OCTOBER 1981 LIDS-R-1156 SURVEILLANCE AND TARGET TRACKING Edited By Michael Athans Wilbur B. Davenport, Jr. Elizabeth R. Ducot Robert R. Tenney Proceedings of the Fourth MIT/ONR Workshop on Distributed Information and Decision Systems Motivated by Command-Control-Communication (C3) Problems Volume I June 15 - June 26, 1981 San Diego, California ONR Contract No. N00014-77-C-0532 PREFACE This volume is one of a series of four reports containing contri- butions from the speakers at the fourth MIT/ONR Workshop on Distributed Information and Decision Systems Motivated by Command-Control-Communication (C3 ) Problems. Held from June 15 through June 26, 1981 in San Diego, California, the Workshop was supported by the Office of Naval Research under contract ONR/N00014-77-C-0532 with MIT. The purpose of this annual Workshop is to encourage informal inter- actions between university, government, and industry researchers on basic issues in future military command and control problems. It is felt that the inherent complexity of the C3 system requires novel and imaginative thinking, theoretical advances and the development of new basic methodol- ogies in order to arrive at realistic, reliable and cost-effective de- signs for future C3 systems. Toward these objectives, the speakers, in presenting current and future needs and work in progress, addressed the following broad topics: 1) Surveillance and Target Tracking 2) Systems Architecture and Evaluation 3) Communication, Data Bases & Decision Support 4) C3 Theory In addition to the Workshop speakers and participants, we would like to thank Dr. Stuart Brodsky of the Office of Naval Research, and Ms. Barbara Peacock-Coady and Ms. Lisa Babine of the MIT Laboratory for Information and Decision Systems for their help in making the Workshop a success. Cambridge, Massachusetts MichaeZ Athans October 1981 Wilbur B. Davenport, Jr.EZlizabeth R. Ducot Robert R. Tenney -1- SURVEILLANCE AND TARGET TRACKING FOREWORD ........................... .......................... iv DATA DEPENDENT ISSUES IN SURVEILLANCE PRODUCT INTEGRATION Dr. DanieZ A. Atkinson ........................................ 1 MEMORY DETECTION MODELS FOR PHASE-RANDOM OCEAN ACOUSTIC FLUCTUATIONS Professor HariZaos N. Psaraftis, Mr. Anatassios Perakis, and Professor Peter N. Mikhahelvsky ................................. 35 DETECTION TRESHOLDS FOR MULTI-TARGET TRACKING IN CLUTTER Dr. Thomas Fortmann,Professor Yaakov Bar-ShaZom, and Dr. MoZZy Scheffe ...............................................41 MULTISENSOR MULTITARGET TRACKING FOR INTERNETTED FIGHTERS Dr. Christopher L. Bowman .......................................4 MARCY: A DATA CLUSTERING AND FUSION ALGORITHM FOR MULTI-TARGET TRACKING IN OCEAN SURVEILLANCE Dr. MichaeZ H. Moore ............................................ 65 AN APOSTERIORI APPROACH TO THE MULTISENSOR CORRELATION OF DISSIMILAR SOURCES Dr. MichaeZ M. Kovacich ......................................... 99 A UNIFIED VIEW OF MULTI-OBJECT TRACKING Drs. Krishna R. Pattipati, Nils R. SandelZ, Jr., and LesZie C. Kramer ................................................ 115 OVERVIEW OF SURVEILLANCE RESEARCH AT M.I.T. Professor Robert R. Tenney ...................................... 137 A DIFFERENTIAL GAME APPROACH TO DETERMINE PASSIVE TRACKING MANEUVERS Dr. PauZ L. Bongiovanni and Professor Pan-T. Liu ............... 149 -ii- DESCRIPTION OF AND RESULTS FROM A SURFACE OCEAN SURVEILLANCE SIMULATION Drs. Thomas G. Bugenhagen, Bruce Bundsen, and Lane B. Carpenter ............................ 17......... AN OTH SURVEILLANCE CONCEPT Drs. Leslie C. Kramer and NiZs R. SandeZZ, Jr .................. 193 APPLICATION OF AI METHODOLOGIES TO THE OCEAN SURVEILLANCE PROBLEM Drs. Leonard S. Gross, MichaeZ S. Murphy, and Charles L. Morefield ......................................... 209 A PLATFORM-TRACK ASSOCIATION PRODUCTION SUBSYSTEM Ms. Robin Dillard .............................. ........ 215 -i-li- SURVEILLANCE AND TARGET TRACKING FOREWORD Almost three full days of this year's Workshop were devoted to surveillance and related issues. The discussion sessions on each day were lively and provocative. Two major themes characterized the re- marks of the participants: 1) the design of surveillance systems and algorithms are critically dependent on the other parts of the C3 system to which they must interface, and 2) there remains a great deal of work to be done before surveillance of a complex environment is well understood. The objective of a surveillance system is to provide an accurate picture of the environment to the other parts of a C3 systems. There are many users of the surveillance information; the only general state- ment that can be made is that different users will require information of different types, of different levels of aggregation, and with dif- ferent priorities. There is no unique point at which the performance of a surveillance system can be measured; it must be evaluated in the context of the other C3 elements and the overall command objectives. On the technical side, there was general agreement that many open questions remain. Single sensor, single target, high signal-to-noise ratio problems are relatively well understood (from a theoretical point of view). Even the addition of either multiple sensors, multiple targets, or low signal-to-noise ratios one at a time produces problems which can usually be addressed with current theory. Taken in pairs or all together, however, particularly when limited, unreliable communications are present, brings one to a theoretical void. In fact, in our opinion, it is not clear that general surveillance issues involving multiple sensors, multiple targets, and fusion centers can be formulated reasonably in a relevant context unless data communica- tions constraints are explicitly included in the very problem formulation, Tiv- The surveillance papers included here are not in the order of presen- tation. Atkinson's paper first provides an overview of surveillance issues and nomenclature. The remaining papers are roughly sequenced along the physical to abstract dimension. Psaraftis, Perakis, and Mikhahelvsky present a normative model on the ocean environment suitable for use in a long range detection problems. The next four papers Fortmann et. al., Bowman, Moore, and Kovacich all address one of the currently popular topics in surveillance, multi-object tracking. Pattipati et. al., then presents a problem formulation which will allow some current adhoc ap- proaches to this problem to be placed in a more rigorous framework. The subsequent papers move to a larger set of issues. Tenney surveys work at M.I.T. on surveillance, communication, and control; Liu and Bongiovanni look at a sensor placement problem. The next pair of papers discuss surveillance systems architectures; Bugenhagen et. al., in an oceanic context, Land Kramer and Sandell in an early air warning setting. We conclude with two papers on the potential role of artificial intel- ligence techniques in surveillance systems; Gross et. al., in a general sense and Dillard based on a specific production rule system. _v_ DATA DEPENDENT ISSUES IN SURVEILLANCE PRODUCT INTEGRATION David A. Atkinson CTEC, Inc. 7777 Leesburg Pike FaZZs Church, Virginia 22043 (CTEC Publication No. SAD-8Z-036 - June 8, 1981) DATA DEPENDENT ISSUES IN SURVEILLANCE PRODUCT INTEGRATION D.A. Atkinson CTEC, Inc. SUMMARY This analysis is focused on the design problems encountered in the integration of ocean surveillance data that arise from the particular nature of theinformation delivered to and required from the system. The starting point is a definition of the ocean surveillance product, its uses, and the requirements specified for it. The general nature of the input data is covered in anoverview of the characteristics of sensors and sources. Problems associated with the distribution of processing, in particular the target classification analysis, in a surveillance network are described. The important issue of multi-source integration (correlation) is covered by raising specific outstanding problems. Finally, the impact of these problems on surveillancesystem design approaches is discussed. 1. THE OCEAN SURVEILLANCE PRODUCT (OSP) The six basic elements present in the OSP are listed in Viewgraph 2, together with a general characterization of the intelligence and operationaluses of this product. These two general categories of use are clearly interdependent, and the seventh element listed as a part of the evaluated OSP indicates one form of this relationship. The basic elements and uses arethose identified in the Navy's Ocean Surveillance Master Plan [1]. Different representations of the basic OSP information appear to be best suited to the needs of specific users. A tabulation of the track representations andclassification levels that would support the needs of specific consumers of the OSP is presented in Viewgraph 3. Operational commanders with theirattention focused on developments in the immediate future can make do with current position and velocity estimates, while intelligence analysts may require historical information. The lower level of geographic resolution for national users is dictated by the volume of information, although they may also require occasional high-resolution inputs. The classification levelsindicated are presented to stimulate further debate, rather than as definitive specifications. Quantitative performance requirements for location accuracy and timeliness may be specified by using the radius of uncertainty (ROU) definedin Viewgraph 4, and illustrated graphically in Viewgraph 5. A significant feature of this parameter is its dependence on the sum of the intervalrequired for processing and transmitting OSP data and the interval between successive observations. The name adopted in the OSMP [El is somewhatinappropriate: "radius of containment" describes this parameter more accurately. The measures of classification performance are rather obvious ones; however, they are quite difficult to determine in practice. The area of data association (correlation) is one in which definitive measures of performance are very difficult to develop, as is illustrated in a number of Navy studies [2], [3], and [4] of this problem. The gate areas andprobabilities of correct choice for specified decision rules, presented in Viewgraph 6, may be used to estimate performance. An interesting feature of these measures is the fact that signature descrimination capability (PSD 1) lowers the effective target density by a simple factor. A formalproof of this intuitive result is presented in the Appendix of a report on an analytic correlation performance model [5]. 2. SENSORS AND SOURCES The large number of sensors that can contribute the basic data used to create the OSP precludes a detailed analysis of the characteristics of individual systems in this report. The overview presented here attempts to identify characteristics that significantly influence system design.Viewgraph 7 presents sensors grouped into three general categories: active; passive (monitoring) involuntary (signals); and passive (monitoring) voluntary (signals). The passive voluntary category includes the sensors vulnerable tothe use of emissions control (EMCON) policies by hostile forces. The basic classification data provided by sensors may be characterized as either image data or signature data (Viewgraph 8). The principles used to arrive at classification using such data are presented in Viewgraph 9. The comparison of images is essentially an interpretive process. In view of the very significant problems associated with automatic image recognition, the processing rate problem arising from the requirement for interactive analysisis likely to persist for some time. Signature comparison is normally made using tolerance intervals or a statistical distance functional defined in the abstract "emissions" parameter space. The sensor position data we refer to here is the information provided by the primitive signal measurements as opposed to information that may beinferred from subsequent processing. Examples such as the fix, line of bearing (LOB), and the hyperbolic line of position arising from a timedifference measurement are illustrated in Viewgraph 10. The primitive classes of position and velocity information are tabulated in Viewgraph 11, with anindication of the processing approach required to incorporate such observations into an updated state estimate. Time and Doppler shiftdifferences differ from the others in the sense that a correlation of observed signals at two sites is required to extract the location information. Inaddition, curvilinear lines of position can yield ambiguous intersections. Statistical filters, such as the extended Kalman filter, can be used for updates when the interval between observations is comparable to or less thanthe typical interval between target maneuvers. The use of this approach in a system delivering time and Doppler difference data is described by Fortmannand Baron [6]. Some systems, in particular mosaic IR detectors, generate track segment observations as a part of their signal detection logic. Thesemeasurements may easily be converted into a complete (position + velocity) observation of the state vector. A major problem arises when the revisit interval becomes much longer than the interval between maneuvers. This is particularly severe for surface traffic in littoral areas where the constraint that "ships cannot walk" causes frequent maneuvers. In these situations theresulting state estimates may not support the operational user's requirements for position projection. 3 The characteristic classification and location information provided by a number of sensors that could contribute to the OSP is tabulated in Viewgraph12. The sensitive nature of information on the accuracy of the data provided by such sources precludes a realistic discussion at this conference.Obviously, accuracy is a very important consideration in arriving at an evaluation of the potential contribution of a source to the OSP. A number of factors associated with the operating principles and designof sensors have a significant impact on the timeliness of the data delivered by the sensor systems. Some of these factors are indicated in Viewgraph 13.Propagation delays are significant for acoustic systems detecting signals at long range. The time required for signal integration in order to suppressnoise and achieve the desired frequency resolution must also be considered, since it can restrict coverage rate potential in large area surveillancesystems. Processing to achieve a position fix and/or to classify the target detected also produce delays in delivery. Data transmission delays are a potential problem in multisite sensor systems that require network coordination and data exchange. 3. SURVEILLANCE NETWORKS A surveillance system is an interconnected network leading from the basic sensing devices through various intermediate processing stages todelivery of an integrated product to the users. The processing stages are distributed in both space and time. The determination of the most effective configuration for an ocean surveillance network involves a vast number ofissues, and our limits of time and competence require a focus on a specific example of a network configuration problem. An abstract outline of data flow from source to user is presented in Viewgraph 14. The sensor/source segment encompasses reception of the basic signal, processing to detect (extract from the noise background) and integrate(convert to the desired form and resolution) this signal, and subsequent processing to refine the OSP elements delivered by the source. This stage isnormally followed by processing at a regional or national center that serves to integrate the outputs of multiple sources (MSI), and to append prepositioned information on the target in certain cases. Preprocessing consists of data conversion and analysis employing only the data supplied by the source and prepositioned intelligence data (eg. a hull to emittercorrelation (HULTEC) analysis). The user may be national command, a fleet or regional command, or a battle group or individual unit. The networkconfiguration issue considered is the location of classification/signature processing in the network. The diagram of the assumed source segment in Viewgraph 15 illustrates two of the options. The source network consists of detectors connected to initialprocessing facilities that extract the signature and location (assumed to be a LOB) data from the basic signal. One option is to place the classification processing at these initial processing facilities. Each of the initial facilities reports to a central source evaluation center. At this center,reports with matching signatures may be processed to obtain a position fix. In addition, the classification processing that matches these signatures with a signature fingerprint file could take place at the evaluation center. The third option is deferral of classification processing until the data aredelivered to a regional MSI processor. 4 The distribution options and some of the factors that must influence the choice are presented in Viewgraph 16. The classification processing, as well as the association analysis that yields a position fix, clearly enhances theutility of the OSP delivered downstream. Processing centers represent potentially vulnerable elements of the system. Thus, early classification will provide the option of delivery of useful information in the event of failure at the source evaluation center or at a regional center. The fullvalue of this option may be realized only if the user is equipped to process the location information (LOB) that can be delivered by an initial processor. Central classification can result in an increased level and likelihood oftarget identification because the information resulting from multi-source association is available. In addition, this option will ensure that users atvarious command levels are using a consistent situation picture. The data transmission loads require a detailed analysis in specific cases, because early classifiaction requires feedback of fingerprint updates while centralclasification may involve transmission of more-elaborate signatures. The basic nature of signature processing has a significant impact on network design. Two basic uses of signature data are outlined in Viewgraph 17, which also discusses the associated data requirements. The differentiation between establishing a track designator and classification maybe blurred by incorporating signature data for unidentified targets in theFingerprint File. The most effective way to assess the confidence of a classification based on signature data is to compute the relative likelihood of alternative matches. This requires relatively complete files of observedsignatures. This data should be regarded as relatively dynamic, with an update frequency determined by the time dependence of the emission parameters and alterations in deployment. The modeling of emission parameter time series for individual targets may be desirable in some cases. Cases in which parametric separation is not sufficient to support unique classification will require a "dynamic signature" approach. The basicprinciple is presented in Viewgraph 18. If a combination of parametric and geographic separation is sufficient to permit a track designator (tracking)analysis, then the ability to maintain classification of a target followsdirectly from the ability to track it. Initial classificaion must be established, but this is the case for any signature-based analysis. Thus,initial classification is always based on multi-source correlation, but trackmaintenance is sufficient thereafter. 