 11 show the time taken for communica-tion for two sample simulation runs with _¿ = 0:005 and¿ = 0 respectively.For the run in Fig. 10, the number of time slots

when time taken to communicate crossed the specifiedthreshold is 5. This is within the 0.01% tolerance limitspecified by the SM. For the run in Fig. 11, the numberof time slots where time taken to communicate crossedthe specified threshold is 124. After the 250th sched-ule, all batteries are exhausted, and the time taken tocommunicate drops to zero.


Fig. 7. Price variation of the first three sensors with schedule number for a sample run with tatonnement _¿ = 0:005.

Fig. 8. Energy usage for the first three sensors for a sample run with tatonement _¿ = 0:0059.

8.2.2. Task DeadlinesFor some high-value targets, users have a strict

deadline to destroy targets after initiating target trackswithin tkill schedules. We conducted experiments wherea certain fraction of the targets are high-valued. Weimplemented a user policy of increasing track bids bya factor k, if the detected targets are high-value. Forour initial experiments, we used a constant value of3. However, in the future, optimal k values could becalculated by using the market data. Users recorded anaverage track time of 7.1 schedules for T2 versus anoverall average of 15.3 schedules, showing how marketscan be used to enforce task deadlines.

8.2.3. Scalability AnalysisFor the current simulation environment, the IP-based

E-MASM formulation can optimally solve problemsizes of up to 10,000 resource bids within a threshold of10 cpu-seconds on 2.3 GHz PIV processor. Using theSGA-based A-MASM approach, problem sizes of up to50,000 resource bids can be solved to more than 98%optimality under the same conditions. Larger networkscan be accommodated by the approximate technique bycompromising on the final optimality. This capabilityindicates that MASM can be used for fairly large net-works, since spatial restrictions often mean that even ifa sensor network has thousands of sensors, only a fewcan be used for a given task at a time.


Fig. 9. Energy usage for the first three sensors for a sample run with tatonement _¿ = 0.

Fig. 10. Time taken for communication for a sample run vs. schedule number for a sample run with tatonnement _¿ = 0:005.

8.2.4. Surplus SharingIt is possible that sensor networks could have users

in a non-cooperative environment, where each agenthas an interest only in maximizing its own utility. Forexample, two different organizations could be sharingthe same sensor network resources. Such scenarios re-quire an incentive compatible auction methodology, tomake truth revelation the dominant strategy, and thus tomake the allocations optimal. For example, a paymentmechanism based on General Vickrey Auctions (GVA)[39] might be used to make truth revelation a weaklydominant strategy. GVA involves computation of n+1winner determination problems for every combinatorialauction to calculate the agent payments. In addition to

the computational complexity, the unique pricing mech-anism used by the CCA protocol to ensure real-time re-sponse precludes direct adaptation of GVA mechanisms.To understand the effect of strategic bidding, a pre-

liminary analysis can be conducted by formulating somesimple strategic bidding formulations and conductingsimulation experiments. For example, Walsh et al. [41]have analyzed the effects of strategic bidding on a com-binatorial auction based supply chain formation algo-rithm. It is important to note that the difficulty of analyz-ing the effects of strategic bidding actually underminesthe benefits of lying about true utilities for users. Anadded advantage with MASM is that there is disengage-ment between the users and the actual sensor network.


Fig. 11. Time taken for communication for a sample run vs. schedule number for a sample run with tatonnement _¿ = 0.

Though the users use the sensor network, the variousnetwork parameters including the prices of individualresources (which might indicate network bottlenecks)and the actual network parameters, like position of sen-sors, etc. is invisible to the individual user. Hence, thethreat presented by malicious entities that have clan-destinely gained access to use the sensor network isminimal. To analyze the effect of strategic bidding, itcan be assumed that agents play Bayes-Nash strategies[19]. However, calculation of Bayes-Nash equilibria isdifficult, except for the simplest of markets. An easiermethod for analyzing market behavior is to devise areasonable strategic bidding policy for users and studyresulting market behavior.A simple strategic bidding policy for MASM users

is to overstate their task utilities. To understand the logicbehind this policy, the pricing policy of CCA should beconsidered. If a MASM user bids a price P for a partic-ular task, the price it has to pay for resources allocatedto its bid is not directly based on P. Instead, for anyresource bundle allocated to the task during resourceallocation, the user usually pays only the sum of theprices of the resources comprising the resource bundle(see round termination step in CCA). Therefore, a userthat overstates its utility has the advantage of gettingpreferential treatment during resource allocation, whilenot having to pay any additional value for resourcesas compared to honest users. To analyze the impact ofstrategic bidding, experiments were conducted where acertain number of agents overstated their utility by afactor, k. The number of targets successfully destroyedby the users during the simulation experiment reflectsthe global performance of the market-based resource al-location. An individual user’s performance is measuredby its surplus defined as the difference between the to-