4. MULTI-SOURCE INTEGRATION (MSI) The MSI problem and the closely associated multi-target tracking problem are technically sweet. This must account, in part, for the large number ofpotential solutions to these problems [3] which have been advanced in recentyears. An analysis of advanced approaches to the multi-target trackingproblem can be found in a review by Bar-Shalom [7]. More recent work isexemplified by the JPDA analysis of Bar-Shalom, Fortmann, and Schaeffe [8] and the highly regarded Bayesian multiple hypothesis system developed by Reid [9]. The basic principles that must be applied to the association of datafrom multiple sources are indicated in Viewgraph 19. Position comparison atsimultaneous or nearly simultaneous view employs relatively simple algorithms,but it is likely to be successful only when the two sources produce very 5 accurate position measurements. Track matching also employs position data, as this is the data element common to nearly all OSP sources. Characteristic consistency tests are particularly useful where signature analysis has provided a partial classification (eg. identified a specific type of emitter). A Bayesian approach to the use of characteristic consistency inassociation analysis has been developed [10. Operational characteristics can also be used in MSI, as they serve to pin point the time of anticipated events. The difficulty associated with adoption of an approach that relies onthe typical operations of targets is its vulnerability to deception and exploitation in times of crisis. An appreciation of the OSP processing at a regional processing center may be obtained from the hypothetical data flow illustrated in Viewgraph 20. The trade-off between time required for preprocessing, such as HULTEC analysis,and the resulting increase in efficiency of the selection of candidate tracks for association is an interesting issue in the design of the processing subsystem. Motion models, which may involve both position and signature projection, are employed in both the evaluation of candidates and the update of state vector and parameters which follows an association decision. These applications are different, and it is by no means evident that the same models should be used at both processing stages. The provision of review proceduresto detect and recover from association errors is critically important-in automatic decision systems. A list of some interesting problems associated with MSI analysis is presented in Viewgraph 21. Problems associated with the impact of targetmaneuvers on position projections have long been recognized. The impact of volitional maneuvers is outlined in Viewgraph 22. The uncertainty ofstatistical distributions may lead to contradictory results when different decision rules are applied [11], and the identification of an optimal rule must be based on empirical analysis. The importance of maneuver detection and adjustment procedures for tracking filters applied to ocean surveillance is well known [12]. Ambiguity is always present when decisions must be based on uncertain information. In considering approaches to this problem, it is useful to distinguish between their contribution to an actual resolution of the ambiguity and their potential to clarify our picture of the ambiguoussituation. The basic considerations are presented in Viewgraph 23, where the rapid growth in complexity and processing demands associated with complex decision algorithms are also noted. A second problem associated with representation is the determination of a form of representation of situation plot ambiguities that is useful to operational commanders Classification chains occur in MSI when data containing different degrees of target identification is associated. Viewgraph 24 presents an interesting example where accurate positional data arising from a radar sensormay be associated with an ELINT report that also has high positional accuracy. Subsequent association with a series of low position accuracy HFDFreports using a combination of characteristic consistency (ie. emitter is compatible with target identified by HFDF) and track matching provides a complete target ID for the radar track. The opimal use of such classification 6 chains to upgrade the OSP is an interesting open problem. This example may also be used to illustrate the potential problems arising from data bias. Ifthe reference frames employed by the three systems are inconsistent, statistical position comparison procedures will prove ineffective. 5. CONCLUSION This analysis is , admittedly, rather long on problems and short on solutions. The time when the design of a surveillance system will be a straightforward engineering task still lies well ahead of us. The objective of this report will be realized if it functions as a primer to introduce you to the problems that can arise in the design of a real world ocean surveillance system. At this point, a few comments on design approaches seem appropriate. * Multiple Data Paths: The provision of multiple data paths fromsource to user will enhance the survivability of the system. The problem with such a design arises from the temptation to use these paths simultaneously. The assumption that association of duplicatereports of an event will be trivial, because the data delivered will be identical, is naive. Simulation Testing: Simulation testing of tracking and decision algorithm performance may be very useful. It is clearly essential in the case of data sources still in the design phase. However,care should be exercised to conduct the tests in a ruthless manner. Significant errors and/or systematic bias should be incorporated in the simulated data. Robustness against such errors is essentialin processing data from deployed sensors. * Automation: High data rates may dictate a hands-off decision logic, but even in such cases an interactive interface is an importantdevelopment tool. These systems can fail in ingenious ways. Sophisticated Decision Algorithms: The need for fancy analytical footwork to achieve some improvement in the processing of ambiguous data indicates a failure in overall system design. An effective ocean surveillance system should provide relatively unambiguousdata. 7 REFERENCES 1. "Ocean Surveillance Master Plan," February 1980. 2. Kullback, J.H., and Owens, M.E.B., "Multisensor Correlation for OceanSurveillance: Problems and Limitations," NRL Report 7258, May 1972. 3. Wiener, H.L., Willman, W.W., Goodman, I.R., and Kullback, J.H., "NavalOcean-Surveillance Correlation Handbook, 1978," NRL Report 8340, Oct.1979. 4. Wiener, H.L., Distler, A.S., and Kullback, J.H., "Operational and Implementation Problems of Multi-Target Correlator-Trackers," Proc. 1979 IEEE Conf. on Decision and Control, Dec. 1979 5. Atkinson, D.A., CTEC Report ITSS-CTEC-7, May 1981 6. Fortmann, T.E., and Baron, S., "Problems in Multi-Target Sonar Tracking," Proc. 1978 IEEE Conf. on Decision and Control, Jan. 1979. 7. Bar-Shalom, Y., "Tracking Methods in a Multitarget Environment," IEEETrans. Auto. Control, AC-23, 618, August 1978. 8. Bar-Shalom, Y., Fortmann, T.E., and Schaeffe, M., "Joint Probabilistic Data Association for Multiple Targets in Clutter," Proc. 1980 Conf. on Information Sciences and Systems, March 1980. 9. Reid, D.B., "The Application of Multiple Target Tracking Theory to OceanSurveillance," Proc. 1979 IEEEE Conf. on Decision and Control, Dec. 1979. 10. Atkinson, D.A., "A Bayesian Analysis of Surveillance Attribute Data,"Proc. 1980 IEEE Conf. on Decision and Control, Dec. 1980 11. Atkinson, D.A., "A Comparison of Probability Gates and Weights in theSurveillance Association Problem," Proc. 3rd MIT/ONR Workshop on C3 Problems, Vol 4, Dec. 1980. 12. Corman, D.E., "OTH/DC&T Engineering Analysis Vol. 8: Evaluation ofSurface Ship Tracking Algorithms," JHU/APL, April 1979. 8 SURVEILLANCE INTEGRATION BRIEFING OUTLINE DEFINE THE OCEAN SURVEILLANCE PRODUCT (OSP), ITS USES, AND QUANTITATIVE REQUIREMENTS CHARACTERIZATION OF SENSORS/SOURCES AND THE LOCATION AND CLASSIFICATION DATA THEY PROVIDE * DISTRIBUTION OF PROCESSING IN SURVEILLANCE NETWORKS -- DATA REQUIREMENTS AND UTILITY MULTI-SOURCE INTEGRATION PRINCIPLES AND PROBLEMS SURVEILLANCE SYSTEM DEVELOPMENT PROBLEMS: A SELECTED LIST DEFINITION OF THE OSP AND ITS USES OSP FLEMENTS 1. TIME (EVENT OR RECEIPT OF SIGNAL) 2. STATE VECTOR (POSITION-VELOCITY COMPONENTS) 3. ACCURACY OF STATE VECTOR (CONTAINMENT ELLIPSOID) 4, CLASSIFICATION (UNIQUE ID, CLASS, TYPE, CATEGORY) 5. CONFIDENCE OF CLASSIFICATION (PROBABILITY THAT CLASSIFICATION IS CORRECT) 6. TRACK DESIGNATOR (UNIQUE ASSOCIATION INDICATOR) OSP USES OPERATTONAL INTELLIGENCE 1. ALTER ALERT STATUS 1. STRATEGIC I&W 2. DIRECT MOVEMENT OF FORCES 2. ORDER OF BATTLE 3. UTILIZATION OF WEAPONS AND 3. TACTICAL I&W SENSORS 4. SCIENTIFIC INTELLIGENCE EVALUATED OSP CONTAINS 7. INFORMATION SUPPORTING TACTICAL INDICATIONS AND WARNING (I&W) (RANGE AND CAPABILITIES OF WEAPONS AND SENSORSJ ETC.) 10 USERS AND DATA REPRESENTATION MINIMUM TRACK USEFUL USER APPLICATIONS REPRESENTATION CLASSIFICATION NATIONAL STRATEGIC I&W T3 C2 COMMAND AREA/FLEET SITUATION MONITORING T2 OR T3 C1 OR C2 COMMANDS SENSOR TASKING AREA I&W BATTLE GROUP TACTICAL I&W T1 C3 UNIT TARGETING ORGANIC SENSOR TASKING KEY TRACK REPRESENTATIONS Ti: STATE VECTOR + VARIANCE T2: TIME ORDERED POSITIONS T3: CURRENT CONTAINMENT AREA CLASSIFICATION C1: COMPLETE ID C2: CLASS OR TYPE C3: CATEGORY (FRIEND-FOE; FISH-FOWL) QUANTITATIVE PERFORMANCE REQU I LOCATION ACCURACY AND TIMELINESS VIA ROU (RADIUS OF UNCERTAINTY) ROUMAX = ( (S * T)2 A 2 )1/2 A = MEASUREMENT ACCURACY S = MAXIMUM TARGET SPEED T = TU + TR WITH TU = MEASUREMENT TO UPDATE INTERVAL TR = REVISIT INTERVAL CLASSIFICATION PERFORMANCE MEASURES F c = FRACTION OF OBJECTS CLASSIFIED PC = PROBABILITY OF CORRECT CLASSIFICATION T C = INTERVAL FROM DETECTION TO CLASSIFICATION (THESE PARAMETERS MAY BE SPECIFIED FOR VARIOUS TARGET TYPES) 12 ILLUSTRATION OF THE ROU CONCEPT RO A - -I - I I I I t Tm Tud Tm Tud Ti Tud 4 TuH 13 DATA ASSOCIATION PERFORMANCE ESTIMATES CONTAINMENT AREAS: PROJECTION AN =2 r Q a 2 (-LN0) FEASIBILITY AM = r ((S TR)2 + (3 cra)2) TR = REVISIT INTERVAL S = TARGET SPEED cra = MEASUREMENT ACCURACY = SIGNIFICANCE LEVEL Q = TRACKING EFFICIENCY PARAMETER (1< Q<6) PROBABILITY OF CORRECT ASSOCIATION: Z = Ap PSD = TRAFFIC DENSITY PSD= SIGNATURE DISCRIMINATION (PROBABILITY THAT SIGNATURES OF DISTINCT TARGETS WILL BE ACCEPTED AS A MATCH) RANDOM CHOICE INSIDE AREA PC1 = (1 - EXP (-Z))/Z CHOICE OF OBSERVATION NEAREST PROJECTED POSITION Pc2 = (1 + 2Zo)-1 WHERE Z T Q a 2 PPsD 14 GENERAL SENSOR TYPES TYPE EXAMPLES SIGNAL DETECTED ACTIVE RADAR E.-M, (HF TO GHz) LASAR RADAR E-M, (IR - VISUAL) SONAR ACOUSTIC PASSIVE OPTICAL E.M. (IR - VISUAL) INVOLUNTARY IR DETECTORS IR HYDROPHONES ACOUSTIC PASSIVE ELINT E.M. (HF TO GHz) VOLUNTARY COMINT E.M. (PRIMARILY HF) 15 BASIC CLASSIFICATION DATA Si 6 ..- ,,7_~,. S2 SIGNATURE DATA 16 CLASSIFICATION PRINCIPLES BASIC DATA ANALYSIS PROBLEMS IMAGE COMPARISON WITH - ASPECT DEPENDENCE OF IMAGES IMAGE FILE - LOW PROCESSING RATE (NORMALLY - HIGH DATA TRANSMISSION RATE MANUAL) SIGNATURE COMPARISON WITH - AMBIGUITY DUE TO OVERLAP IN FINGERPRINT - SIGNATURE SPACE FILE - TIME DEPENDENCE OF SIGNATURE DATA 17 POSITION DATA EXAMPLES FIX S LINE OF BEARING . -.-- . TIME DIFFERENCE S2 // / \C /// / S = SENSOR LOCATION 18 FORMS OF POSITION AND VELOCITY INFORMATION POSITION PROCESSING CONSIDERATIONS FIX + ELLIPSE KALMAN FILTER UPDATES OF STATE VECTOR LOB + ACCURACY EXTENDED KALMAN-FILTER REQURIED FOR UPDATES HYPERBOLIC LOP EXTENDED KALMAN FILTER REQUIRED FOR (TIME DIFFERENCE UPDATES + ACCURACY) REQUIRES MULTISITE ASSOCIATION VELOCITY PROCESSING DOPPLER SHIFT + RADIAL VELOCITY INCORPORATED IN ACCURACY FILTERS MEASUREMENT UPDATE DOPPLER DIFFERENCE REQUIRES MULTISITE ASSOCIATTON + ACCURACY TRACK SEGMENT FROM SENSORS WHICH CREATE A TRACK AS PART OF THE SIGNAL DETECTION PROCESS 19 EXAMPLES OF SENSOR CHARACTERISTICS BASIC LOCATION SENSOR TYPE DATA CLASSIFICATION DATA RADAR A FIX SIGNATURE E.G., CROSS SECT I ON HF RADAR A FIX + DOPPLER CROSS SECTION IMAGING RADAR A FIX IMAGE (SAR CR ISAR) PHOTO OR VISUAL PI FIX IMAGE ACOUSTIC ARRAYS PI LOB (AT,A D) SIGNATURE (ACOUSTIC FREQUENCIES) ELINT PV LOB (AT) SIGNATURE (EMISSION PARAMETERS) HFDF PV LOB SIGNATURE OR MESSAGE CONTENT KEY: A = ACTIVE LOB = LINE OF BEARING PI = PASSIVE INVOLUNTARY AT = TIME DIFFERENCE PV = PASSIVE VOLUNTARY A D = DOPPLER DIFFERENCE 20 FACTORS INFLUENCING TIMELINESS SIGNAL PROPAGATION DELAYS: SIGNIFICANT FOR ACOUSTIC SYSTEMS (C 2900K) SIGNAL INTEGRATION TIMES: INTERVAL IS DETERMINED BY REQUIRED FREQUENCY RESOLUTION * MULTISITE ASSOCIATION PROCESSING: FOR SOURCES WHOSE BASIC MEASUREMENT IS NOT A FIX' * CLASSIFICATION PROCESSING: IMAGE INTERPRETATION OR FINGERPRINT MATCHING * TRANSMISSION DELAYS 21 SURVEILLANCE NETWORKS SENSOR/SOURCE SEGMENT RECEPTION ALTERNATE PATHS DETECTION USER AND INTEGRATION , ' _ = RECEIVE ASSOCIATION FOR FIX DATA CLASSIFICATION _UPDATE PROCESSING PLOT PREPROCESSING fMSI DECISION PROCESSING PROCESSING CENTER 22 A SOURCE NETWORK S S P @ D&I . D&I D&I D&I Class Class Class Class EVALUATION CENTER POSITION FIX PROCESSING ! \ I CLASSIFICATION\ PROCESSING ! - DISSEMINATION TO REGIONAL CENTERS AND USERS KEY: S = SENSOR D&I = DETECTION AND INTEGRATION CLASS = CLASSIFICATION PROCESSING 23 CLASSIFICATION PROCESSING OPTIONS CLASSIFICATION PROCESSING CAN TAKE PLACE: IMMEDIATELY AFTER SIGNAL INTEGRATION AT A SOURCE CENTER IN CONJUNCTION WITH FIX DETERMINATION AT A REGIONAL PROCESSING CENTER CONSIDERATIONS: EARLY CLASSIFICATION - REDUCES DATA LOAD DOWNSTREAM FROM THE SOURCE - CAN REDUCE VULNERABILITY IF THE USER CAN PROCESS THE BASIC POSITION DATA - INCREASES THE FEEDBACK DATA LOAD BY MULTIPLYING FINGERPRINT DATA BASES CENTRAL CLASSIFICATION - CAN REQUIRE TRANSMISSION OF ADDITIONAL SIGNATURE DATA FROM THE SOURCE - INCREASFS THE LIKELIHOOD AND LEVEL OF CLASSIFICATION THROUGH MULTI-SOURCE DATA ASSOCIATION - ENSURES A CONSISTENT REGIONAL PICTURE 24 SIGNATURE PROCESSING TWO BASIC USES: * COMPARISON OF REPORT AND TRACK SIGNATURES TO ESTABLISH A TRACK DESIGNATOR CLASSIFICATION OF TARGETS VIA A MATCH WITH FINGERPRINT FILE DATA REQUIREMENTS: * FEEDBACK FROM REGIONAL AND NATIONAL CENTERS TO MAINTAIN CURRENT FINGERPRINT FILES. (UPDATES INITIATED FOLLOWING EACH ASSOCIATION DECISION.) DETERMINATION OF CONFIDENCE OF CLASSIFICATION IS CRITICALLY DEPENDENT ON COMPLETENESS OF FINGERPRINT DATA. DATA ON GENERAL AREAS OF CONTAINMENT FOR TARGETS (POPULATION ANALYSIS) CAN BE USED TO ENHANCE THE ASSOCIATION ANALYSIS. TIME DEPENDENCE OF PARAMETERS, ARISING FROM CHANGES IN ASPECT OR ALTERED SOURCE CHARACTERISTICS, MUST BE MODELED. 25 DYNAMIC SIGNATURES ABILITY TO TRACK (ASSOCIATE SUCCESSIVE TARGET OBSERVATIONS) ABILITY TO MAINTAIN CLASSIFICATION FOLLOWING INITIAL ID GENERALIZED SIGNATURE = EMISSIONS SIGNATURE + GEOGRAPHIC CONTAINMENT 26 MSI PRINCJPLES POSITION COMPARISON AT SIMULTANEOUS VIEW TRACK MATCHING (A MORE COMPLEX POSITION COMPARISON) CHARACTERISTIC CONSISTENCY TESTS (E,G., ONLY CLASS A HAS RADARS OF TYPE B) OPERATIONAL CHARACTERISTICS (E,G, SHIPS NORMALLY COMMUNICATE WITH A TRAFFIC CONTROL NEAR THIS LOCATION) 27 DATA FLOW IN A REGIONAL PROCESSING CENTER CONTACT REPORTS PREPROCESSING HULTEC ANALYSIS CANDIDATE SELECTION EVALUATION I MOTIONPROCESSING I e i MODELS DECISION | TRACK UPDATE PROCESSING TRACK REVIEW FUSION OF PREPOSITIONED INTELLIGENCE 28 SELECTED MSI PROBLEMS * MOTION MODELS: VOLITIONAL DYNAMICS * AMBIGUITY: RESOLUTION AND REPRESENTATION * CLASSIFICATION CHAINS * ROBUSTNESS: ALGORITHMS VS. BIAS 29 VOLITIONAL DYNAMICS ANALOG OF NEWTON'S FIRST LAW "TARGETS MOVE AT CONSTANT SPEED ALONG STRAIGHT (RHUMB) LINES, EXCEPT WHEN THEIR COMMANDERS DECREE OTHERWISE." MANEUVERS: - INVALIDATE STATISTICAL DISTRIBUTION ASSUMPTIONS - NECESSITATE EMPIRICAL TESTING OF MOTION MODELS TIME SCALES: TM = TIME TO EXECUTE A TYPICAL MANEUVER A r= MEAN INTERVAL BETWEEN MANEUVERS FILTER CHOICE IS BASED ON RELATION BETWEEN REVISIT INVERVAL AND THESE TIMES. 30 AMBIGUITY RESOLUTION IS DESIRABLE, BUT THE IMPROVEMENTS WHICH CAN BE ACHIEVED WITHOUT NEW DATA ARE SEVERELY LIMITED REPRESENTATION IN A FORM USEFUL TO OPERATIONAL COMMANDS REMAINS AN OPEN PROBLEM- MULTIPLE HYPOTHESIS SYSTEMS CAN CLARIFY THE REPRESENTATION OF AMBIGUITY, BUT FACTORIAL GROWTH OF STORAGE AND PROCESSING REQUIREMENTS IS A PROBLEM 31 A CLASSIFICATION CHAIN _ -F t - RAIDA Associated by simultaneous consistency (unique ID) 32!nI ! ~ _ o Al Ascae ysmlaeu soitdb / prilcasfcto)takmthn 32 SELECTED DEVELOPMENT PROBLEMS MULTIPLE DATA PATHWAYS: THEY SHOULD EXIST BUT NOT NORMALLY BE USED. TRACKING UNDER ADVERSITY: SIMULATION TESTS SHOULD BE CONDUCTED IN A RUTHLESS MANNER. AUTOMATION: AN INTERACTIVE INTERFACE IS AN ESSENTIAL DEVEI OPMENT TOOL, EVEN FOR "HANDS- OFF" SYSTEMS, SOPHISTICATED DECISION ALGORITHMS: THESE MAY REPRESENT A TACIT ADMISSION OF FAILURE IN OVERALL SYSTEM DESIGN. 33 34 MEMORY DETECTION MODELS FOR PHASE-RANDOM OCEAN ACOUSTIC FLUCTUATIONS HariZaos N. Psaraftis Anastassios N. Perakis Massachusetts Institute of Technology Cambridge, Alass. 02Z39 Peter N. MikhaZevsky Naval Underwater Systems Center New London, CT. 06320 (Work on this paper has been supported by the Office of Naval Research (Code 431) under contract N000Z4-79-C-0238) 35 Q1981 IEEE. Reprinted, with permission, from ICC '81, International Conference on Communications, June 14-18, 1981/Denver, CO. MEMORY DETECTION MODELS FOR PHASE-RANDOM OCEAN ACOUSTIC FLUCTUATIONS Harilaos N. Psaraftis Anastassios N. Perakis Peter N. Mikhalevsky 1 Massachusetts Institute of Technology, Cambridge, Mass. 02139 2 Naval Underwater Systems Center, New London, CT. 06320 AB3STRACT described in the following. The basic problem of the ocean acoustic detec- 2. THE CONTINUOUS-TIME DETECTION MODEL tion process is formulated analytically under the assumption of fully developed saturated phase In this paper, detection at time t is defined random multipath acoustic fluctuations. Detection is defined as occuring whenever p, the root mean through a specified threshold level p, as follows: square pressure at the receiver, exceeds a speci- through a specified threshold level p fied threshold level p0 A continuous time model p(t) = P (t) 3 (1) is first developed for obtaining the probability density functions of the time between two success- The unconditional probability of detection, ive detections and of the time p is above n (hold-- denoted by A dt is b-denoted by 1 dt, is by definition the probability ing time). This model is then compared wi the that (1) is satisfied at some instant of time in extensively used (A, a) model and with data. The the interval (t, t+dt) where t is random and dt model is seen to exhibit similar long-term behavior is small. but markedly different short-term characteristics It is straightforward to show [3, 8, 143 that as compared to the (X, a) model, a fact which is if a 2 is one half the long time average mean due to the memory of the process. A comparison 1 of these models with data obtained from various square pressure and is the root mean squaresingle path phase rate (in rad/sec), then X (in field experiments demonstrates, in most cases, a secl) is given by the following formula: 1 significantly improved prediction capability over the (A,a) model. Subsequently a two-state anda four-state discre- P, p 2 2 (2) te-time Markov models are derived and closed-form X ex(-p expressions for the probability mass functions of 1 the corresponding interarrival and holding times are developed. The results obtained using the Eq. (2) holds under the assumptions of the phase latter models are favorably compared with both random multipath model [4], in which p has a the continuous-time model and the data, the great- Rayleigh PDF, p a Gaussian PDF, and p,p are est improvement lying in the much lower computa-ndeendent tional effort involved. We call A the "unconditional detection rate" or the "per unit time unconditional probability of 1. INTRODUCTION detection"; X is the average number of detections (or "arrivals' )per unit time.Much of the work in the area of acoustic detec- The conditional probability of detection, deno- tion in the ocean has traditionally been based on ted by 4(t)dt, is defined as the probability that the so-called (A, a) model characterized by the a detection occurs at some instant of time in the "relaxation time" 1/X and the standard deviation o interval (t, t+dt), given a detection occurred of the signal-to-noise ratio" [1]. Use of this at time 0. model has relied heavily on parameter estimation It is again straightforward to show [14] that from field experiments without application of the if f . . (P ,P2,l, 2 ) is the joint PDF of relevant physical and probabilistic structures of P1P2 1P2 2 the process. p 1,p2 , where subscripts 1 and 2 refer to This situation has been improved recently by times 0 an~ t respectively, then d(t) (in sec- 1) is the efforts of several authors ([2] to [8]). Under given by the following formula: the assumption of a fully developed saturated multi- path phase-random field, probability distributions 1 for several random variables such as p, the root (t) A P f pPlP2 2 mean square pressure at the receiver, p = dp/dt, . 0 1 2 the phase and $ = dW/dt, as well as many joint (3) probability distributions have been derived. This new knowledge has enabled us to develop new conti- nuous and discrete-time detection models, which are J6~~~~~~~~~~~~~~~~~~~~~~ We call ?(t) the "conditional detection rate", refinements of this model were based on al approxi- or the "per unit time conditional probability of mation to the definition of detection using a modi- detection". If the detection process had no memory, fied but equivalent detection criterion that produ- 4(t) would be equal to A for all t. Evaluation ced very accurate results with significanitly reduc- of (3) for the phase random process has however ed computational effort [14]. shown that this is not the case. Results showed We now proceed to briefly describe the current- that 4(t) - 0 as t + 0 and 4(t) = Al for t > t ly used (X, a) model and put it in a form compati- to is the decorrelation time of the process and ble to our model: was observed to be in the neighborhood of 3/2vv The basic assumption of the (A, a) model is to 4/2rv . For t between 0 and t0, the behavior that "detection opportunities" are generated in of 4(t) was observed to be either monotonically time according to a Poisson process of parameter increasing up to t , or to have a peak higher than X [1]. The reciprocal of X is known as the "relax- A at some intermediate value of t. The existence ation time" of the process and its value is usually ol the peak was seen to depend on the selected taken arbitrarily from empirical considerations of value of a1, v and p . A conditional probability the process, and without any explicit relationship similar to 4(t) is q/t), defined as the probability to the detection threshold level. At any particu- that a downcrossing through p occurs at some lar detection opportunity, a detection occurs if instant of time within (t, t+St), given that an the level A in dB ( = llog p2 ), which is upcrossing occured at t=0O. The formula giving assumed to be normally distributed, exceeds a +(t) is similar to (3), the following: specified threshold level Ao. It is argued in [14] that a common basis of t (r comparison between the (X, a) model and our conti- 1 0 J 1 2 P1P2 (4) unconditional probabilities for both models. It0 - P (4) turns out that in order to satisfy this require- ment, the values of X can be no longer be takenWe now proceed to evaluate the PDF's of the inter- arbitrarily, but according to the formula: arrival and holding times: We use the term "interarrival time" to denote A v (7) the time between two successive detections. The exact evaluation of the PDF of the interarrival 1 2 time seems to be very difficult. Rice [11], Longuet-Higgins [12] and McFadden [13] have presen- The above value of X is called "equivalent A". ted several approaches to this problem in the It is the value that X should take in the (A, a) general context of axis-crossing of random funct- model so that this model has the same average ions. We propose here the technique used success- number of detections per unit time with our new fully by Psaraftis [9], which states that a good detection models. Hence, all comparisons of the way to approximate the above PDF is by the function: (A, a) model with the continuous-time model impli- t citly assume that A takes the above value. If ,(t) exp(-f0 4(x)dx) t>O this is the case, the (X, a) model has, in terms F(t) = { (5) of our previously defined terminology, the follow- 0 t<0 ing properties: We now define "holding time" to be the interval a) 4(t) = 4(t) = const = A1 (no memory) between any upcrossing through the threshold p b) 'PDF for interarrival time: Negative exponen- and the first downcrossing through p0 that follows. tial given by: f (t) = tl(exp(- By analogy, the PDF of the holding time can be t 1c) PDF for holding time: 1/2-order Erlang (or approximated as follows: Gamma) given by: {f(t) exp(-f p(x)dx) t> 1 0 t<0,we.r.e, - ~ i obtaned b 4We will now see how the continuous-time and where b(t) is obtained by (4). The determination of the interarrival and hold- the (A, a) models compare with the data that was ing time PDF's exhibits a non-trivial complexity. made available to us. The reason is that the joint PDF inside the inte- From the analysis of the CASE experiment [7, Thegrals ista the jointP ) is itself difficult 16] which was done at ranges varying from 200 to2gras f 2 2is itself difficult 400 km and frequencies of 15 and 33 Hz, we choseto evaluate. In [4] & t11] it is shown that p and to present Record 21. Although a record of low [ 1 signal-to-noise ratio, this was chosen for beingP 2 are linked in a rather complex fashion, involv-ing Bessel functions. Moreover, the correlation one of the few records examined that satisfy the of p with f and of pM e r is not a well-esta- phase-randomness assumption, an assumption on whichof P with .2 and of, p2 with l1 is not a well-esta- our detection models are based. In Figs. 1 (Inter-blisned function. bli d ,fun.ction. s arrival Times) and 2 (Holding Times), a thresholdThe exact evaluation of 4(t) and p(t) involves of = 175 = 7 volts a 2 = 15.95 volts2 and the execution of a total of 3 nested numerical 0 1 1v = .1734 Hz were used. In both cases, the (A, a) integrations. This was the reason why the computer model, unlike our model, fails to match the implementation of the model originally presented observed experimental histogram for small values significant computational difficulties. Subsequent ob served for small valuesof t (t3O). On the other hand, both models predict ~~~~~~7 ~ ~ s s, equally well the behavior for times greater than P prob(p>p at time TIpp0 at time 0) = a Then Pi(n)= b (-a) al-b k-l)/o, orUD p p 0 1\nI kl (1-a) aZ- /Z, or ln-1PUU = prob(pBp0 at time TIpp 0o at time 0) = 1-a PI(n) = ab Z (1-a)k-l( 1-b)n-k - 1 (13) k=l 38 0.16 I I I I Pl P22 P33 P44 O0.14 Continuous Time Model _ O.14 J.2( )123 34 0 o.12 - 3 10.12 Markov Model 0.l0 , P21 - 3 2 P4 3 Fig. 5: Four-state Markov Model 0.08 -v 0.08 the PMF's of the interarrival and holding times P.,n14 - because now more involved, because "up" states are now states 3 and 4 and "down" states are states v0 ~ c 1 and 2. X .4 Defining PH(k) and PI(n) as in the two-state 0.02 model, and also V0 5 10 15 20 25 31) PN(n-k) = prob (non-holding time = (n-k)T)0 5 10 15 20 25 30 N t (secon-s) prob (upcrossing between n and n+ll downcrossing between k and k+l)Fig. 4: Interarrival Time, 2-State Markov Model then it can be seen that Figure 4 presents a comparison of the interarrival p P(k P-k) (14) time PMF of the two-state Markov model with the P(n) H N nk=0 interarrival time PDF of the continuous-time model for Record 21 of the CASE data.(The data histogram These PMF's can be evaluated by working in the z- for this record is shown in Fig. 1). The agreement It of this very simple Markov Model with the data as can be shown that the holding time PMF is given well as with the continuous time predictions is quite satisfactory. The only parameter that has by the formula to be calibrated by the user is the time step T, P (n) P [AnBnp (An-l n-l-B )1/(A-B) (15) which can be seen in Fig. 4 to be T = .6 sec. The H 32 44 results of a cursory sensitivity analysis showed that the form of the PMF predicted by the Markov and that the non-holding time PMF by the formula model does not change appreciably over reasonable (less than one order of magnitude) changes in T. PN(n) = P23 [Cn-D - (Cn-Dn-)]/(C-D) (16) The holding time prediction of the two-state model sometimes compares with the data even better than where A,B,C and D are constants given in terms of that of the continuous-time model. Still, its the transition probabilities as follows: form does not fully satisfy our intuition, since 2 it has its most probable value in the first time A = /2[P 3 3 +P P 33 -P44 ) + 4P 4 3 P34] increment (similarly to the (X, a) model having 2 its mode at t=O). This somewhat counter-intuitive B = l/2[P3 3 +P44-/(P 33-P44) + 4P4 3 P3 4 result, together with the apparent simplicity of /2 the two-state model in its description of the dyna- C = 1/2[Pll+P2 2 +P 1 1- 2 2 + 12P21 mics of the underlying process, has motivated us to D 2 formulate a more refined Markov model, discussed D = 1/2[Pll+P22-V(Pll-P22) + 4P12P21 ] below: 3.2. Four-State Markov Model Hence P (n) can be evaluated by substituting (15) and (163 into (14). Introducing two additional states in the two- Figure 6 is similar to Figure 4, with the ' state model, we obtain the model of Figure 5, where difference that here the continuous-time modeiJis we have: compared with the four-state Markov model. (same CASE 21 Record as in Figs. 1 and 4). The resultsState 1: 0 P < Pt 0 of the four-state model are only slightly superior State 2: PO , p < p0 to the ones of the two-state model. Here p' aad State 3: <" ', as well as T, have to be selected by the user. SP0 t : 0O In Figure 6, T = 6 sec, p 3.5, P 14, with State 4: P0 6 P < P0 = 7. volts, 01 = 15.95 volts2 and v = .1734Hz- same as in the two-state discrete-time and the Here pad p " are artifical thresholds, chosen by continuous-state model comparisons. A brief sensi-. the user. As before, the user also selects the tivity analysis has shown that better results are time increment T. As it can be seen from Fig. 4 obtained for levels of p', po" not close to p the system is a birth-death process. (namely,p'

:m 0 L. ~ *' (~0 S_ 0 o ,3- 0 . 4i) v r -.. I: S- ' (A2j to 0 a) _ c c C o U i-' , -.C 4 c cu *- o-) -v 4- .- ) cm 4J -- ?- ' z ' LLJ C- v L. CDa6 4 0 0 2_ - C 4V 4. V) In 4.n I C C) 0) C 0) 0) 0) 0 r: 4 g tot 0 0 In o S. . r J .J -? 0 -) ) * , : *. 4_n *t I In ^ 4.)E S .= 0 x c to 4 cS. 0 0 2I to 0 o- * to 4I 0 0 _ = * a c0 r =- V 69 Thus, MARCY's two running modes have complementary capabilities: the user runs MARCY in batch mode whenever necessary to initialize or re- initialize the estimated tracking scenario; and he runs her in recur- sive mode whenever possible to update track estimates handed off to her by batch mode. 2.2 MARCY's Inputs and Outputs MARCY's inputs and outputs depend, of course, on the particular multi-target/multi-sensor tracking environment to which she is applied. We here describe the algorithm's inputs and outputs for the case of the ocean surveillance environment. 2.2.1 MARCY's Inputs In An Ocean Surveillance Tracking Environment As figure 1 shows, MARCY's inputs are of two types: o values of control parameters; o target-related data. MARCY obtains values for her control parameters either directly from the user (there is an extensive user/MARCY input dialogue for specifying parameter values) or from a previously created control parameter data file on disk. Table 1 shows the most important control parameters for which MARCY needs values. After MARCY has a complete set of parameter values, the user may instruct her to print ,these values out for inspection. An example of such a printout is given in section 4. MARCY then offers the user an opportunity to change values of one or more parameters (without, of course, requiring that he respecify them all). The user may, at his option, save parameter values on a disk data file for his later use. MARCY reads target-related data from a previously prepared binary disk data file. MARCY is, in principal, applicable in any multi-target/multi-sensor environment and thus can, in principal, accept any kind of target- related data. For each data type, MARCY must be supplied with the appropriate "measurement equation"--that is, the equation that relates that type of data to the target's position and motion. 70 CONTROL PARAMETERS TARGET LOCAL. MANUAL DATA ASSOCIATION/CLUSTERING MANUALRELATED DATA FUSION/CORRELATION MODIFICATION DATA I . CAPABILITY TRACKS OUTPUT OUTPUT BATCH//0* MANUALIMODE I MODIFICATION DATA ASSOCIATION/CLUSTERING ,:SOFTWARE fl I CAPABILITYDATA FUSION/CORRELATION CONTROL PARAMETERS TARGET 4 l > LOCAL MANUAL TAR ELAT ED I ~ DATA ASSOCIATION/CLUSTERING MODIFICATIONRELATED DATA FUSION/CORRELATION C IIT-f DATA CAPABILITY I I f. TRACKS' OUTPUT OUTPUT I { r MANUAL MODIFICATION GLOBALDATA ASSOCIATION/CLUSTERINGRECURSIVE I CAPABILITY DATA FUSION/CORRELATION MODE SOFTWARE Figure 3. A High-Level View Of MARCY's Functions And The Flow Of Control 71 The current MARCY, however, has been developed with the ocean sur- veillance application foremost in mind. Thus, the current MARCY's data has most often been that of the ocean surveillance community, viz: o positional information - radar measurements; - pairwise coherence measurements; - active sonar reports; - pilot sightings; o bearing-only information - SOSUS lines, of bearing; - HF/DF lines of bearing; - passive sonar reports. Indeed, much of the early work with MARCY has been with a single data type--that of pairwise coherence measurements. (Table 4, which appears in section 4 of this paper, shows an example of a dataset consisting of 48 pairwise coherence measurements, each line on the table being one such datapoint.) 2.2.2 MARCY's Outputs As figure 3 shows, a user of MARCY has two major opportunities for output: one after MARCY's Kalman filter has solved the data association/clustering problem in a "local" manner, via data fusion/ correlation; and the other after MARCY's integer program has solved the data association/clustering problem in a "global" manner. At either point, the user may request any of eleven types of information, of which seven are tabular in character and four are graphic in character, about all or a selected subset of the clusters or tracks found by the algorithm. Table 2 lists these types of output information. Examples of some of these types of output are given in section 4. After viewing any type of output the user may recycle through the output dialogue and make other output selections. 2.3 Versions of MARCY The algorithm MARCY that we describe in this paper is applicable in a general multi-target/multi-sensor environment. As such, MARCY may be seen as the progenitor of a number of existing versions of the algorithm each one tailored for use in some one special practical situation. 72 Table 2. MARCY's Output Types Type Number Information Content 1 Water times, peak numbers 2 Water times, peak numbers, probability scores 3 Water times, peak numbers, probability scores, chi-square scores . 4 Water times, peak numbers, probability scores, chi-square scores, measurement residuals 5 Water times, peak numbers, state vectors 6 Water times, peak numbers, state vectors, covariance matrices 7 Water times, peak numbers, position coordinates, courses and speeds 8 Plot of feasible tracks in geographic space 9 Plot of feasible tracks in measurement space 10 Plot of residuals versus time 11 Plot of covariance matrices versus time 73 Most of these special versions of MARCY have been created to support ocean surveillance experiments by the Advanced Research Project Agency (ARPA) of the Department of Defense. The ARPA experiments in which a version of MARCY has provided tracking support are: Date Experiment Name Fall, 1978 Semi-Alerted Search Experiment (SASE) Spring, 1980 Broad Area Search Experiment I (BASE-I) Spring, 1981 Broad Area Search Experiment II (BASE-II) Summer, 1981 Pathfinder All of the versions of MARCY that have been created to support these experiments are equipped only with batch mode (i.e., none of them are equipped with recursive mode), and all operate on real data in real time. 3. HOW MARCY WORKS MARCY works by solving the data association/clustering problem in two stages: a "local" stage and a "global" stage. MARCY first solves the data association/clustering problem in a "local" manner, by which we mean that MARCY: o considers (at least implicitly but not necessarily explicitly-- see below) all possible partitions of the dataset into can- didate clusters; o scores, via data fusion/correlation (which is carried out by a Kalman filter applied to the data in each candidate data cluster), each data cluster for the degree of agreement between: a) what the data in the cluster would imply about the behaviour of a target under the assumption that the datapoints in the cluster are all due to that target; and b) a simple model of target motion, namely, one of constant course and speed.* o rejects candidate data cluster whose scores are sufficiently poor--along with (implicitly) all other candidate data clusters that could be formed that contain such poorly scoring clusters as subsets.*#!*# * A user of MARCY can arrange for the algorithm to accomodate maneuvering targets as well. He does so by supplying parameter values that specify a high uncertainity in the constant course and speed model (i.e., he specifies a large amount of "system noise"). *#!*# This idea, which as a moment's reflection will show is entirely rigorous, is what keeps MARCY computationally feasible. This idea is due to Charles L. Morefield of VERAC Corporation. 74 Figure 4 illustrates this process. The numbered circles in the figure represent individual datapoints (of any type). The columns of circles in the figure represent copies of the entire dataset. The figure shows MARCY building candidate clusters in a bottom-up fashion -- trying first all possible one-point data clusters, all possible two-point data clusters, all possible three point clusters, etc., rejecting (as indi- cated in the figure by solid vertical bars) all such candidate clusters that score poorly (along with, implicity, all other clusters containing the failing cluster as a subset). The surviving candidate clusters are then fed to the global data association/clustering process described below. Note that it is possible, and indeed often happens, that, because data clusters are scored in a "local" manner (i.e., independently of each other), surviving data clusters can have datapoints in common. Such cases of surviving data clusters that overlap are bothersome because one then has instances of a datapoint being associated with more than one target! Thus, it is necessary to consider the surviving data clusters as a whole -- that is, in a global manner. MARCY next addresses the data association/clustering problem in a "global" manner, by which we mean that MARCY: o notes cases of local solutions to the data association/clustering problem and the data fusion/correlation problem in which surviving candidate data clusters have data- points in common; o selects, via a 0-1 integer program, the "best" subset of non- overlapping surviving data clusters. Here, by "best" we mean that subset of non-overlapping data clusters the sum of whose local scores is better than the sum of scores of any other subset of non-overlapping surviving data clusters.* Figure 5 illustrates this global data association/clustering process. The blobs on the left-hand side of the dotted line in the figure repre- sent clusters of data that survive the local data association/ clustering and data fusion/correlation process, and the numbers shown in those blobs represent their scores. Since there are cases of overlapping blobs, we must select a subset of them that do not overlap. We (more generally an integer program) find that the subset of blobs shown on the right-hand side of the figure is the "best"** such selec- tion of non-overlapping blobs. * This idea of a globally best subset of non-overlapping data clusters is also due to Charles L. Morefield (1). ** "Best" here means that the sum of the scores of the (non-overlapping) clusters is as large as possible. 75 p- S- g- L/- l- C C0-,o~-, E F- rO -I C: r- L-) LL.-4 ~S-t- 0 0 U, LI. I p-- 4-' 0-~u- 0 (-) Fr CM C..)cal ~ ~ c 0- a- -JOLLJ Q) -0 0C)~~L) S.-4-, [~ _ F- C 0 7.,00I- Cd,O c-e~~~ca,~~~a z ~~~I- S.. E= 0-~~? -I-~ 76 (-'U--~~~~~~~~~~~~~C ,r'- -o 0 '.--l~~~~~~~~~~ (I) a.,f ~~~~~~~~~o I ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~e tt v~~~or-.~~~~~~~~~r C 0, ~/a t-f \ o CU /~ Ci o~~~v, a~~~~~~U o\t1 i J , 7 w ., L-\.S ~~~~~~~Q) i~~ i. \ if~~~~~~~iJ \ ~~~~O ; l 0~~r-~~~~~~~.~V 77~~~~~~~~~v 4. AN EXAMPLE Consider the hypothetical two-target/three-sensor ocean surveillance scenario shown in figure 6. Two targets move on constant courses at constant speeds: target #1 on course 1700 at speed 20 knots, and target #2 on course 1900 at speed 30 knots. These data, along with their initial positions at hour 0 as indicated, will result in the targets passing through one another at hour 10.