TABLE IVMarket Performance with Strategic Agent Behavior

Number of strategic agents 0 2Honest consumer surplus 1.22 ¡:012Strategic consumer surplus NA 2.3

tal utility it obtained from destroying the targets and thetotal price it paid to SM for buying resources during thesimulation.Two sets of experiments were conducted. In the first

set of experiments, all the users bid honestly. In thesecond set of experiments, two out of the five usersoverstated their utilities by a factor of two. The averagesurplus achieved by the honest and strategic agentsis shown in Table IV. As expected, strategic agentsbenefited from overstating their bid prices and theiraverage surplus increased from 1.22 to 2.3. The averagesurplus of the honest agents has decreased from 1.22 to¡0:012 as a result of strategic bidding.As shown by the preliminary analysis, CCA proto-

col encourages strategic behavior in a non-cooperativeenvironment. However, the benefits of strategic biddingcan be mitigated by using surplus sharing mechanismswhere the SM charges the users a fixed percentage oftheir surplus on each task bid. For example, a variant ofCCA, CCA-SS (Combination Auction Algorithm withSurplus Sharing) has been implemented where users arelevied an additional charge of fifty percent of their ex-pected surplus as calculated from their task bids. Underthis mechanism, the overall surplus to the strategic agentdecreased to ¡1:83. That is, they fare worse than thehonest users.Though surplus sharing provides an effective mech-

anism that works as a disincentive against overstating


task utilities, detailed experimentation is required to an-alyze the complete implications of strategic behavior forsome particular environment.

9. CONCLUSION

Market-based approaches provide a valuable frame-work for designing systems that must consider com-plex tradeoffs in their decision-making. Although muchwork has been done on individual aspects of market-based systems (e.g., auction algorithms, agent biddingstrategies, etc.), there is very little work on developing acomplete system in a complex real-world domain suchas sensor management. Therefore, one of our contribu-tions is to assess the design implications and compo-nents involved in building such a system. To addressthe issues that in the past have prevented the use ofmarket algorithms use for sensor networks, we have de-veloped techniques including auction mechanisms foraligning scheduling with mission objectives, approxi-mate techniques for handling computational complexity(A-MASM), pricing mechanisms for enforcing resourceconstraints and surplus sharing mechanisms to reducethe impact of strategic behavior. We have also shownthe system’s efficiency in a simulation environment, bycomparing with a weighted information-theoretic sensormanager. A crucial advantage of the proposed mech-anism is its flexibility. The proposed mechanism canbe easily adapted to an alternate sensor network sce-nario without much additional work by suitably creat-ing QoS charts for relevant network tasks (see Fig. 2)and by adapting price equations to reflect utilization ofthe appropriate network resources (see Section 5). Thiscontrasts with the current approaches based on approx-imate dynamic programming that are based on domain-specific and cumbersome formulation.We are currently working on implementing these

mechanisms using real-world sensor data. We are alsoworking on developing an alternate non-myopic sen-sor management approach for comparison with MASM.Two possible choices are i) Maximum marginal return(MMR) [4] sensor management approach that is ex-tended to incorporate non-myopic scheduling behaviorand ii) an approximate dynamic programming based ap-proach, similar to [37, 43, 44], that can handle multi-sensor, multi-task sensor management problems. In fu-ture, we plan to extend our market-based approach tosmart dust environments, which do not have a cen-tralized sensor manager. We also plan to extend ourinfrastructure to more effectively allocate informationgoods and develop a market-based situation assessmentcomponent. Current auction algorithms are generallydesigned to handle the allocation of tangible goods.However, we must adapt these e-commerce algorithmsto deal with information goods in the sensor-fusiondomain. Information goods (e.g., observations/reports),

and the sensors/processes that generate them, may beshared between agents to effectively complete compat-ible tasks, where applicable. For example, if two usersare engaged in the same sub-goal, and want the sameinformation, then a single commodity can be commu-nicated to both agents and will satisfy both of them.We plan to make use of the current research in digitalauctions [10], but will need to apply it appropriately toour domain.We would also like to develop a market-based sit-

uation assessment component that learns valuable sit-uation assessment cues from the market bids and priceinformation in the system. The situation assessment thatusers perform typically relates only to their immedi-ate surroundings and pertains to local information only.A global perspective can be obtained by observing theoverall market trends in the sensor manager’s situationassessment module. For example, a sudden increase inthe volume of bids from the users in one particular re-gion of the environment could suggest an impendingenemy attack in that region. Current prices can conveyinformation about when resources are in high demandand/or scarce. As part of this effort, collaborative filter-ing approaches, similar to those used by Amazon [23],could mine information from a combination of goal andbid behaviors to detect strategic patterns. Eventually,one could imagine a sensor management system thatrecommends a new information product (e.g., a targettrack) based on what previous users in similar situationshave selected.

REFERENCES

[1] A. Andersson and F. YggeManaging large scale computational markets.In Proceedings of the Thirty-first Hawaiian InternationalConference on System Sciences, Software Technology Track,Jan. 1998, 4—13.

[2] V. Avasarala, T. Mullen, D. L. Hall and A. GargaMASM: A market architecture for sensor management indistributed sensor networks.In Proceedings of the SPIE Defense and Security Symposium,Mar. 2005, 5813—30.

[3] V. Avasarala, H. Polavarapu and T. MullenSGA: An approximate combinatorial auction winner deter-mination algorithm for environments with strict real-timeconstraints.In Proceedings of the IEEE-WIC/Association for ComputingMachinery (ACM) International Conference on IntelligentAgent Technology (IAT-06), 571—578.

[4] C. F. CamererBehavioral Game Theory.Princeton, NJ: Princeton University Press, 2003.

[5] D. A. CastanonOptimal search strategies for dynamic hypothesis testing.IEEE Transactions on Systems, Man and Cybernetics, 25(July 1995), 1130—1138.

[6] D. A. CastanonApproximate dynamic programming for sensor manage-ment.In Proceedings of the Thirty-sixth Conference on Decisionand Control, Dec. 1997, 1202—1207.


[7] S. Cheng, E. Leung, K. Lochner, K. O’Malley, D. Reeves,J. Schvartzman and M. WellmanWalverine: A Walrasian trading agent.Decision Support Systems, 39, 2 (Apr. 2005), 169—184.

[8] R. V. Denton, E. Alcaraz, J. Knopow, J. Llinas and K. HintzTowards modern sensor management systems.In Proceedings of the Sixth Joint Service Data Fusion Sym-posium, June 14—18, 1993, 659—678.

[9] Y. Fujishima, K. Leyton-Brown and Y. ShohamTaming the computational complexity of combinatorialauctions: Optimal and approximate approaches.In Proceedings of the 16th International Joint Conference onArtificial Intelligence (IJCAI), 548—553.

[10] A. V. Goldberg and J. D. HartlineEnvy-free auctions for digital goods.In Proceedings of the Fourth Association for Computing(ACM) Conference on Electronic Commerce (EC-03), 29—35.

[11] D. L. Hall and J. LlinasHandbook of Multisensor Data Fusion.Boca Raton, FL: CRC Press LLC, 2001.

[12] D. L. Hall and S. McMullenMathematical Techniques in Multisensor Data Fusion (2nded.).Norwood, MA: Artech House, 2004.

[13] K. J. Hintz and E. S. McVeyMulti-process constrained estimation.IEEE Transaction on Systems, Man, and Cybernetics, 21, 1(Jan./Feb. 1991), 434—442.

[14] K. J. Hintz and G. McIntyreGoal lattices for sensor management.In Proceedings of the Signal Processing, Sensor Fusion andTarget Recognition VIII: SPIE International Symposium onAerospace/Defense Sensing and Control, Apr. 1999, 249—255.

[15] K. J. Hintz and J. MalachowskiDynamic goal instantiation in goal lattices for sensor man-agement.In Proceedings of the SPIE, May 2005, 93—99.

[16] J. S. JordanThe competitive allocation process is informationally effi-cient uniquely.Journal of Economic Theory, 28, 1 (1982), 1—18.