* The three sensors in the figure are SOSUS stations. They are fixed in space** and are arranged in a triangular configuration as shown. Table 3 describes this scenario in numerical terms. Table 3 also indicates the levels of "process noise" and "measurement noise" that will be injected into the scenario by a data generator in computing artificial target data based on the scenario in figure 6. Table 4 shows a collection of 48 artificial pairwise coherence measure- ments that were computed from the data in table 3. Thus, the data in table 4 are synthetic data that represent the surveillance scenario shown in figure 6. This data comes to MARCY without any indication of the number or characteristics of the targets which produced them (the target numbers shown in the second column of table 4 are, of course, stripped away before the data is given to MARCY). MARCY's job is to unravel the data in table 4 so as to recover, as closely as possible, the scenario shown in figure 6. Figure 7 shows the dialogue that a user of MARCY had with the algorithm in running MARCY on the data in table 4. Our description below of this dialogue is keyed to the figure by symbols of the form A , B , etc. The rectangles on the dialogue show entries made by the user. At point A , the user is logging on to his account on the computer system. At point B , he sets the width of the terminal to 132 charac- ters, thereby allowing for outputs with wide formats. At point C , MARCY is invoked by hitting "MARCY". The algorithm is then loaded into core, and execution begins. MARCY first announces herself and then prints the running notes (these are overall reminders to the user) as shown. At point D , MARCY determines whether or not the terminal in use is a Tektronix terminal. This information is necessary in order to be able to automatically clear the display, and for graphic output purposes. * We chose this case so as to make it harder for MARCY to assign data- points to targets than would otherwise be the case. ** MARCY can handle situations involving mobile SOSUS stations. 78 INITIAL POSITION \ INITIAL POSITION FOR FOR TGT #1 TGT #2 TGTS CROSS AT TIME 10 SENSOR #3 __,,SENSOR #2 TGT #1: COURSE 170, SPEED 20 KTS TGT #2: COURSE 190, SPEED 30 KTS Figure 6. Hypothetical Ocean Surveillance Scenario. 79 eD G(o(sg I L W iv ( o ? UL 04 4lCd I * C. hi I : .C Co CO U M W C M c e 0 w W M 0 C -: .C% LO 4A M C) M M %jo M b W LO - -v C w vI M w v - v -M w --* * M- * * v v v , M , w M M' w M L) O) CI . -C CZI _j Lu___ ru- c c M U x o a, C4 . -- - -4 -4 .4 . - --- 4J .m M4 w w M M M M M v V V- V))C X ( CocunuO} C' aUC - atr, tO'W* -rran-4 0L,)at aalin?.traaf rar, awc-n e o .- 1 -U. -t.4.4 CC a - 2 WEL r z a -a it L o? a a _t Li tc a- )C a ~ I ww ?C~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~. At point E , MARCY asks the user to specify the (spherical) coordinate system with which he desires to deal with the tracker. The user, as he is working in the North Pacific, naturally selects north latitude and west longitude as being positive. All coordinate inputs and outputs are then regarded by the user as being in this coordinate system. At point F , MARCY asks the user for the speed of sound in knots. At point G , MARCY asks for data concerning surveillance stations. The user, wishing to run the artificial data case with three immobile sur- veillance stations as illustrated in Figure 6, enters the station data as shown. At point H , MARCY asks the user to supply the names of two scratch disk datafiles that will be used to store parameter values for batch and recursive modes. The user responds as shown. The user has now finished entering the "universal parameters" (so called because they apply to both batch and recursive mode). MARCY prints these parameters out starting at point I , and asks the user to verify them. The user indicates that they are satisfactory. (If the user had made an error in entering one or more parameters or had he for any reason been dissatisfied with the values of the universal parame- ters, he could have indicated this fact to MARCY. MARCY would then have recycled through the relevant part of the input dialogue, thus giving the user the opportunity to change these values). At point J , MARCY asks the user to indicate whether to run in batch mode or recursive mode (or to stop). The user chooses batch mode, as shown. At point K , MARCY advises the user that it has begun the process of constructing tracks in batch mode. At point L , MARCY asks the user whether he wishes to provide batch-mode parameters by restoring them from an old parameter datafile or by entering them manually. The user, not having already created a suitable parameter datafile, indicates that he will enter them manually. At point M , MARCY asks the user to specify the first parameter -- the name of the target datafile. The user responds with "PKSF5" (i.e., the datafile whose contents are displayed in table 4) as shown. At point N , MARCY asks the user to specify a lower bound and upper bound on the length of the tracks he is seeking. At point 0 , MARCY begins a series of questions aimed at establishing what constraints, if any, the user wishes to place on tracks. 82 Basically, the user has an opportunity to constrain tracks by means of each type of data in the datafile: station numbers, bearings, fre- quency gins, times of receipt, time differences, frequency differen- ces, and coherences. The user also has the opportunity to constrain a track's components by datapoint numbers. The first question concerns whether or not the user wishes to constrain a track's components by datapoint number. The user, not wishing to limit the tracker's atten- tion to any particular datapoints in the datafile, answers "no". MARCY then goes on to ask the user whether he wishes to constrain tracks in other ways as shown. The user responds "no" to all of these questions except for the one involving the "times of receipt" (TOR) of datapoints where, as he wishes to limit the tracker's attention to the datapoints coming from the first five hours of the scenario shown in Figure 6, he does so as shown. At point P , MARCY begins to ask for parameters that will be needed for local data association/clustering and data fusion/correlation (i.e., the Kalman filter). The first item is the measurement noise covariance matrix R. The user responds as shown. Note that MARCY advises the user of the units in which to respond. At point Q , parameters 21 and 22, which are needed to determine the process noise covariance matrix Q, are asked for and received; again, desired units are provided. At point R , the minimum water time and maximum water time between adjacent components of tracks are asked for and received. Point S is concerned with the specification of initial state vectors and covariance matrices for MARCY's Kalman filter. The dialogue here is self explanatory (to those familiar with Kalman filtering); At point T , MARCY asks for and receives the three types of thresholds by means of which local data association/clustering and data fusion/correlation is cut short. Here the availability of default options as shown make it easy for the user to specify standard values. At point U , MARCY advises the user that values for all parameters have been received and asks the user whether or not he wants these parame- ters printed out for his inspection. The user, after indicating that he does, receives the printout as shown. At point V , MARCY asks the user to indicate whether or not the batch- mode parameter values are satisfactory. Upon inspecting the printout of parameter values, the user decides that he would like to change the threshold for stagewise chi-square scores from 9 to 10. He indicates this fact to MARCY. MARCY responds to this by listing, beginning at point W , all the areas in which the user might wish to change parameter values. The user makes appropriate selections and, after MARCY recycles through the relevant part of the input dialogue, makes the desired change at point X 83 At point Y , the user indicates to the tracker that he is now satisfied with the values of the batch-mode parameters. MARCY responds to this by saving these values in the batch-mode parameter datafile. At point Z , MARCY asks the user whether or not he wants to look at the data in the specified peak datafile. The user responds "no" as shown. Had he responded "yes" instead, the user would find himself looking at table 4 (less the second column as mentioned previously). MARCY's output in this case (i.e., for the target data shown in table 4 and the control parameter values shown in figure 7 above point Y ) is shown in the remaining tables and figures: type 4 output is shown in table 5, type 7 output is shown in table 6 and type 8 output (a geographic graphical display) is shown in figure 8. As tables 5 and 6 show, MARCY produced six data clusters (each with 5 or 6 datapoints) in the local data association/clustering process. These six clusters divide into two sets of three clusters each. As figure 8 shows, each of these sets of clusters is traceable to one of the two targets in the hypothetical scenario of figure 6. 84 L Lng ; M L W L,- C- W - to to J t "a t. ? at It---t\)fU I IUC I t tI 'I ? C) C - > >) G CD.e 43,.) 0:3 << 09e9 e0t f. -CA-o ? . C, o C- - M f N (t r? r - q..e , ? * e L , t> It . . ? .. oU t . 'A I CO I I I I I n I to I' I I .S a IaY ~~Pv~ QO~wee eastP e oC0-ta Qass, eaty; ,; '" C.'. '.r 'f ' - . .- 1'- -* L 'S M ,e I.. E -. )- CNs krn X I r- ul r t *~t,_ UwbA M _'A L'aW.. I I I I C- A n( V ru Kco '' LAD rU CD ; O *D I U -W S C) DC> _U Ca* n ,} . . , . . . . . . . . . . . * * . . * C. . * .0 C.) .- MS.... a% ..d OJ Ir~40 r IL.q 4C Li.E r "=2:~ 5tta u?>t (ttJ~- t@7t Uf~Y tW:D Ot > OQm~ ~N~N ~NU~ ~P~ 85+m~m~n s~ LL u .J -J Z- -- 00 o) cn 0U= L - . .LI C.- CAO >L0_ >- C) ZC , " -I ), -4 CL - _ I U) > -O L. C) I ) - . X ,.-! , -- a) 0X O- O U :LL.CD(I-0 t -- _C ? @ ~ '0-- L- Z LL )z =o oI. -I I n 5- Cl = -i.rr U0 - E+ r - :: C-:3 Z ) O )a9) C0 CD a 2: 2 J * f-U h V cna~~~~ e- _j =0 ~- - _ , LD * 0 rI CY ( -m 0 CK C O- U- g. -i S 0 S 0 -a ( LLO 5: O 0 a) C) OC - C I CU L -= = C, S J O -e OL zJ CI LL ,Ln v, L H t4 H H H H H 4 H O H H Ht.4 H t 4 H H H * H .4 H H 4 o H H4 H H H H t4 M H H H H H H H H )f tS H {...~." .. .,,,Y . .~ - t- _MMS HH+4 .w ~- .. - ,,.f KM w > > e _ +M ?~,~ ).e - C ta ~- ? W VI a. :H H MM * *. 7 M Ct w* , "w O . * Itt M#,,. - ' 4- C MM rw oH ffw u ".I ,*4 *4 - MM ' rMC t) 4 CZ W4 ( ,.. n2 Cc MM ' C z ) MM C i MM MMr I- IiiCf MM KM 3 4 Ce 0 KM t f O _ t wL I_ t4 f 4t ffM Y ( * O Z O.J ? MW {^" I MM MM"^f^F* 0. A C MM M t4 " ff ::X-. - FF < o t^J 4 a O *2_ W_ o- a 9 nJ rMM MM t- 4 y J * OIAJ I Li ~~ eu evw aw rl S *n X L1 ' MM~n~~ MM 0 J4 _ - W C) W MM~Y~CY~~ ~ 4 _ . Z I InO -M~MMMMMH O IC Ct_ KMHMM 4 4MMM*4M 4 4HH~S~L)+tl4 _ I- I-' Z a HMn i*MHHMHMMMMM Li *- I t t - t I *- HMMMMHMHHMMMHMcr 44 01* (* In I-'- nu-v . 4MMMMMMHMMHMMM c: D3 0 Li 3 Ot. v- a s- MMMMMHMHHHMMMH~ro S C 14 2: (.3 1 OCX '-"- a- ct 0 (1 i a- .. -c cc - - 3) C) lC (H c s TO 4-a 0 0 r.0 4ir0 44cO 0 cr5 w ix. C. d e- o~~~~~~~~~~~~~-- 4-)' W 0 : wl w= = ,' .ta ta lra ~ WT(o, l oOw ~~~~~W qs) X CA C 0 c-~ ~ ~ ~ Wa c EQ t- - . - < a L' ta4 0 bil la. W!e C) a- aia c ('1U.-4 - L C> ) 0 1 Wi W (C Li> z2r: 4) 3Z7 L S '4 iL_ a- Li.0. C., W )- m .j .H~~~ mN+ st~~a a x la. * -.a w OYC) 5-() () (3 C LL. t- a i Z Z Z 0 A / Z O - LLJU LU - a (A L.) C CI F - CY Z C? < LLU C LUl F- F- F- F- >c C> O >: C ( V) (A L. F- LLU C) L Z I Cl! LL F- C C) CO C D u LL Cl ^ LU ,, Ci -.LI) ( ( ) V ( () F- LLU - 4-u F F- F- C) i F- 0 ! I< - - 0 CD 2 Mr LU 0 - C(O LLJ i F- F- F- F- F- C I V . ' (A v, C < >< E LU Z: o C\J CM) ( V)A c C C) C) C) an uI VA L. Z D C X C =) D LU Z (AK a C C >-> -> > >- U C 0 3 i: LL C-) C _ Z Z Z Z . I I C) LlN-J -tC _ C) ( (F LL L. LL LL LL i LL Z I Z II S- z 4-' LL 44 - > LLJ C !L - -- __ 3 u) C- C- c-) (_) c go ) C) M -: cnLU LU LL LU LU 01 L LL LLU rt- 1 0. CL Q. CL C : QJ 0 C) LL CV) C) C) C) C) CZ 0 0 C) LU.i L F- F- F- F- F- X F- F- < . F Ul 2-i t S S0 F- F- F- F- - F- C C) C F- C) LI C< < :3 3 0 0 C\J C\J 0 3: _-M3J C) C CD Q C)- 0@ C) C) C) C) C) 0 C0z 0 C) C - C < c C >< >- >- >- >- >- U LU LLJ >- LU LU ^ LLS C: L aL c CL=C)i C jCt C LU J D C L LULI LU (A ()A LUJ C) C)89 89 Q0 -) O uA rrt0..0 p4~- is' l -- ,. qt.l - X inisi rci I C 0 C.J-3 cr . - , . 1 ;Ca ' , ,e 1 < 4tz 0. c ci4D 0 Z w - C r - -zL m-4,- < zq4 c'" c c z . , . e-e-a WL::41 0aw .- CZ< 3 > OQ Ob tbb- w lr n~ n W- - -- -- OW OC O . 4.) Li - 4-. 4 O W O .1,? O CW ) O ?4J -o W 4)%J L LJ w Al t_ C.) n C, C) .' C) -J L J - 0. u. 40 o - t kn _J?4, OW.t ,- ?4 C.t W. XW 90 - 0IU b.0 La .. O ZtO1 Q biQJ cZn 0 - L s Z .. 0.J.. J 3 2: 'r .i:= t_- :t C " t'" ii t.c o-. 3O- E (Z: -Z) 0: a :0.) L, c CAW < 3 ; nG U ww CIL : :> ta-0 - I-- Z I-- Z 0 0Z C o3 2 : o-' 2: z o > ,(fl '.4) C oOct CE oI Y 5 CE Li. -. '.__ Z,.-1 ,- . ',.- . - I- X t LO .- (3 o 5 a 2 0 o C ( ('L a Z .- --- ; C)---- Ii 0 C 0) _-4u. 0 M9- 0 L Z, 1: 0 WtA J -) 0 1.3 .J.--4- - -0 2- Lo 0n oo C) *0 * I O O O O " : I- CL (3 ( ( j (3 0 C (A 0 " tu ?) & & W O [ ,JLulLi Li *iU(us3 ' ( * J. W L -O *- ) * E 0 * wE to2 IL H . . V CE CE aE i-Xt-- -- us Z u a m C- ? . QfL C W I - - ) to U to Z .- o' . LI- r C S C C Z > . . ... it l W a CD D :CE C O I W v - a- - - o F - L m C.- C), Ch - - 4 3-. C e w 0 C -- hI (l I(t X(S *) IL CE C . ZX 2 : a * 00 CE CE * C. I- 0 ? ( 4 " CE 4 l 3 0 X-a c 92 4-) 0 CD U) 3,~~~~cd Ce .~V -- X LICe oCe Z--4 U - C - (D( 0 . ta to I I--gC.) X w '" U - -' n: oe o =l (xo C (3Ce ) M- z U) us z C. .4 0.- ) ) '. 0 )IC N .- C : " Cgt C UO b~0- 0 (- 3 0 ct< a. UJat f C- a QI <9 _ 7 J O{ O.a: &d 0.z , S ., z - OC t t? '. ? I s-, I-. 7L Zi- tn 0 I O XC U : Z W " --0 'XW ZV Z: fuf _ _ O A -:I: -- .O W-(3 (4 o z L * - (f) XC '.4 Li. : , . W 7: Li'0 7 u I ti '-?eW ?00- 0 - 1 : C 0 ).(00.9-Ce LI<. CC7U. Ll.J > L.O UJ 3 t L&IL.J04-LHi'-'(3 ~ CELJ7: '4 kr 0 O r '4 L C a: VJ fl)(Q U9-( 0 0 rCC _ ( Of '.C4:: w OM o: C 3 --o WOMIe M w eM - LI:CE IU ---- .Wow LIt. O SOcr C z U))I 0. O tr I U..L -t - ()-ti U) . (L -U)U. -_ C. (U < zrozrsrwrxrc< . :7- 0/I t,.. c~r.4.I-4- I--4I.- C I-- a- g3 : .0 ~ 7I 5-.a 03 W WO ?C H C] , - , )o aa 0I EL m C - W 420 a C) , Z 0 .- W d 0 C. -c II. 0 - - -. 0 0 (x C - abi W*'Q'U)< > 5- C 0 Li (z L - (?C1 0 3 x cr 0 m VCE ox'C -o- - j W CA 0-' a = ce #--0a A. to 4 ',,), 0 -a c3 E >-I Iz ta 9A )I I.- IA1t- CE In 0 - Li Vi W * O C U 0C CE .-i C. zr ix .- U) re 0 0 0 aC3 0 Ud OA s o 0 0 (C C 4 tau - n 3 &La I- C-- + cm (U C. 0 000000 .3 E L c . ei C- C (C * - F - i L (ai C. a-' LI %ZC, 0 0 03 0- 0 0 W Ua. C. 0? 'l I C-'I 4 E Li E -'A cc Q2 0 * 0 cr 0aIA "G.Q a1i0 J (C X-5- 0.4 (C(~~~~~n C -A .A Uft In IA %m ~ a 4VI- IA 11511 a . CE IF-> z n a F - r.- r- r-- - 0 - U W Ir ra W WI '4* ? 4* '-* Z0 9* 00 F Sal 0. Li ni C t 5&- Cz C (C r- is CL Q. IC z qj Q tj %J V~~~~~~a C4azj xc3 a i-WYVO aa ? WI aO~4tO~C ~W *4 LL C10 O a OQI~t~i~01fn ~~ 49 o..-.to r--~I~ ttctr P-! -1~ ~ -- - c'crXb- - - u -es - - -Il us z O~J rD to -- -, --a t ,.t to 4.-00 (U- - _t3~~~~~~~~~~~~~~~~~~~~~~~~~~~ 4h i I I"I mm t,'ZwN <>1 C _m- C3 _ w Vt'nC W tAvvv- CO *u (0 to t~tz* I- o - IF- 3 W:9 LnLw co t ', t lD t.I) r Cb kl LI > -r in co t% pi tpt ru : t- V) ED t'W t-- C u o a f~ ru nurumninjui? tonininin w co u.-Usu t7mramri tv - (n7 I CtLi ( -fL f,~ 4m---th-- C -- r l.,m WCC a ?,-; -t f -- uq C>n (DLNt . - D" coS tl-1 r- L t-sonLA C;~t V_?,7X ., 9 t~C*Q C n@G7? Mn v Xsa -o t7> I u i=>.7t tA t~ 4r)4 ttrlc~lt, r; W cLa1- t-,7nO ~ s)t AZ J__O C tU 0JQ U SutfS1 SU U X cu lu t u p u u f7 ru r Utlu IV pi nnivU l tJ1 tf CKa cox *w) -rC- c f sn ub C) no Txzwr--t r- t I in t m :* 0 La r -w CR tN fn - tN) t-tVp _-_tN < >-V --- t t7---4 m |,t -LR'U)-t _ltn-8E bO t>N Ns-OF btws>^s nt~aa@ o 0 H_H If I-.~ 1 a-. a- . 0 a- z 0 \~~~~~ o C.~~~~~~~ H <\~~~~~~~ Z? ~?' \ -Je~~~~~~~~~~~~~~~~ 0 LUrrw ~~~~~~~~~~~~- T0) \3) ~~~~~~~~~~~110~~~~~C1FL ---- -- -- 0~~~~~~~~~ 0~~~~~~~~~~~ 0 C 0p~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~029~~~~~~~~~~~~~~~~: Q: u z0 z- H -J LUa.~~~ LLI0 0 0 CL LU CI4 ~ ~ ICi: z0 z 4, 11 Li i LU 0. ~~~~~~~~~~~~~~~~~~~~ Li.e C, 111 v ,,,,'C e+ ~-+ V C 1-0c~ .X !-- C _ c .. I C I H X r tox nt __~a H 0 <- _rr~ LL0 r l- . I ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~~ rF- zu wi, 112 C14n o 0LL. m\~~ v \ m XZco C 0- ~". , QL " \ 0 At z6o C 5-o~ o ~ ~GUT C a< - - X o o,~ux0 5- . <-~ C Oi -- .1U 1 3 L - 0 a WDL z <0o< 0 co THE CORRELATION PROBABILITY: o SUPPORTS MULTISENSOR CORRELATION OF DISSIMILAR SOURCES * RECURSIVELY CALCULABLE a PRESERVES MEMORY (TRACK HISTORY) e TENDS TO 1 FOR CORRECT PAIR * BUILT IN FAULT TOLERANCE o 'COMPARISON' PROPERTY * EASILY, DIRECTLY INTERRPRETED FIGURE 7 SOME PROPERTIES SO FAR OBSERVED 114 A UNIFIED VIEW OF MULTI-OBJECT TRACKING Krishna R. Pattipati Nils R. SandeZZ, Jr. Leslie C. Kramer ALPHATECH, Inc. 3 New England Executive Park Burlington, Mass. 0803 115 A UNIFIED VIEW OF MULTI-OBJECT TRACKING by Krishna R. Pattipati Nils R. Sandell, Jr. Leslie C. Kramer ALPHATECH, Inc. 3 New England Executive Park Burlington, Mass. 01803 1. INTRODUCTION In many practical situations, dynamic systems are subjected to abrupt structural and parametric changes at random instants of time. A representative sample of such switching environments include: multi-object tracking problem in surveillance theory with realistic features such as uncertain observations due to false alarms caused by low signal to noise ratio or clutter; missing measurements due to detection probability less than one caused by noise or noise- like interference; non-uniform media; and changing target character- istics due to maneuvers [1]-[21]; * system failures caused by sudden external disturbances such as occurs in a power network subjected to transmission line trippings, genera- tor shut-downs, malfunction of protective equipment, and the like [22]-[26]; * modeling uncertainties produced, for example, by the linear, finite- dimensional approximation of nonlinear, distributed and time-dependent dynamics of a chemical process or a power plant [27]-[29]. In this paper, we provide a general mathematical framework for classifying the existing state estimation and hypothesis testing problems C[1-[291 arising in systems subjected to random structural and parametric disturbances. The mathematical approach is based on an event-driven, linear stochastic system model comprising a hybrid (i.e., continuous and discrete) state space. It is shown that the problems of multi-target tracking in surveillance theory, Markov chain- driven systems, estimation under uncertain observations, maneuvering target tracking and system failure detection are special cases of this general formu- lation. It is generally well known that implementation of the optimal (in the sense of minimum mean-sauare error) state estimate for these problems requires an exponentially growing number of hypotheses and, hence, memory and computational resources. Therefore, the main thrust of the previous research has been to devise techniaues for reducing the number of hypotheses with little or no com- promise on optimality. The general problem formulation of the present paper provides a common intellectual framework for comparing numerous existing (and proposed) hypothesis reduction techniques, and facilitates the development of a general purpose package for state estimation in switching environments. The computer algorithms permit a convenient comparison among various approaches in common numerical terms. 116 The paper is organized as follows. In Section 2, the estimation problem is formulated in the framework of an event-driven linear stochastic system model with hybrid state space. The optimal Bayesian estimate is derived in Section 3. The techniques for reducing the exponentially growing number of hypotheses are discussed in Section 4. Finally, the issues involved in the choice of a programming language for the implementation of the estimation algorithm are discussed in Section 5. It is concluded that the programming language PASCAL affords a flexible, general purpose implementation of the estimation algorithm. 2. PROBLEM FORMULATION We consider a discrete-time, event driven system of Fig. 2-1 described by the state and measurement equations: x(k+l) = Ax(k) + w(k) + e(k) (2-1) z(k) = Cx(k) + v(k) + f(k) (2-2) where the matrices A and C, and the deterministic bias vectors e(k) and f(k) are functions of a discrete-state stochastic process, q(k) s Q(k). The number of levels (i.e., number of discrete states or hypotheses), nk in Q(k) can be stage dependent. We call the set of hypotheses Q(k), the local set of hypoth- eses at stage k. The noise sequences w(k) and v(k) are zero-mean, white- Gaussian noise sequences with covariances W(k) and V(k) dependent on the discrete-state process, q(k). Clearly, the complete state description of the system at any stage k entails the mixed Markov process {x(k), q(k)}. It is required to find the minimum variance estimate of x(k) given the measurement set zk A [z(l), z(2), ..., z(k)]* i.e., the conditional mean: x(kjk) = E{x(k) zk} = f x(k)p(x(k) tz )dx(k) (2-3a) x(k) and the conditional error covariance: (kIk) = E{ x(k)-x(klk) [x(k)-x(k k)]' (2-3b) The complexity of the event-driven system,and consequently of the estima- tion problem, is determined by the dynamics of q(k), i.e., number of levels n of q(k) s Q(k) and the nature of dependence of q(k) on the past measurement history, zk-l and on the discrete-stochastic process history, qk-l = [q(l), q(2), ..., q(k-l)]. A partial list of problem formulations subsumed by Eqs. 2-1 and 2-2 is provided in the balance of this section. 2.1 MARKOV CHAIN-DRIVEN SYSTEMS The process q(k) takes cn values 1, 2, ..., N and is described by the Chapman-Kolmogorov equation r (k+l) = w (k)P (2-4) The problem formulation may be generalized to include noisy observations on the discrete-state stochastic process, q(k). In surveillance context, this general- ization corresponds to data transfer from various field commanders to the data fusion process. 117 where r_(k) = [71l (k), 72(k), ... T(k)] is a row vector of discrete Markov state probabilities with 7i(k) = Ptq(k)=i} and P = [Pij] is a matrix of (possibly time varying) transition probabilities. That is, Pi = P{q(k+l) = jjq(k) = i, qk 1, zk} = p{q(k+l) = jlq(k) = i}. The matrices i, C, and the bias vectors e and f are known for every q(k) = 1, 2, ... N. It may be noted that when q(k) is absorbing Markov chain (i.e., P=I), the formulation reduces to the well-known multiple model adaptive estimation (MMAE) algorithm, first derived by Magill [27]. 2.2 STATE ESTIMATION UNDER UNCERTAIN OBSERVATIONS The problem of state estimation when there is a nonzero probability that the state cannot be observed has been studied extensively in the literature [1]-[6]. In this case, the stochastic process q(k) is a binary random variable and affects only the observation Eq. 2-2. Thus, C = q(k)C ; q(k) = 0, 1 (2-5) and V(k) = (l-q(k))V 0 (k)+q(k)Vj(k) (2-6) The dynamics of q(k) are represented by a two-state Markov chain described by an equation of the form (2-4), or as a binary independent random process. In the latter case, we have P(q(k)=lk-l,zk-l)= p(q(k)=l) = 6 ; p(q(k)=0) = 1-6 (2-7) 2.3 TRACKING IN A CLUTTERED ENVIRONMENT The basic problem of tracking multiple targets is as follows: "given a set of returns (i.e., measurements) on the targets in each scan (stage) k, associate the measurements with the correct targets and determine an estimate of each target's state." Various solutions have been proposed for this prob- lem, and [7]-[20] represent the state-of-the-art in multitarget tracking. In a multitarget tracking problem, the q(k) process affects only the measurement subsystem. In order to specify the q(k) process for the tracking problem, we first consider a single target with cluttered measurements. If Mk is the number of measurements at stage k, and given that only one of them could have originated from the target of interest, then q(k) assumes one of the following nk = Mk+l values: q(k) = 0 if none of the returns is correct (2-8) i if the ith return is correct, i=1,2 ,...,Mk The dynamics of q(k) may be represented in various ways. For example, Bar- shalom and Tse [8] hypothesize a missed detection probability of a for the target, and then assume that no inference can be made on which of the Mk returns is correct from past data; so that k-l k-l 1 k-1k-! p(q(k)=Oq k,z ) = k; and p(q(k)=iql z-) _ - (2-9) Mk Clearly, other representations are possible. Note in particular that the pro- blem of state estimation under uncertain observations is mathematically equivalent to single target tracking in clutter with Mk=l. (Compare Eqs. 2-7 and 2-9). In a multitarget tracking problem with Mk returns, the process q(k) E Q(k) represents a particular assignment of Mk returns among all the possible targets. These are Nf false targets, Nn new targets, and Nd of the previously hypothesized Nt targets (confirmed and tentative) such that Nf + Nn + Nd = 4k . With the assumption that each target can be associated with at most one measurement for each hypothesis (assignment) per scan, Q(k) is the set of all possible assignments of Mk measurements among all the possible targets. The number of possible assignments (i.e., levels) nk are: min(Mk,Nt) Mk-Nd Nk = Z __Z 0 ndn (2-10) Nd=0 n where -N. N. (n = -( d) N -N (! N!NdNn) (2-11) The above specifications of nk, where new targets can be initiated, corres- ponds to Reid's algorithm [15]. He specifies the q(k) process as follows: a particular assignment containing Nd of Nt p(q(k)qk-l /Z ) = P previously hypothesized targets, Nf false targets N new targets such that Nd+Nf+Nn=Mklq ,z ) an assignment containing Nd, Nf, and Nl all} assignments containing Nd , N, Nf P{NdNf,Nnll k-,zk- } k-l k- d qI k z } * P{N, k- zk- (2-12) Reid assumes that all the assignments containing the same Nd, Nf, and Nn are equiprobable, i.e., Pdn = 1/ndn. Also, the number of targets detected at stage k, Nd is assumed to have a binomial distribution: P{NdI k-i k- jNt d Nt-N d PN = (2-13)d| = ) PD (1-P D Nd where PD is the probability of detection. Finally, both the number of new targets and false targets are assumed to follow Poisson distributions given by: Nn P Y) ? exp[-a Y]k-1 k-1 (a * exp[ y] n (2-14) n Nf k-l k-l (ScY) ? exp[-f Y]k-l (fy) [f (2-15) P{Nf 5 _ N,! 119 Thus, Eq. 2-12 simplifies to: N N k( Y) (Y) Nt-Nd P{q(k) q :z PD dPD t d exp{-(f+Sn )Y} (2-16, When no new targets are allowed, Eq. 2-10 simplifies to: min(MkN't) Nt! n= =O NtNd (2-17) and corresponds to Bar-Shalom's [91 version of the multitarget tracking problem. For example, when Mk = Nt = 3, we have nk = 86 and nk = 34. Thus, the number of possible assignments at each stage k are, in theory, considerably larger in Reid's version of the multitarget tracking problem [15] than in Bar- Shalom's [9]. When clutter density is high, the difference (nk-nk ) could be quite large (zMk 2Mk-l assuming Mk >> 2 and Nt = 1). This suggests that the initiation of new target tracks via an operator-interactive process may be preferable to an automated track initiation! 2.4 MANEUVERING TARGET TRACKING The usual analysis of a tracking situation consists of describing the target dynamics by the state-space equations of the form 2-1 and 2-2 and designing a Kalman filter to provide the conditional mean estimate x(klk). This type of analysis works well until the target suddenly changes course or speed. One method of modeling a maneuver, suggested by Moose [21], is to let the bias vector e(k) in Eq. 2-1 assume N distinct (known) values, eI, e2, ..., eN. The transitions between any ei and ej are modeled by a Markov process.* qc(k) = i<------e(k) = e.; i=1,2,...,N (2-18) The dynamics of q(k) are described by Eq. 2-4. 2.5 SYSTEM FAILURE DETECTION Recently failure detection and identification (FDI) has been the subject of considerable interest [22]-[26]. An important subclass of FDI problems is the detection and estimation of soft failures, viz., the bias errors and changes in the noise levels. The failure events are modeled as additive dis- turbances. The failure detection case can be modeled by x(k+l) = Ax(k) + w(k) + q(k)e (2-19a) z(k) = Cx(k) + v(k) + q(k)f (2-19b) where q(k) = 1 for failure and a(k) = 0 otherwise. The bias vectors e and f are, in general, unknown and have to be estimated online. However, we will not address this question here, as there is a large literature on this subject [24], [35]. It should be noted that Moose assumes a semi-Markov process for the transitions, but actually uses a Markov process model in the final implementation. 120 This concludes the list of special cases of our problem formulation. We now turn to the derivation of optimal (Bayesian) state estimate. 3. OPTIMAL BAYESIAN ESTIMATE The sequence of measurements zk = [z(l),z(2),...,z(k)] is a function of the particular sequence of the random process qk = [q(l),q(2),...,(k)]. Therefore, if we define Qk = Q(1) x Q(2) x ... x Q(k) = j } (3-1) as the global set of hypotheses at stage k, the cardinality of Qk is Qk1 kN (k)= |Qk | = H n. (3-2) i=l The conditional mean, x(k k) is given by x(kjk) = x(klk ,ik)P(q zk) (3-3) qksQk - In Eq. 3-3, x(kjk;jk) is the conditional mean of x(k) given zk and a parti- cular sequence of the stochastic process, jk and P(qklzk) in the posterior probability of the sequence qk given zk. Clearly, x(krk) is the convex com- bination of the conditional estimates x(klk;qk). The conditional error covariance, Z(kjk) is given by Z(kfk) = Zk ' ?k aQk qkF~~~~~~~~~~~~k~~~ ~(3-4) where Z(klk;ak) is the a posteriori estimation error covariance matrix for a given qk There are two important features of Eq. 3-4 that are worth noting. First, the conditional error covariance, I(kIk) is a function of measurement data, due to the presence of terms P(qklzk), x(klk;qk) and x(klk). Second, Z(k k) is not just a convex combination of the covariances of individual terms, E(k k;qk): it includes additional terms of the (positive semidefinite) dyadic form [x(klk; q)-x(klk)] [x(klk;2k)-x(klk)]'. This shows that the covariance in- creases by the presence of terms whose estimates are significantly different from x(klk), weighted by P(qklzk). The structure of the optimal Bayesian estimation algorithm is shown in Fig. 3-1. The conditional mean and covariance are evaluated as follows. Since the density p(x(k) zk,qk ) is Gaussian, the density p(x(k+llzk;qk) is also Gaussian The corresponding means x(klk;qk) and x(k+l k;)k) are given by a Kalman filter of the form x(k+llk;q k) = Ax(kIk;qk) + e(k) (3-5) x(k+lfk+l;atk+l) = x(k+lj1;) + Kr(k+l;_k+l) (3-6) where r(k+l;ak+l) is the innovation process: r(k+l; +l z(k+l) -Cx(k+lk+lk;q ) - f(k+l) (3-7)r~k~l;~ ) = ~k~l_) - - Here, K is the Kalman gain calculated from the following equations: 121 Z(k+1Ik; k) AZ(kjk;j )A' + W(k) (3-8) S(k+l;q ) = CZ(k+llk;q )C' + V(k+l) (3-9) K = Z(k+llk;k )C'S- (k+l;ak+ l) (3-10) Z(k+lik+l;q ) (I-)C)7(k+llk; ) (3-11) It should be noted that A, e(k) and W(k) are functions of q(k), and C, f and V(k+l) are functions of q(k+l) in Eqs. 3-5 through 3-11. Here, Z(k+l)lk;qk) is the a priori estimation error covariance matrix for a given ak. The innovation process r(k+l;jk+l) is zero mean, with covariance S(k+l;qk+l) and is normally distributed. The only remaining item in Eqs. 3-3 and 3-4 is the a posteriori probability P(qklzk). Using Bayes' rule, this probability is given by P(q kZ k) a -aK (z ki )p(P ) (3-12) where ak = p(zk) is a normalizing constant, pCzk qk) is the likelihood function, and P (q) is the a priori probability of the sequence qk. The likelihood function p(zkfqk) may be evaluated recursively from kzi k k-l k p((k) k-l k-l) (zk-lj qk-1) Ik1k-1 (3-13) Since in most applications q(k) is independent of zk- l, the likelihood func- tion simplifies to k jk) k zk l k-1) k-l)p(zk) = p(z(k)la Z ) p(zkl ) (3-14) Using the normality of innovation process, the negative log-likelihood function may be computed recursively from - k q k-1l Jj+ (.; qk)S - (k; rk; k X(q) = -ilg p (z |q) = A( ) 2 r r(k; ) + dim(z(k))log (2r)+log S(k;)k (3-15) with X(qi) = 0. Since r is Gaussian and white, X(qk) will have a chi-square distribution with k=l dim z(i) degrees of freedom. The hypotheses for which Xexceeds a certain threshold may be dropped [13], [16]. A more direct recursive formula for P(qkczk) can be derived as P( p(z(k) qk k-l k-l k-l 1 -1) (3-16) where Ck = p(z(k)izk-l) is a normalizing constant. Again, the hypotheses for which P(qak zk) is less than a certain threshold may be deleted [14], [15]. As mentioned earlier, the complexity of the estimation algorithm is determined by the quantity P(q(k)aki-l,z - in Eq. 3-16. Equations 3-3 through 3-16 constitute the recursive relations for the optimal Bayesian estimate. There are two important characteristics of this estimate that are worth noting. First, the optimal (condition mean) estimate 122 is a nonlinear function of the measurements due to the terms P(~ zk). Second, the computation of x(kjk) requires an ever-growing number of filters with an associated growing memory and computational requirement. Thus, at stage k, we need Nc(k) (cardinality of Qk) elemental estimates, x(kjk;qk); NC(k) covariance matrices, E(klk;qk); and Nc(k) posterior probabilities, P(ak1zk). This is clearly impractical and hence, techniques must be devised to reduce the number of hypotheses. These are discussed in the next section. 4. HYPOTHESIS REDUCTION TECHNIQUES The basic idea of hypothesis reduction techniques is to transform the global set of hypotheses Qk into a smaller global set Qk such that the memory and computational requirements are minimized, while maintaining the "near" optimality of the estimation algorithm. The various available techniques may be categorized into the following five groups: (1) use of a simplified dynamic model of the q(k) process; (2) hypothesis deletion; (3) hypothesis merging; (4) decoupling of hypotheses (cluster formation); and (5) use of system con- straints to advantage. A practical estimation algorithm may employ one or more of the above reduction techniques. Also, note that the hypothesis reduction techniques have close analogy to the methods of aggregating Markov chains [30], [311. This analogy is pursued elsewhere [32]. 4.1 SIMPLIFIED MODELS OF THE q(k) PROCESS When q(k) is described by an N state Markov chain with N absorbing states (i.e., the transition probability matrix P=I), the global set of hypotheses Qk is independent of k with a cardinality of N. The optimal Bayesian estimate can be implemented by a bank of N Kalman filters, each corresponding to one of the N absorbing states. The resulting estimation algorithm is the well-known multiple model adaptive estimation (1MAE) algorithm [271]-[29]. A more realistic description of the q(k) process, however, is via a weakly coupled Markov chain. In this case, the transition probability matrix, P=I+EB, where s is small. That is, P is diagonally dominant. Note than in this case, even for small _, the optimal Bayesian estimate x(k'k) is the weighted sum of Nk elemental estimates x(kjk;qk) as in Fig. 3-1. However, intuition suggests that as long as a/s >> max (TRi), where a = (tk-tk-1) is the time step and TRi is the "response time" of the ith Kalman filter in the multiple model, then the MMAE algorithm should be "nearly" optimal. That is, the posterior probability of hypotheses with nonidentical elements is negligible and the number of hypotheses that are almost identical, and that have nearly identical estimates and co- variances is large. The former set of hypotheses can be deleted, while the latter can be merged (as shown in the following) so as to reduce the computa- tional load of the estimation algorithm. Thus, the weakly coupled Markov chain formulation provides an analytic framework to study hypothesis reduction tech- niques and is a subject for future research. 4.2 HYPOTHESIS DELETION The basic idea of hypothesis deletion is to simply prune unlikely hypothe- ses in view of the measurements or via a heuristic technique. A straight- forward heuristic pruning procedure is to simply limit the number of hypotheses to be included in the estimation algorithm. This is the approach employed 123 by Singer, et al. [7] in developing an N-scan (stage) filtering algorithm for a single target tracking. At each stage k,kthis algorithm keeps hypotheses corresponding to the last N scans, viz., i=--N+l ni hypotheses, where ni = Mi + ., and Mi is the number of returns at stage i. A remarkable conclusion of their simulation was that with only N=l, near optimal performance was achieved. However, this conclusion is not, in general, valid in the multi-target tracking problem. For example, Keverian and Sandell [16] found that it was essential to have N>l in tracking targets with crossing tracks. Pruning on the basis of measurements may be accomplished using the likeli- hood function p(zklak) (or equivalently, the negative log-likelihood function X(jk)),or the a posteriori probability p(ck zk). The former approach was employed by Smith and Beuchler [131, Fraser and Meier [20], Sittler [111] and Keverian and Sandell [16]. The method of Keverian and Sandell [161 computes the likelihood function after each branching and keeps only the M most likely hypotheses at each stage. This prevents hypotheses with probability less than the current minimum of the M hypotheses from being extended. The choice of the parameter M is critical and determines the complexity versus optimality tradeoff of the estimation algorithm. The pruning technique based on posterior probabilities was employed by Reid [15]. He eliminates all hypotheses with a probability less than a threshold, a. Morefield [143 uses an optimization- based integer programming formulation to delete unlikely hypotheses; this is a batch processing approach. An illustration of a typical pruning technique is shown in Fig. 4-1. If Qk is the set of hypotheses remaining at stage k, then the correspond- ing prior and posterior probabilities must be normalized to have sum equal to 1. Thus, () = P( [ k ( ) (4-1) q sQ and a similar expression for the posterior probabilities. k forms the basis for enumerating the subseauent set of hypotheses Qk+l = Qk x Q(k+l) at stage k+l. 4.3 HYPOTHESIS MERGING Hypothesis merging is the process of combining hypotheses in a "suitable" manner. The hypothesis merging techniques may be categorized into single- stage (also known as zero-scan or nonback-scan) and multistage (or multiple scan) algorithms. The multistage algorithms can be subdivided into fixed- scan (or fixed-depth) algorithms [115] and variable-scan (or variable-depth) algorithms [16]. A single-stage algorithm allows only one hypothesis to remain after each stage. Prime examples of the single-stage algorithms are the prob- abilistic data association filter (PDAF) of Bar-Shalom and Tse [8], the approxi- mate Bayesian estimation algorithm of Sawaragi, et al. [2], and the algorithm of Athans, Whiting and Gruber [3]. The algorithlms of Reid [15] and Keverian and Sandell [16] are representative of multistage algorithms. We briefly summarize these two types of hypothesis merging below. 4.3.1 Single-Stage Algorithms The single-stage algorithms of [2], [31, and [8] make the fundamental assumption that the conditional density p(x(k)lzk-l) is Gaussian with mean 124 ^n S -. ' in _ ,-, - so that .-- 1 .= ! ?- z) sc .hct the set or hvctheses at ce x () = ) hs n .l._e,: ts; el (3) hs condit-o..al 17s-. --k~,^ .J' -. 1 -re sLmily p(z(k) ;z-l ',C(k)) and h'ave kro-Cm form. In the prababilistic data association filter (?DA-), ,(k) takes on values 0,1,2,... ,Mk as in Eq. 2-8. Thereforef, the posterior probability Eq. 3-16 simplifies to P(q(k)=i ) = Ski = bk p(z(k)lz ,q(k)=i)P(q(k)=iJ z ) (4-2) where bk = P(z(k) zk-l) is a normalization constant. Usina the model of P(q(k)=iz-k!l) in Eq. 2-9, and the assumption thlat P(x(k) Izk-l) is Gaussian, Skis are easily computed. Since the q(k) process affects only the measurement subsystem and since p(x(k) zk- 1) is assumed to be Gaussian, a single Kalman-like algorithm (with a data dependent covariance matrix satisfying a Riccati equation) is adequate to implement the PDAF algorithm. The derivation is straightforward and may be found in Refs. [8], [32]. The structure of hypothesis merging is displayed in Fig. 4-2. The algorithms of (2], [3] have precisely the same form as in Fig. 4-2, but somewhat simpler due to inherently simpler assumptions on their structure. 4.3.2 Multistage Alaorithms The basic idea of a multistage algorithm is to combine only those hypotheses that have "similar" state estimates and covariances. Two hypotheses qk and k"of the same length are said to be similar, if their corresponding state estimates x(ktk;,k'), x(kjk; ak) and their covariances Z(k1k;ck'), Z(kik;k ) satisfy the following criteria: lx i(kik;k )-i (k ik;a) < Zii(kjk; ) + ,(kk; and (4-3a)* |Z ,(ktk;q )-r (k k;q )d t kr < (4-3b) Zii(kk;q ) This situation may occur, for example, when two hypotheses are nearly identical except for a few stages back and there is a limit on the total number of stages to be considered in the algorithm. When these earlier stages are dropped, Eq. 4-3 is satisfied. Then, qk and ck" may be combined into qk such that the hypothesis ak has the following properties: k k' k"q = q or a (4-4) k( = P(q' ) + ?(a ) (5) Note that this assumes independence cf and qk", is generally not vai id. Recently, Bar-Shalom [36] -as provided a formulation that relaxes this asis=1rnton at the cost of solving an additional matrix Iinear equation for each pair of -seuences. 125 kj k = P/k' k kl ? kP(a zk) = p(k Zk) + p(ak *z) (4-6) 1z )T _(k ka ) = --- (c zk) (4-7a) [(klk;qk) )p(kk; ak)kZ (k k;Q ) Pt Pk 1 zk kl zk k " k" (4-7b) * P (qk zk )+ j k fc' k' kit, . k lP((kjlka)-x(klkk) x(kk;q )-x(klk;crk )}j P(a _z) . Equations 4-7a and 4-7b assume that the sum of two nearly identical Gaussian densities is Gaussian and that the mean and covariance of the resulting dis- tribution should be the same as the mean and covariance of the sum. Thus, hypothesis merging eliminates the redundant enumeration of the hypothesis tree as shown in Fig. 4-3. 