[17] M. Kalandros and L. Y. PaoCovariance control for multisensor systems.IEEE Transactions on Aerospace Electronic Systems, 38, 4(Oct. 2002), 1138—1157.

[18] K. KastellaDiscrimination gain for sensor management in multitargetdetection and tracking.In Proceedings of the IEEE–Systems, Man and Cybernetics(SMC)/International Association for Mathematics and Com-puters in Simulation (IMACS) Multi-Conference on Computa-tional Engineering in Systems Applications (CESA’96), July1996, 167—172.

[19] J. D. KattarA solution of the multi-weapon, multi-target assignmentproblem.WP-26597, The MITRE Corporation, Bedford, MA, 1986.

[20] M. Kolba and L. L. CollinsInformation based sensor management in the presence ofuncertainty.IEEE Transactions on Signal Processing, 55, 6 (June 2007),2731—2735.

[21] L.WalrasElements of Pure Economics.Homewood, IL: Irwin, 1954.

[22] V. Lesser, C. L. Ortiz and M. TambeDistributed Sensor Networks: A Multiagent Perspective.Boston, MA: Kluwer Academic Publishers, 2003.

[23] G. Linden, B. Smith and J. YorkAmazon.com recommendations: Item-to-item collaborativefiltering.IEEE Internet Computing, 7, 1 (Jan. 2003), 76—80.

[24] G. Mainland, D. C. Parkes and M. WelshDecentralized, adaptive resource allocation for sensor net-works.In Proceedings of the Second The Advanced ComputingSystems Association (USENIX)/Association for ComputingMachinery (ACM) Symposium on Networked Systems Design& Implementation (NSDI’05), Apr. 2005, pp. 7—13.

[25] G. A. McIntyre and K. J. HintzAn information-theoretic approach to sensor scheduling.In Proceedings of the SPIE: Signal Processing, Sensor Fu-sion, and Target Recognition V, Apr. 1996, 304—312.

[26] G. A. McIntyre and K. J. HintzSensor management simulation and comparative study.In Proceedings of the SPIE International Symposium onAerospace/Defense Sensing & Control: Signal Processing,Sensor Fusion, and Target Recognition VI, Apr. 1997.

[27] J. MoffatComplexity Theory and Network Centric Warfare.CCRP Publication Series, 2003.

[28] T. Mullen, V. Avasarala and D. L. HallBringing the market to level 4 fusion: Investigation ofauction methods for sensor resource allocation.In Proceedings of the Military Sensing Symposia (MSS)National Symposium on Sensors and Data Fusion (NSSDF),May 2005.

[29] T. Mullen, V. Avasarala and D. L. HallCustomer-driven sensor management.IEEE Intelligent Systems, 21, 2 (2006), 41—49.

[30] R. Muraleedharan and L. OsadciwDecision in a building access system using swarm intelli-gence and posets.In Proceedings of the Conference on Information Sciencesand Systems, Mar. 2004, 414—419.

[31] J. M. NashOptimal allocation of tracking resources.In Proceedings of the IEEE Conference on Decision andControl, Dec. 1977, 1177—1180.

[32] J. Ostwald, V. Lesser and S. AbdallahCombinatorial auction for resource allocation in a dis-tributed sensor network.In Proceedings of the The 26th IEEE International Real-TimeSystems Symposium (RTSS’05), Dec. 2005, pp. 266—274.

[33] P. Padhy, R. K. Dash, K. Martinez and N. R. JenningsA utility-based sensing and communication model for aglacial sensor network.In Proceedings of the Fifth International Joint Conference onAutonomous Agents and Multiagent Systems (AAMAS’06),May 2006, 1353—60.

[34] P. A. SamuelsonFoundations of Economic Analysis.Cambridge, MA: Harvard University Press, 1947.

[35] T. SandholmAlgorithm for optimal winner determination in combinato-rial auctions.Artificial Intelligence, 135, 1—2 (Feb. 2002), 1—54.

[36] W. SchmaedekeInformation-based sensor management.In Proceedings of the SPIE: Signal Processing, Sensor Fu-sion, and Target Recognition II, Apr. 1993, 156—64.

[37] M. K. Schneider and C.-Y. ChongA rollout algorithm to coordinate multiple sensor resourcesto track and discriminate targets.In Proceedings of the SPIE Conference on Signal Processing,Sensor Fusion and Target Recognition, Apr. 2006.