4.4 DECOUPLING OF HYPOTHESES (CLUSTER FORMATION) The basic idea of this technique is to decompose the set Q(k) of nk levels at stage k into Zk independent sets Ql(k), Q2 (k), ... QZk(k) with smaller number of levels nlk, n2 k, ... nZkk, respecitvely. There are at least two ways to decompose Q(k) into Qi(k), viz., additive and product forms.Zk In additive form, the levels nik, n2k, ... nZkk are such that nk = i1 nik and is appropriate when q(k) process is represented by Zk decoupled Markov chains. The multiple model adaptive estimation algorithm falls into this category with Zk=N, the number of states in the Markov chain and nik=l for i=1,2,...,N. In product form, however, a level in Q(k) represents a particular way of combining the various levels of Qi(k), Q2 (k), ... , QZk(k), taking one level from each. The product form is appropriate in a multitarget tracking problem, where the entire set of targets and measurements can be divided into independent clusters via gating (validation region) criterion [83,[9] , [15], and [16] based on resi- duals or the actual observations. In order to illustrate the effect of the decomposition on computational requirements, assume for the present, that .k=Z, nk=n, and nik=ni, i=1,2,... ,: at each stage k. The decomposition of Q(k) into 2 independent sets Ql(k), Q2(k), ..., Q%(k) implies that the set of hypotheses Qk of cardinality Nc(k)= nk can be decomposed into Z independent sets of hypotheses Q, i=i,2, ..., each with smaller cardinality Nci(k)=n4- Thus, the total set of hypotheses is reduced to Qk with cardinality N(k) given by N (k) = E N (k) (Additive Decomposition) i=l (4-8) N (k) = H N i(k) (Product Decomposition) i=l In actual implementation, the order of computation should be 7,x, p(q Iz3) or P(s ) for maximum efficiency. 126 Typically, Nc(k)< min J(u,v) u u BY CHOOSING v AND v, OWNSHIP CAN BE SURE OF A0 PERFORMANCE NO WORSE THAT F(vo ,vl ). 162 CONDITIONS FOR OPTIMALITY NECESSARY CONDITIONS FOR v , 1 AND v DEFINE H = -e - + Pl (VX1 - v sinCp) + P2(vx2 - v cosCp) iIx 114 X ea HU 11- 1P = 1 - .14 Ilull) - VP2 IIx allu II- ~v p = 2 (4 .14 Ilu |1) -p llx ll P.(tf) = -Yx.(tf) i = 1,2 THEN: aH ** *(x,u ,v,p) = 0 FOR ALL t. 3H C---- (x ,v dt 0v -6 0t 163 NUMERICAL EXAMPLE y P(tf) Es~~~~~~o)~tf) tP0 E(0) P (0), x TRACKING TRAJECTORY FOR OWNSHIP AND TYPICAL EVASION MANEUVER FOR THE TARGET TWO PHASE TRACKING PROBLEM ? TWO SIX MINUTE LEGS * EVADER RUNS AWAY FROM PURSUER ON A COURSE OPPOSITE TO THE BEARING LINE. * a = .14, 8 = .10 AND y = .0002. 164 PERFORMANCE CRITERION AND THE NECESSARY CONDITION AS A FUNCTION OF TRACKIN SPEED Initial Bearing = 450II x(O)I = 4 Kyds y = .0002 c = .14 B = .10 .002 - .0002 * h(v)J(u,v) h(v) .001 _ .0001 [30-60 5 1 10 2 -.001 -. 0001 -.002 - -.0002 J[15-60 -.003 -. 0003 J[0-601 165 PERFORMANCE CRITERION AS A FUNCTION OF TRACKING SPEED FOR TWO DIFFERENT STRATEGIES (TWO SIX MINUTE LEGS) Initial Bearing 450 Ij = 4 Kyds y = .0002 = .14 = .10 .002 J (u,v) .001 0 5 IG0 15 20 25 v, Knots -.001 [15-90] -.002 -.003 \ 15-60] 166 U) O \ O ? D 11 \I 0)- - CD71~~~~~~~ C \ -- o L, \ r II'xcDO H \ U)o~~~~~~~~~~~ 6 H \LLLr'~ _ 16- 7 ~'~ ~ ~ ~~ 167 Q~~t- C.tY~~~~~~~~~~~~~~~~~~~~~~-Cw -~ __o c-> CDO ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ Cm N 0 Oi *- \xl CD \ 0 L\o\ U- ' >~0CD \W LL \-- I L%ar \ kD LU U) )\`- D3~~~13 X x0 Cr) U] 2: U] 03t\\ C2:v 0a 0 Nu CDO II t LI a II ,I 0 0 0?? O LLI \II d 0CD O II UO CD LULtJ ? X \\ \_ CDD I - z E Z \CDO~~~~ \v16 C) 01 uJ 0 OD tn~I ID II L ' +- a) ? mn I c.OJIW vLU xl -=o .i- o - (\I ii- I II \-- 11 II CD \ Z O O 00Ln CD LU 0 o LUJ LL_ -- U) Z\o x > >: 3 Z I -- LU 'L C.C 0 1 DESCRIPTION OF AND RESULTS FROM A SURFACE OCEAN SURVEILLANCE SIMULATION Thomas G. Bugenhagen Bruce Bundsen Lane B. Carpenter Applied Physics Laboratory The Johns Hopkins University Laurel, Maryland 20810 (This work was supported by the Naval Electronics Systems Command under Task G3AO of Contract N00024-8Z-C-530Z with the Department of the Navy). 171 Description of and Results from a Surface Ocean Surveillance Simulation T. G. Bugenhagen, B. Bundsen, and L. B. Carpenter The Applied Physics Laboratory, Johns Hopkins University Laurel, Md. 20810 ABSTPRACT As a part of the Over-the-Horizon/Detection, Classi- fication and Targeting (OGH/DC&T) Engineering JAnalysis Program performed by The Johns Hopkins University Applied Physics Labora- tory (JILU/APL) in support of the Naval Electronic Systeaa; Co^-_and (NAVtELEXSYSCOM) (PKE 108-2), it was necessary to establish ocean surveillance systen requirements in a realistic environment. Thne area tracking and correlation (ATAC) model was developed and used to perform analyses for defining these surveillance requirements in a setting where high interest ships interact with merchant ships traveling in shipping lanes. This work was supported by The Naval Electronics Systems Command under Task C3AO of Contract N00024-81-C-5301 with the Department of the Navy. 172 1. INTRODUCTION Adequate surveillance of an ocean area may require tracking large numbers of targets over a broad area using a variety of sensors. Often, the density of background ships in the area is high enough to cause interference with the tracking of high interest targets. The high interest targets themselves may execute maneuvers at unkrnown times, further complicat- ing the situation. As a part of the OTH/DC&T analysis effort, it was necessary to determine the requirements of ocean surveillance system parameters for this setting. In some previous analyses, the merchant ship traffic was assumed to be uniformly distributed throughout the area. This is quite different from the actual situation where the merchant ships travel in shipping lanes. To define the ocean surveillance requirements in a realistic environment, the area track- ing and correlation (ATAC) model was developed. It is a simulation of the ocean surveillance situation including: a. Generation of merchant and high interest ship positions as they move across an arbitrarily defined ocean area in defined shipping lanes; b. Development of simulated sensor reports from a variety of sensors; c. Initialization of tracks; d. Correlation of the reports to previously established tracks; e. Continuation and projection of the tracks to any chosen time; and f. Measurement of the accuracy of the correlations by comparing with the actual (ground truth) tracks stored in the computer. In the present version of the ATAC model, three different sensors were modeled and\ used in the analysis: two active radar sensors and a passive sensor that is assumed to give reports only on the high interest ships. The passive sensor is assumed to also give a unique identity of the ships on which it reports. Of the two active radar sensors, one is assumed to give only posi- tion reports, while the other gives position and velocity measurements on the ships. A later version of the ATAC model also includes sensors that give line- of-bearing measurement reports. 173 2. THE AREA TRACKING AND CORRELATION MODEL Figure 2-1 illustrates the functional flow when the ATAC model is used in a Monte Carlo loop with simulation of ship traffic and sensor reports. Each block of the diagram will be explained, in turn, in both Sections 2 and 3. The three basic steps to the model are: (1) developing sensor reports, (2) assigning sensor reports, and (3) scoring. The first step is subdivided into three parts: scenario, process noise, and measurement noise. In the scenario, the ocean area is defined (by four corners in latitude and longitude), the structure of the shipping lanes is established, and the density of the merchant shipping is set. In the process noise subdivi- sion, the ship heading errors and velocity errors are defined, after which specific simulated ships and their motion can be generated. Measurement noise takes into account the sensor parameters to generate sensor reports, which are perturbed from the ground truth (actual) positions according to the position error associated with the sensor. In the correlation phase, which makes up all of step 2, the sen- sor reports are associated (correlated) with existing ship tracks in a track file by using Kalman filtering and probabilistic decision making. The time step loop in this part of the block diagram is meant to show that as sensor reports are periodically generated and received (based on the sensor update interval), the correlator associates the reports to tracks. At each time step, certain measures of effectiveness are calculated and then averaged at the end of the time period for which the model was run. After running the ATAC model with one choice of the process noise and sensor parameters, a new set of values is chosen. The process is repeated as shown by the Monte Carlo loop. The measures of effectiveness are collected for each such iteration and used to give the grand averages used in the analysis. ATAC is currently programmed in the PL-1 programming language and run on an IBM 3033 computer. It uses about 1.5 megabytes of storage and about 10 seconds of comptuer processing time for one iteration involving roughly 20 ship contact reports every hour over an area of 40,000 nmi2 in a given 10-hour time period. 174 Merchant shipping lane structure Military ships: numbers and courses Scenario. Step 1A Merchant shipping de nsit . Generation of ship tracks with heading errors, aC and velocity errors, aV Process DevelopingProcess Step lB noise sensort noise reports Ground truth (specific ships and motion) Sensor definitions: Detection prob, Pd False alarm rate Update intervals, -t Measurement Step 1C Position errors, an noise Monte Carlo loop Sensor reports Correlator control Time parameters: step Avg speed assumptions Threshold settings AssigningCorrelation Step 2 sensor .__________________________ reports Correlator: Association of reports to tracks 1OE'sB .utput a^ IStep 3 Scoring MOE's Fig. 2-1 Area Tracking and Correlation Model Monte Carlo Functional Flow Diagram 175 2.1 BACKGROUND SHIP TRAFFIC Background ship traffic in a region of interest (ROI) is gene- rated along shipping lanes defined as a sequence of geographic coordinates (turn points). Each lane has a width established by assuming a normal dis- tribution about a mean path with a standard deviation that can be input. Within each shipping lane, individual ships of a given type are generated from a departure rate, which is either input or calculated on the basis of ship density considerations. Ship tracks consist of a time history of geo- graphic coordinates between which motion is maintained with constant course and speed. The time spacing of the points is small enough so that inter- mediate positions may be determined by interpolation. 2.2 HIGH INTEREST SHIP TRACKS The procedure used to generate the tracks of high interest ships is different from that used for background ships. A master file contains data for a number of sample paths, each representing a ship maneuvering randomly to avoid detection while attempting to maintain a particular average speed and course. The state equations from which these tracks are generated are written in a latitude-longitude system, and the ship follows a rhumb line between ob- servations. Input parameters are: a. Initial ship latitude and longitude; b. Maximum ship.speed permitted; c. Mean ship velocity desired; d. Mean target course (angle); e. Standard deviation of maneuver speed, and beta, a parameter describing the relationship between a random change in course and a change in speed. 2.3 GENERATION OF SENSOR REPORTS After the ship paths and movement along the paths have been established, sensor reports (using data from this simulated ship traffic) are generated. Simulated ground truth position data on ships, within some geographical region of interest, is perturbed according to assumed sensor error distributions. An input detection probability permits missed detections, and the capability of generating false targets is provided. 176 2.4 CORRELATION OF SENSOR REPORTS The correlator attempts to make the associations from the sequence of sensor reports that have been developed. As the sensor reports are periodi- cally generated and received (based on the sensor update interval), the correla- tor associates the reports to tracks. Once the associations have been made, a ship tracker (based on a Kalman filter) is used to develop the individual tracks. This procedure works well as long as none of the ships are maneuver- ing. When a ship maneuvers, a contact report's distance measure, obtained from the tracker, may be large. This may cause the report to be thrown out because it either failed a threshold test or was not assigned in the best hypotheses. To overcome this problem and to efficiently assign contact reports from maneuver- ing and non-maneuvering targets, the correlator has been designed to process four different types of ship motion in stages, with one type of motion processed in each stage. The four stages and the corresponding motion processed are listed in Table 2-1. The correlator processes through all four stages each time a set of contact reports is received. Table 2-1 Correlator Stages for Processing Ship Motion Models Stages Ship Motion Processed 1 Slow, Straight 2 Fast, Straight 3 Slow, Maneuvering 4 Fast, Maneuvering When all the possible associations between contacts and tracks have been made in one stage, the contacts and tracks are removed as candidates for successive stages. Since most of the merchant ship traffic will be follow- ing a constant heading (rhumb-line sailing) at moderate speeds, they will usually be processed in the first or second stages. When they are removed, the remaining contacts and tracks can be associated in a more efficient manner. 177 2.5 TECHNICAL DESCRIPTION OF THE CORRELATION OF SENSOR REPORTS The correlator developed for use in the ATAC simulation is a tech- nique for assigning reported contacts on the ocean surface to existing ship tracks or to previously reported points (to initiate a track). In general, the sensors are assumed to be giving information on many contacts at approximately the same time with the results being correlated with the current track file. In this sense, ATAC does report-to-track correlation. 2.5.1 FEASIBILITY TABLE The construction of hypotheses begins with the construction of the feasibility table. The table is simply a listing of the current track file versus the new set of contacts and an indication of which of the contacts are feasible to be associated with each of the tracks. Feasibility is defined in terms of possible ship speeds and sensor errors. A maximum ship speed, v?max is used to determine the distance a ship could travel in the time, At, since the tracks were last updated. If d.. is the distance from the last position of track i to contact j, then the association is, by definition, feasible if d . < F(V At) +3 + cr3 j ma- max nJ where a . is the sensor error associated with the tast position of the ith track and a no s the sensor error associated with the j contact.nj After determining which contacts are feasible for a given track, the figure of merit is calculated for each of these contacts. The figure of merit used is log fi (Z.) where tiR function f is given by the distribution of the expectedtRosmtioA of the i track after time, At, since the last update and Zj is the j contact. 2.5.2 UNIQUE ASSIGNMENTS There is one exception to this rule of calculating the figure of merit. If there is only one feasible contact for a track and if that contact is not feasible for any other track, then that contact is arbitrarily assigned to the track. This procedure is called "Unique Assignment" in the program. 178 The figure of merit is not calculated. For all remaining locations where a contact is feasible, the figure of merit is calculated. The resulting table of values is called the distance table in the program. 2.5.3 THRESHOLD TESTS The values for the feasible contacts are then subjected to a threshold test to further reduce the number of ambiguities for each track. Specifically, if log fi (Zj) < Th where Th is the arbitrarily chosen threshold, then the contact Zj is retained as one of the possible contacts for track Ti. 2.5.4 UNIQUE MINIMUMS Before going into the hypothesis generation routine, one further attempt at reducing the number of ambiguities is made. The attempt is to try to find contacts that give "unique minimum" values of the figure of merit for some of the tracks. The rows of the distance table represent the tracks currently being carried and the columns represent the new set of contacts to be assigned. For each row, the contact is found that yields the smallest figure of merit. These are called "row minimums." Likewise, for each column, the track is located for which the figure of merit is smallest. These are called "column minimums." When a row minimum is also a column minimum, then an assignment of that contact to the corresponding track will be made if neither of the following two condi- tions is violated: a. There must be no other row minimums for that column (contact), and b. There must be no other column minimums for that row (track). 2.5.5 THRESHOLD LOWERING AND ITERATION ON HYPMAX The number of hypotheses that could be generated when initiating tracks with the first two sets of reports is a function of shipping density and the circular area within which a ship could have moved in the time interval, 179 AT, between the reports. If the ships in the area were uniformly distributed throughout the area, the number of contacts from the second set of reports that could be associated with each contact from the first set is N p7r cAT + 3 /nl + n2 where p = shipping density, v = upper bound on ship speed,S AT = time between reports, and Cnl On2 = sensor accuracy of 1st and 2nd reports. This number by itself can become large. However, the number of hypotheses that can be generated can become very large. As a result, when the update interval or the location errors increase, the correlator processilng time increases dramati- cally. This is shown in Figure 2-2, in which the actual processing time for one replication is plotted versus the update interval. A similar curve for the-sen- sor accuracy would reach the upper limit when o % 4 nmi.n To keep the processing time and costs within bounds and still ob- tain results when AT > 2.0 hrs or a > 4 nmi, it is necessary to reduce the thresholds that limit the size of tie uncertainty areas and, therefore, the number of hypotheses that can be generated. An estimate of the number of hypotheses that could be generated is made. If this number is less than an upper bound (called HIYPMAX), hypothesis generation begins. If the estimate is greater than HYPMAX, the threshold, Th, is reduced by a set amount, and any entry in the feasibility table larger than the new threshold value is removed. A new estimate of the number of hypotheses is made and compared with HYPMAX again. This process is repeated until the estimate is below HYPESAX. The result is called the ambiguity table. Thus, for each track that has not been assigned a unique contact, the ambiguity table lists those contacts that are both feasible and have a figure of merit less than the final threshold value resulting from the iterative reduction process. 180 50 40 30 - (One 10-hour replication) 20 p = 40 ships/100,000 nmi2 40,000 nmi2 A010 0 0.5 1.0 1.5 2.0 nmi 2 AO3.0 0 0.5 1.0 1.5 2.0 2.5 3.0 AT 1 radar sensor update interval (hr) Fig. 2-2 Computer Processing Time vs Radar Sensor Update interval 1 81 2.5.6 GENERATION OF HYPOTHESES To generate the hypotheses, the ambiguity table is used. A pointer is set at the first unused contact for each track (row) starting with the last track in the table. The resulting assignments constitute the first hypothesis. If there are unused contacts remaining, they are saved as possible beginnings of new tracks. If there are insufficient contacts, then one or more of the tracks must have received a detection miss. The technique for generating the hypotheses in these cases is discussed in paragraph 2.5.8. Successive hypotheses are generated by incrementing the position of the pointer for the first track to the next unused contacts. When the last con- tact is used for the first track, the position of the pointer for the second track is incremented once and the process repeated for the first track. After all contacts have been used for the second track, the process is repeated for the third track and so on for the remaining tracks. The overall process can be compared to an odometer with each track representing one wheel of the odometer. As the last digit (contact) is used on one wheel (track) of the odometer, the next wheel is incremented one or more digits. Methods are included to keep from generating illegal or inconsistent hypotheses (i.e., ones in which a contact is used more than once). 2.5.7 TRACKER A basic part of the correlator is the tracker that is described in Reference (b). The tracker is used to give a figure of merit for relating a contact report to a track. The state equations of the tracker are outlined below. If (LK, 1 l) are the latitude and longitude in radians of a ship contact at the time of the (K-l)th observation, then the coordinates of the ship, at the time of the K-th observation tK hours later, will be given approximately by VNK LK = LK-1 + R tK + WL,e EK + H = MK-1 + R cos L K + WK e K-l where R is the radius of a spherical earth, VNK is the average velocity of thee ship North, VEK is the average velocity East, and wL and wM are noise terms. In 1382 addition, it is assumed that the rate of change of the latitude and longitude (here called AK and DK) are vNK VEK PK R cosL w,e LK-1 where wA and we are velocity noise terms that allow for random minor deviations in velocity and for maneuvers. If we define 0 tK 0 MK 01 0 tK w =XK ' K = ' WK-1 = ;KO O 1 ) w VNK VEK K-1 R and K-1 Re cos LKe e K-! then these may be written as XK - K XK-1 + WK-l' It is assumed that the noise terms are white and Gaussian with zero mean and covariance matrix QK. 183 2.5.8 EVALUATION OF HYPOTHESES If X is the state vector of a track and X (+) is the updated estimate at time K, then the projected estimate of the state at time K + I is XK+1(-) = K XK(+) where 0K is the transition matrix. When the targets are not maneuvering or maneuvering only slightly, the measurement errors can be approximated as Gaussian so that the probability density, f(XK) , is effectively Gaussian. This allows the Kalman filter projection routines in the tracker to be used, which then can give figures of merit for associating reports to track. Techniques showing how to handle large maneuvers are dis6ussed in paragraphs 2.4 and 2.5.9. The uncertainty in the estimated position given by the projection routines of the Kalman tracker is represented by the covariance matrix P(-). The distribution of XK is given by " !(X ( - f(XK) = 21 I /2 exp 2 (XK ) P(- ) ( where MK = E(XK). If there are n tracks being kept and m contact reports, ZK received at time K, with m < n, then fl (Z K1) ' f2 (Z K2) f (ZKm) is the likelihood function for the hypothesis that, for all i, the report ZKi should be associated with the track whose projected uncertainty is given by fi' The likelihood of this particular assignment or hypothesis is 184 m Lj = log ir fi (ZKi) i=l and the assignment that produces the maximum Lj is called the maximum likelihood assignment or hypothesis. If we define ZKi Z Ki- xi( then 1 n TL -n An 2 2T -2ZPn C PKi + Ki 2 i=l Ki When reports are missing (i.e., when there are less reports than tracks), then each hypothesis has one (or more) fewer terms in it. However, no one hypothesis is favored more than another since each will contain the same number of "zeroes." The last equation is used to evaluate each hypothesis with a certain number of the best ones retained. When only the single best hypothesis is kept, it will give the preferred tracks at that time. 2.5.9 INITIATION In the ATAC model, track initiation is accomplished with the aid of the Riceian distribution [References (g) and (h)]. This distribution is the result when an object, whose initial position has a circular normal distribution, moves with a constant speed v in an unknown direction for a time t. All direc- tions are assumed to be equally likely. Thus, if the distribution of the initial position is given by f(r) = 1 exp - r2/2c2 27L then the distribution of the final position after a time t is given by f(r,t) = exp (r2 + u2 t2)/22] IO rt 185 thwhere Io is the ordinary Bessel function of zero order. For a given time interval t, the distribution of target positions has spread outward into an annular region. It is not only at the beginning of the simulation that tracks must be initialized. Even after tracks have been established, there are new ships entering the area and, therefore, tracks to be initialized. In either case, the log of f(r,t) is used for each hypothesis that assigns a second report to a point saved from a previous set of sensor reports. The distance between two such reports is r in the last equation. Speeds of 12 and 24 knots were assumed for two type ships, slow and fast, as described in paragraph 2.4. The value of a assumed is 2 2 2 2a = a +a t + + 1 v 2 where 21 = variance of the first position measurement, 2a, = variance of the second position measurement, a = variance of the speed estimate, andv t = time between the measurements. Besides its use for track initiation, the Riceian distribution is also used for maneuvering targets in the ATAC model. If a particular track is not updated in one of the first two motion models (non-maneuvering) described in paragraph 2.4, then the Riceian distribution is applied again from the last point of the track to see if possible contacts exist outside the normal projected uncertainty ellipse that could be associated with the track. If there are no such contacts, then the track is given a miss for this reporting time. If there is at least one such contact, then the one that minimizes the likelihood equation from the previous section is chosen to update the track. In effect, the track is reinitialized from the last point. As before, two speeds (12 and 24 knots) are assumed for the two maneuvering motion models. 2.6 MEASURES OF EFFECTIVENESS AND SCORING The technique used to establish requirements for the ocean sur- veillance parameters in Reference (a) is similar to that used in other require- ments analysis. First, a nominal case value is chosen for each of the parameters in the problem. Then, a sensitivity analysis is done by varying the value of each parameter, one at a time, over a specified range and calculating certain measure of effectiveness (MOEs). 186 2.6.1 TRACK PURITY This MOE is intended to measure how pure a track is in the sense of the track having the same ID number for each contact report. It is the number of previous time steps that a ships' identity number agrees with the current identity divided by the number of time steps up to and including the current one. 2.6.2 MEAN RADIAL PREDICTION ERROR At each time step, the tracks are projected 1 hour ahead, The predicted position is then compared with the ground truth position at the future time and the error between them calculated. 3. RESULTS OF THE PARA1ETER SENSITIVITY ANALYSIS The results of the ATAC Parameter Sensitivity Analysis are presented in the following graphs of the MOE as a function of the parameter under considera- tion. The MOE being considered can be plotted against only one of the parameters at a time; therefore, it is necessary to select a set of nominal values for the parameters. The selected nominal values are listed in Table 3-1. In Figure 3-1, the track purity index decreases rapidly. When the active sensor is aided by the passive sensor the situation is much improved. When the radar sensor update interval is increased to larger values the result is few reporting times in the overall period. Consequently, there are propor- tionately more passive sensor reports which provide identification of the high interest ships. Since the passive sensor reports are correlated with few errors, there is less chance for error overall and the purity index remains high. The prediction error 1 hour in the future from the sensor reporting times is plotted in Figure 3-2. In general, the errors for the background ships increase as the update interval increases while that for the high interest ships fluctuates around a constant value. Some of the fluctuations are due to the coincidence of the radar sensor observation times and the high interest ship maneuver times. When the passive sensor is added, the result is that the error for the high interest ships is not very dependent on the update interval. In Figures 3-3 and 3-4 the radar sensor accuracy is varied. The main characteristic of these curves is the improved track purity for the high interest ships when the passive sensor is used. 187 r* * OC)r 0) 4. > E4 E4 c > r-4 C) $ $4 CO ,1) r4 ro o 44 , E-- P IJ , o Ci r4 00 d C.o Oo ,-O *,-4 O O c - I ,.4 r-4 W Co cv 4i Z 0 (1.0o 0 . o b o uO CO *: 4 P 4 . r o a4 a0 *) o4 z o . o hi i I ~ P 00 P " 4J .f o i ' -4- u) HU0 0 0 a 4 C)~I ~ C O O a ) O C 0 1 * 0 0 1.4 " n r- Z .Cu -H 4rU-I,! P r CZ 4/ cali O r I I I C O I Sw tl zC u H I0 $ ,4C4 v.A - ? CY ? q r. 4 4 $ . P 1 P44 Pq 0 C , 0l ,-) $4 .. Z Im C) I I$ 4J r-u " $4 .1 4 U) i-4 ., 1 U) i, , r, Cu c, C ]u 0, $ 4. CO4a) ) 0 d- 0 Cu Ci r-4 P.4 w Cu ~ COO R C 7 C r 4 H d) a H C ) 0 4 c<3 Cb : 0 E to Cu < Cu C t 4-P44 Cu Co a a bO rc vl ~1Coc b4 0g04 .7 4$ $4 $4 U)c ~ m U) CO U) P4 $>) 1.0 ? SHIPS Lu O HIGH INTEaESTSHIPS 0.2 .-- -- RADAR SENSOR MESURING PON ONLY. - -L)- -- PASSIVE SENSOR ADDOED TO RAOAR SENSOR. 0.5 1.0 1.5 2.0 2.5 3.0 5 4.,0 AT 1 . RPDAR SENSCIR UPDATE INTERVAL, HRS Figure 3-1 Track Purity vs Radar Sensor Update Interval 25 - a r IGH INTERAST SHIPS -5- ::= 10~...RADA/ SENSOR MEASURING POSITION ONLY. -'-Q- -- PASSIVE SENSOR ADDED TO RADAR SENSOR. 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 LTU, RADAR SENSOR UPDATE INTERVAL., HRS Figure 3-2 Prediction Error vs Radar Sensor Update Interval 1 89 I.- -. S0. < t;: 0.4 - HIGH INTEREST SHIPS 0.2 -. 4-~---- RADAR SENSOR MEASURING POSITION ONLY. - , - PASSIVE SENSOR ADDED TO RADAR SENSOR. 1.0 2.0 3.0 4.0 5.0 *nl' RADA.R SENSOR ERROR IN NMI Figure 3-3 Track Purity vs Radar Sensor Accuracy 25_ X 20- ' ;4Z \ INTEREST SHIPS Z 15 C -J 10 -, ALL SHIPS E RADAR SENSOR MEASURING POSITION ONLY. --- FJ--- PASSIVE SENSOR ADDED TO RADAR SENSOR. 0 1.0 2.0 3.0 4.0 5.0 on1, RADAR SENSOR ERROR IN NMI Figure 3-4 Prediction Error vs Radar Sensor Accuracy 190 4. CONCLUSIONS The correlation approach used here was adequate to handle the size problem analyzed in determining ocean surveillance requirements. However, it was necessary to incorporate some threshold reduction techniques to keep the number of hypotheses from getting too large. This is done by limiting the size of the uncertainty areas. The costs incurred are more errors in correlation. The ATAC model could be extended to include a weapon targeting model that simulates the launching of OTH weapons. It could then be used as a tool for defining and developing improvements to the information col- lection, processing, and distribution portions of the U.S. Navy's OTH-T capability. Previously generated requirements could also be compared with Fleet exercise data in the model. 191 Section 5 REEREdECES (a) The Johns Hopkins University Applied Physics Laboratory Secret report FS-80-076, dated March 1980, "OrI/DC&T Engineering Analysis, Volume 9, System Requirements (U)." (b) The Johns Hopkins University Applied Physics Laboratory Unclassi- fied Report FS-79-032, dated March 1979, "OTH/DC&T Engineering Analysis, Volume 5, A Kalman Filter Rhumb-Line Ship-Tracking Algorithm," by T. G. Bugenhagen and L. B. Carpenter. (c) The Johns Hopkins University Applied Physics Laboratory Confident- ial report FS-79-276, dated December 1979, "OTH/DC&T Engineering Analysis, Volume 8, Evaluation of Surface Ship Tracking Algorithms (U)," by D. E. Corman. (d) The Johns Hopkins University Applied Physics Laboratory Unclassi- fied report CIA 1597, dated August 1.980, "Computer Programs for Area Tracking and Correlation Model." (e) Culbertson, J. S., Unpublished notes, Preliminary Draft of Back- ground Ship Motion Model, dated 5 July 1978. (f) The Johns Hopkins University Applied Physics Laboratory Secret report FS-77-118, dated September 1977, "Clipper Bow for Targeting Anti-Ship Cruise Missiles (U)," by T. G. Bugenhagen, B. Bundsen, D. V. Kalbaugh, J. H. Walker, and M. B. Williams. (g) Rice, S. O., "Mathematical Analysis of Random Noise," Bell System Technical Journal, 1945-46, Volume 23, pp. 282-332 and Volume 24, pp. 46-156. (h) Office of the Chief of Naval Operations, Washington, D.C., B. O. Koopman, OEG Report No. 56, 1946, "Search and Screening." 192 AN OTH SURVEILLANCE CONCEPT Leslie C. Kramer Nils R. SandeZZ, Jr. ALPHATECH, INC. 3 New EngZand Executive Park Burlington, Mass. 0803 (This work was supported by the Office of NavaZ Research and the NavaZ Ocean System Center under contract number NOOOZ4-80-C-0309). 193 Presented at the Fourth MIT/ONR Workshop on Command and Control, San Diego, CA. June 1981. AN OTH SURVEILLANCE CONCEPT* by Dr. Leslie C. Kramer Dr. Nils R. Sandell, Jr. ALPHATECH, INC. 3 New England Executive Park Burlington, MA 01803 1. INTRODUCTION In this paper, we describe a brief study into techniques for detecting and providing warning of low-altitude airborne threats approaching United States Marine Corps units in the field. Such threats are difficult to detect since they can be masked by terrain features until quite close if sensors located within or above U.S.-controlled territory are used. If sensor alti- tude is increased to look over terrain features, the sensor's vulnerability is increased and its deployment typically becomes quite expensive. The alter- native approach, viz., deploying sensors forward into enemy-controlled terri- tory, is promising but difficult from the viewpoints of delivery, sensor vulnerability, and information retrieval. The purpose of the research reported here was to examine these issues in light of modern technology and its probable advancement through the late 1980s and early 1990s. Three objectives for this work were defined: 1. To identify promising sensors and sensor systems which could provide USMC forces with adequate warning of impending attack by low-flying aircraft. 2. To characterize the capabilities and the relative advantages and disadvantages of alternative concepts. 3. To interpret the results of this work in terms of its implications for future research directions. We will show in this paper how the ground rules and scenario defined prior to beginning this work, coupled with the characteristics of the classes of sensors physically feasible, clearly lead to a preferred approach to solving the problem. *This work was supported by the Office of Naval Research and the Naval Ocean System Center under contract number N00014-80-C-0309. 194 2. SYSTEM MISSION AND SCENARIO The mission of the over-the-horizon (OTH) surveillance concept sought was to detect, to localize, to count, and to identify the threat in a timely man- ner. More specifically, the surveillance system should: 1. Provide 5 to 10 minutes warning of approaching threats of all types: fixed- and rotary-wing, subsonic and supersonic speeds. 2. Count threat aircraft at a coarse level, i.e., differentiate between a few (1 or 2) and many (10 to 20 or more). 3. Locate the approach sector to about 10? relative to a reference point (e.g., an antiaircraft battery). 4. Identify the threat as to aircraft type, if possible. Specific threat aircraft were not identified, in fact, results were developed parametrically versus target speed, which was the only number of importance given the level of detail of this analysis. For the sake of evaluating parti- cular cases when desired, the following speeds were adopted as typical and nonspecific: * Helicopters: 130 kt = 70 m/s * Subsonic Airplanes: 500 kt = 260 m/s = M 0.8 * Supersonic Airplanes: 975 kt = 500 m/s = M 1.5 Appropriate bands about these speeds were then considered when needed. Friendly and hostile forces were visualized as residing on opposide sides of the forward edge of the battle area (FEBA). The friendly territory was defined to be 20 km wide and 20 km deep. The hostile territory was defined to also be 20 km deep but somewhat wider (30 km) to provide some latitude for the threat approach azimuth. We found that restricting attention to an area this size made it impossible to satisfy the warning time requirement in all situ- ations, so again a parametric approach was adopted as we illustrate below. 3. METHODOLOGY The purposes of this research were to quickly identify promising OTH detection concepts suitable for USMC applications and to recommend appropriate further effort. These purposes would not be served by detailed consideration of the multitude of physically realizable sensors. Rather, the approach adopt- ed was to consider generic sensor characteristics (such as range, target handling capability, data rate, etc.) in order to determine what characteris- tics were required to meet the surveillance objectives defined for this study. With these characteristics identified, particular physical sensors or classes 195 of sensors could then be considered to determine how to best meet the mission objectives. This is not to say that existing or projected physical sensors were ignored as requirements were developed. In this regard, we are particu- larly indebted to individuals at ERADCOM and MERADCOM who provided valuable information regarding present surveillance concepts. The various means shown in Table 1 for detecting aircraft were considered, and the sensors associated with these means of detection were considered from the viewpoint of the characteristics listed in Table 2. The operational envi- ronment faced by the USMC was a very important consideration at all times. The USMC mission environment cannot be overemphasized as a factor determining the types of equipment suitable for Corps use. An equally important factor is the USMC system environment; any novel system proposed must operate compatibly with other USMC assets, either existing or projected. In the remainder of this paper, we will present selected technical results developed in the course of this investigation, and we will describe the sur- veillance concept recommended for further consideration. Additional detail may be found in reference [1]. 4. DETECTION AND WARNING TIME In order to minimize the delay between the time when the threat enters the surveillance system coverage and actual provision of warning to a decision- maker capable of responding, one must understand the source of the delays. Figure 1 illustrates the situation. The warning time is determined by the time it takes the threat to fly from where it is when you find out about it to where you are (line segment B in the figure). The time lost, so to speak, is the time that elapses from the instant the target emits an observable with- in surveillance range to the time when you are notified (line segment A). The components of this delay (i.e., the lost time) are observable propagation time and the time needed for signal processing, data processing, data intrepretation, and data dissemination. We are interested in determining when various con- stituents of this delay are dominant. For electromagnetic observables, the propagation time is essentially zero; the observable travels at 3 x 108 m/s. For seismic and acoustic observables, the speed is much slower: 1,000 to 2,000 m/s for seismic (under certain con- ditions, seismic signals can propagate as slowly as 200 m/s) or 300 to 350 m/s for acoustic. If one considers a typical maximum range for acousto-seismic sensors, say 10 km, then the observable propagation time could be as long as 5 to 30 seconds. While appreciable, this is a relatively small part of the minimum warning time requirement. How long should the data handling take? Electronic signal and data pro- cessing should (with advanced technology) only require a small fraction of the 5- to 10-minute warning desired. We conclude that only one situation arises in which the warning differs significantly from the threat travel time from entry point into the surveillance coverage to the defended location: the situation in which human data handling is carried out in a system employing 196 (D~~ c~aM~( ( C (2 La D C H. 0n Di a C? MD I-' (2 CD ~ ~ C CD( I- Di C? >9 5 CD Ht D~i 10 t"(2 H- (3 (2 ('3 t~ (3 H- CD) H- c D rC Z 0 En 5 - (2 k~ (2 ~ CD ~Q CD H- - ? a 0 ?< CD:aC IQ C 0 CD H-H-CD H- ~ 11 0 (2 C . - D: (2rt.I i CD C 0 C? DD (n H- CD CD 0 li H D i- CD 0 rt- 0H-< - ' I- :3 CD 1 Di 0 c- I-' -0 D t -. C H - ( (C? Z N D P (D 0 C (1 C C 0H H c < - H- ( D H- ' ~ . - O H- n " = H- H- o P a 3 ct r r ~ ;'C? ' rC D CDCr r 0 - H D a< Di a < Ce = Ct td CH- CDD z < - H' H- 11 -~~~~ CD 4 (n Di w i -? t O 0! CD I-I C? Ii 01 ( t D9 H- ~ H- H- rt -D = Di D Ct t-1 C 0 (2 0 cnz CD H' 4 CJ CD 40 z Di z 0'C ~~ CIL D 11 Di DiJi r?a C rt 0 . D CD C? 0 CD H- i -, --- H- ;~ :: I- =~ H' '-"aCD 0 H- rt CD H ~~~~~~~~~~~~ i i u 3 ? : P ~'CD ~~ CD t - H ' r cn ( c t = l< n- CD H-t H- ( ?- CD C19'I ii CD 2 CD, 3 (f 'CD - ~~~~~1 HC?