[38] J. B. Shoven and J. WhalleyApplying General Equilibrium.New York: Cambridge University Press, 1992.

[39] H. R. Varian and J. K. MacKieMasonGeneralized Vickrey auctions.Department of Economics, University of Michigan, 1994.

[40] C. A. Waldspurger, T. Hogg, B. A. Huberman, J. O. Kephartand S. StornettaSpawn: A distributed computational economy.IEEE Transactions on Software Engineering, 18, 103—117.

[41] W. E. Walsh, M. P. Wellman and F. YggeCombinatorial auctions for supply chain formation.In Proceedings of the Association for Computing Machinery(ACM) Conference on Electronic Commerce (EC’00), Oct.2000, 260—269.

[42] W. E. Walsh, M. P. Wellman, P. R. Wurman and J. K. MacKie-MasonSome economics of market-based distributed scheduling.In Proceedings of the International Conference on Dis-tributed Computing Systems, May 1998, 612—621.

[43] R. Washburn, M. Schneider and J. FoxStochastic dynamic programming based approaches to sen-sor resource management.In Proceedings of the Fifth International Conference onInformation Fusion, 608—615.

[44] J. L. Williams, J. W. Fisher, III and A. S. WillskyAn approximate dynamic programming approach to a com-munication constrained sensor management problem.In Proceedings of the Eighth International Conference onInformation Fusion, July 2005.

[45] F. YggeMarket-Oriented Programming and its Application to PowerLoad Management.Ph.D. thesis, Department of Computer Science, Lund Uni-versity, ISBN 91-628-3055-4, CODEN LUNDFD6/(NFCS-1012)/1-224/(1998).


Viswanath Avasarala obtained his bachelor’s degree from the Indian Institute ofTechnology, Madras in 2001. He has a Ph.D. in information sciences and technologyfrom Pennsylvania State University. In his Ph.D. thesis, he investigated the useof market-oriented multi-agent frameworks for real-time and distributed sensormanagement.His primary research interests are in data mining, market-oriented programming

and multi-agent collaboration. Since 2006, he has been working as an informationscientist at GE Global Research. At GE, Dr. Avasarala uses artificial intelligencebased methodologies to solve research problems in machinery diagnostics andprognostics, fleet logistics and health management, stochastic simulation, multi-objective optimization and medical imaging. He has published 3 journal papers, 1book chapter and more than 10 conference papers in these areas. He has also chairedand/or acted as an organizing committee member of several reputed internationalconferences in soft computing. He is an active IEEE member and the current chairof Schenectady chapter of the IEEE Computational Intelligence Society.

Tracy Mullen is an assistant professor at the College of Information Sciencesand Technology at The Pennsylvania State University. She received her Ph.D. incomputer science from the University of Michigan.She has previously worked at Lockheed Martin, Bellcore, and NEC Research.

Dr. Mullen’s research is focused on computational markets and multiagent systemsto support decentralized decision making. Specific interests include market-basedresource allocation for sensor management, prediction markets, and trading agentsand markets. Her research papers have been published in ACM Transactions onInternet Technology, Decision Support Systems, Electronic Commerce Research, IEEEComputer,Mathematics and Computers in Simulation, and Operating Systems Review,among others.

David L. Hall is a professor in the Penn State College of Information Sciences andTechnology (IST) where he leads the Center for Network Centric Cognition andInformation Fusion. Previously he was the associate dean for Research and GraduatePrograms in Penn State’s College of Information Sciences and Technology, wherehe was responsible for overseeing IST’s Ph.D. and Master’s Degree programs aswell as research grant administration and leadership of IST faculty research efforts.Hall joined IST from the University’s Applied Research Laboratory (ARL) where heserved as associate director and senior scientist. At ARL, he oversaw the 150-personInformation and Network Systems Office, comprised of four divisions: InformationScience and Technology, Navigation Research and Development, Systems andOperations Automation, and Communications Science and Technology. He hasmore than 25 years of experience in research, research management, and systemsdevelopment.He is the author of more than 200 papers, reports, books and book chapters,

and he has delivered numerous lectures on his research, research management, andartificial intelligence. Hall has been named as an IEEE fellow for his contributionsto data fusion and he was awarded the DoD Joe Mignona National Data Fusionaward for his contributions and leadership in data fusion.


with the advent of small, inexpensive, low-power a market ... · be a promising path for...

Documents