( CD Di -H- 0 If~ ~~( C? C? CD CD Cl ) Ij~ ~~( CD - - ;a P, tl a ~~~~~~~~~~~~~n I H~~~I- 3 ~ ~ D IA C/)tarnii rn rr,~~19 long-range acoustic sensors. Thus, we prefer completely automatic data hand- ling and deployment of acoustic sensors (if used) close to the anticipated threat. The threat time-of-flight (TOF) is simply related to its speed as shown in Figure 2. The solid lines indicate the threat speed (plotted along the ordinate) which results in a 5-, 10-, or 20-minute transit time for a given range (plotted along the abscissa). Below eaqh solid line is a dotted line indicating the corresponding speed if a 100-second delay is introduced into the acquisition cycle. Higher speeds mean shorter warning, of course. One sees at once that for 5 to 10 minutes of warning, high-speed threats must be detected well before they enter the 20- to 40-km deep region of interest de- scribed in Section 2. Only helicopters can be detected in time within this region. The implication is that either forward sensor deployment, long sensor range (as much as 100 to 200 km in some cases), or both is necessary to ade- quately warn of high-speed threats. 5. RESOLUTION AND ACCURACY Spatial resolution refers to the sensor system's ability to determine that two objects located close together are in fact distinct. Accuracy refers to the system's ability to locate a single object along various measurement "directions" with small error. We quote the word "directions" here because we mean to consider measurements such as angle or range rate as well as up, down, left, or right. Resolution and accuracy are typically related since if an object can be described as occupying a given resolution cell, that is, a box in measurement space of size such that objects outside it can be distinguished from the one inside, the object's position is known to the cell size at worst. In many practical situations, the error will actually be a small fraction of the re- solution cell size. For example, radar accuracy in range, angle, or doppler (i.e., range rate) is frequently described by an equation of the form a= K (1) where a is the error (the standard deviation of the random process represent- ing the measurement), A is the resolution, K is a factor near unity, and SNR represents the signal-to-noise ratio [2]. Thus, for SNR equal to 10 to 100 (10 to 20 decibels), which represents typical "fair" to "good" measurements, the error a will be on the order of 1/3 to 1/10 of the resolution A. Much finer range, angle, or doppler "splitting" is physically feasible at even higher values of SNR. We wish to consider whether resolution or accuracy is the driving design consideration for the systems and mission under study. We will show that in general, a sensor with adequate resolution to allow coarse threat counting 198 will automatically have adequate accuracy to define the threat's position and heading as required. This conclusion follows from an assumption that the surveillance system would be adequately able to count the threat if it could resolve aircraft separated by 100 m or so. This would also permit the system to adequately judge the size of a formation in the event that many aircraft were flying together with less than 100 m separating adjacent ones. We will consider below the implications of this assumption. The requirement that the threat be located to within a 10? sector can be interpreted several ways. In the previous section, we showed that the threat must be detected several tens of kilometers from its target if adequate warn- ing is to be provided; a 100 sector will be many kilometers wide at these detection ranges. Thus, if the sensor system has spatial resolution on the order of hundreds of meters, it will certainly be capable of localizing the threat to within a 100 sector. An alternative interpretation of the 100 requirement, based on threat velocity, is that the system be able to determine the threat's heading to 10?. A simplified means of analyzing this question is indicated in Figure 3. The analysis is based upon assuming that the system computes velocity from finite differences of position. A realistic system would use a more sophisticated approach, thus the results we derive below are conservative: they upper bound the likely course direction estimation error. Suppose position measurements are made of the threat at two locations separated by a distance d as shown in Figure 3. Then the (straight-line) course of the threat must pass through the two error volumes centered at the position measurements as the figure indicates. The angular error in the course direction can be related to the position measurement error in the cross-range direction as shown in the graph, which is a plot of the equation y = 2 *- d 2 2 where y is the cross-range error and 8/2 is the cone half-angle. The approxi- mation tan(e/2) 3 9/2 is good to 1 percent for e less than 200. One can see that if the threat is observed over a range band of 1 to 5 km (i.e., d = 1 to 5 km) a cross-range error of about 90 to 450 m is adequate. If cross-range resolution is on the order of 100 m to permit threat counting as suggested earlier, the cross-range error will typically be on the order of 10 m (one- tenth the resolution) as we indicated earlier. Thus we again conclude that resolution adequate for counting the threat will automatically result in accuracy adequate for determining its angular sector. Having concluded that requiring 100-m spatial resolution will result in both adequate resolution and adequate accuracy, we consider the implications of requiring resolution on this order. We will focus on determining the angle resolution provided by typical sensors of different types in order to identify those approches which appear to match the needs of an OTH surveillance system. Attention will be restricted to angle resolution because range resolution of 199 the order required is generally easy to obtain from a technical viewpoint if sensors are used which provide range resolution at all (i.e., active radar). Note, however, that we have not concluded that range resolution is essential. The angle resolution A associated with a physical sensor having aperture dimension s and operating at a wavelength A is generally given by an equation of the form A = K (3) where the angle A is measured in radians, A and s are measured in common units, and K is a factor near unity which depends on several considerations, for example, the aperture shape. By solving this equation for s as a function of X and A, taking K = 1, and combining with the equation y RA (4) to obtain the cross-range resolution in meters (y) in terms of the angular resolution in radians (A) at range R expressed in meters, one can develop Figure 4. Here curves of constant cross-range resolution are shown in the plane described by range-to-target along the abscissa and aperture size along the ordinate. Several ordinate scales are shown, corresponding to several typical sensor wavelengths: IR (10 microns), millimeter wave radar (X = 3mm, corresponding to 100 GHz frequency), microwave radar (A = 10 cm, corresponding to 3 GHz or S-Band), and an acoustic sensor operating at 100 Hz. If typical operating ranges for various sensors are selected as well as typical aperture sizes, one sees that the cross-range resolution requirement of about 100 m is met for alternative sensors as shown in the figure. 6. SENSOR TYPES AND CHARACTERISTICS In this section we discuss general characteristics of several sensor classes as applied to the OTH surveillance mission. In many cases, we make assertions about sensors which we do not support by detailed calculations or references yet which may be controversial if taken out of context (e.g., "long- range radars tend to be expensive"). We believe that within the context of the discussion our assertions are reasonable, and that taken overall our ulti- mate recommendations given below are justified. RADAR Radar sensors generally provide excellent quality data at a relatively high "generalized cost." By this we mean that radars can provide functional characteristics of the type listed at the top of Table 2 at a level that far exceeds the requirements of the mission defined in Section 2, but that several disadvantages accrue relative to some of their physical and operational char- acteristics as listed at the bottom of the table. 200 Radars can be built using modern technology that have long range and fine resolution. Accuracy can be quite high. Many targets can be processed simul- taneously. Signal processing technology is well understood and reasonably easy to implement. The physics of radar signal propagation are well understood and system performance can be predicted with some confidence. These characteristics imply that radars are excellent sensors for some applications. We argue, how- ever, that they are not very well matched to the low-altitude aircraft detec- tion mission. Several factors lead to this assertion. First, radars are vulnerable to countermeasures because they are active: they broadcast in order to receive. This makes radars vulnerable to anti- radiation missiles which home on their emissions. Also, adaptive jammers can be built which listen to radar emissions and adjust their interference output for maximal effect. By use of sophisticated signals and the associated signal processing technology, these vulnerabilities can be minimized. Such measures increase the cost of the radar, however. Second, the apparent long range available with radar sensors is illusory for this mission due to the effect of terrain masking. Low altitude air threats will be hidden from view by hills if the radar is at low altitude over friendly territory. If it is at high altitude or deployed forward over enemy territory, vulnerability to attack is increased and deployment cost is in- creased as well. The third and final factor to consider is that radars tend to be expensive in general terms. If only a few are deployed to fulfill a mission, they must have long range (and thus large power-aperature product) for broad coverage. Also, an ability to handle many targets is needed. If many short-range sensors are deployed to fulfill a mission, the number required will result in a large total cost even if unit cost is modest. PASSIVE OPTICS The obvious advantages of passive optical sensors are that they can be quite small and light physically and that they can provide very high quality data (in the sense of fine spatial resolution). Their small size and weight make them well suited to applications requiring high mobility, however they require careful design if they are to be rugged and reliable. Their primary disadvantage is relatively short range in bad weather or when faced with op- tical countermeasures such as smoke. Also, search is difficult due to their typically narrow field of view. We examined optical sensors (both active and passive) at some length during this study, primarily from the viewpoint of sensitivity calculations. The goal was to determine the current state-of-the-art; the details are col- lected in [1]. While we are able to make first-order judgments regarding the range reasonably expected under various circumstances, we could not develop definitive results without a more detailed threat signature character- ization than that available to us. Our results are therefore largely para- metric, as shown in Figure 5. Here we plot the noise-equivalent temperature difference (NETD) at various ranges for a typical passive IR sensor with various values of atmospheric attenuation. The NETD indicates the temperature 201 difference between object and scenic background which results in a signal just at the noise level; for a signal-to-noise ratio of x, the object-minus-back- ground temperature difference must be x times NETD. The sensor parameters are given in [l]. They are not meant to represent any particular current sensor; they are, however, typical. We assume a sensor operating in the 8 to 14 micron waveband with a 3000 K background. The figure shows that an operating range of up to 5 km or so is reasonable to expect for these sensors. ACOUSTO-SEISMIC SENSORS Acousto-seismic sensors were discussed extensively with individuals at Lincoln Laboratory, ERADCOM, and MERADCOM. We found that the potential dis- advantages of acousto-seismic surveillance are limited range, a poorly under- stood signal environment, and an immature system technology. The consensus we observed was that the technology of acousto-seismic (i.e., mechanical-wave or MW) sensors and sensor systems is in its early stages of development but that this technology is very promising. The real thrust of the research in the MW arena is the analysis of system questions rather than transducer ques- tions. Very sensitive microphones and seismic transducers are already avail- able; the real issue is how to process the resulting signals and data in order to associate the measurements with particular physical emitters in a multi- target environment and to extract the information about them which is desired. The range expected of MW sensors is not yet clear. Several individuals interviewed quoted maximum ranges of 10 km or so for both acoustic and seismic detection of single aircraft under quiet, well-understood measurement condi- tions. Under realistic battle conditions, one would expect the operating range of MW sensors to be shorter. On the other hand, their cost may be low enough and their size and weight may be low enough that large numbers could be delivered to cover the required surveillance zone, even with maximum range of 5 km or so. A more serious question surrounds the utilization of MW sensors in a complicated, noisy battle environment. This environment requires very sophisticated signal and data processing and interpretation. Compounding this problem is the fact that the signal propagation physics for MW sensors is not well understood (compared, for example, with electromagnetic signals) so that the "known" signal being sought in the noise is not very well known at all. In contrast to the disadvantages faced today by MW surveillance systems, there is a potentially huge payoff in mission utility if the problems are successfully resolved. The components can be small and inexpensive. They can be put where needed by several means. Their capabilities (e.g., in terms of spatial resolution) are well matched to the requirements for this mission, and their physical characteristics (e.g., size, weight, ruggedness) are well matched to the needs of a highly mobile USMC force. For these reasons, we recommend a more extensive investigation to define a system using MW sensors as the primary threat detection mechanism and to identify and catalog the issues associated with it. 202 7. CONCLUSIONS AND RECOMMENDATIONS As a result of our studies, we recommend that an OTH surveillance approach encompassing two classes of sensors be considered: numerous hearing-like sensors ("ears") for fulfilling functions requiring only coarse spatial re- solution coupled with a much small number of vision-like sensors ("eyes") which would carry out tasks requiring high resolution. Our immediate recommendation is that this concept, which we describe more fully below, be defined in suf- ficient detail that the USMC can make a well-informed judgment as to whether or not to develop some or all of it. The concept we propose is described as follows: We envision the ears to be acousto-seismic sensors deployed beyond the FEBA. These would be delivered by artillery, by remotely piloted vehicles (RPVs),or by manned aircraft.* In the last case, sensor delivery would be a new task assigned to existing air assets; no new aircraft can be dedicated to this task, and in fact, a delivery mechanism which does not interfere with existing tasks would be required. The information collected by these sensors would be relayed back to the USMC C3 system via smart repeaters: repeaters with a certain amount of internal data processing capability. These would reside on both sides of the FEBA as required. The eyes would be RPV-borne infrared (IR) sensors. (Millimeter-wave radar sensors are a possible alternative.) For most situations, these sensors would have to be on station when required; the time-to-station is typically excessive if one contemplates launching these vehicles on demand. Often, a logical loca- tion for prestationing such sensors is obvious, e.g., near an enemy airfield or along an important approach route. In operation, the system we propose would behave much as a human does: the ears cue the eyes. One hears a noise, which results in a head turn to identify the cause in the event that further data is needed. An important aspect of the two-sensor-class system we envision is a nonheirarchic, distributed information processing and dissemination mechanism. Such a mechanism is shown pictorially in Figure 6. The architecture depicted allows data fusion (i.e., the combination and interpretation of data from several sensors or transducers) at several levels, with the concomitant feature that information can be identified and routed to relevant users as soon as it is derived from data, thereby circumventing further delays in the system. We believe that this system structure offers several advantages. It is clear, for example, that the USMC contains a heirarchy of decisionmakers whose information needs vary over a wide span. The individual in charge of a single antiaircraft battery is concerned with a much smaller part of the threat than the officer responsible for the entire task force. A surveillance system which can simultaneously fill the needs of decisionmakers at each end of this spectrum obviously offers distinct advantages. Benefits accrue in terms of reduced reaction time, for example, if warning information can be provided to the force capable of responding to it without passing through *In an advanced scenario, battlefield robots could be used for sensor emplacement. 203 intermediate stages of handling while simultaneously presenting higher level echelons of command with suitably condensed or abstracted reports of the threat and the reaction to it. The architecture we suggest can accomplish this. Developing in detail the means to accomplish it, both from a hardware and a software viewpoint, is a present focus of C3 research. A major advantage of the concept we propose is that it is a natural ex- tension of the already existing Remote Battlefield Sensor System (REMBASS) which is being readied for deployment by the U.S. Army and the USMC in the early 1980s [3]. The approach to aircraft detection we propose would add new elements to the mission for which REMBASS is designed, however we believe that many aspects of the research and development which have gone into REMBASS and its predecessor systems bear on the OTH aircraft detection problem. The surveillance concept description just presented is obviously quite sketchy. We believe the concept is sound. Nevertheless, many issues come to mind immediately. We recommend that these issues be addressed in subsequent work. Among these issues are questions related to utility, operations, phen- omenology, and architecture. By utilitarian issues, we mean to ask: does the system tell you what you want to know, when and where you want to know it. To resolve this, further work needs to be done to define the threat and to determine precisely what information is needed by different USMC decisionmakers. By operational issues, we mean to ask: what sort of components make sense from a USMC operational viewpoint, particularly with regard to sensor delivery, to system mobility, and to interface with the existing and planned USMC C3 network. By phenomenological issues, we mean to ask: What do the targets of interest look like and sound like to the sensors we propose. Some data was examined in the course of this research. It appears that, for example, acousto-seismic data has very high potential for providing target identifica- tion, however a great deal of processing is required to extract the information. Defining this information extraction process is a major focus of several re- search projects. Many system issues (as compared to one-target, one-sensor issues) such as data association among targets and measurement correlation among transducers remain to be resolved. By architectural issues, we mean to ask: how should the information processing be distributed. What sorts of "computers" are required at what physical and functional locations within the system? Is packet radio the proper technology for information retrieval? The issues described above are relevant to USMC surveillance system research and development in general as well as to any future consideration of the particular surveillance concept we suggest. Whether or not our con- cept is pursued further, these questions should continue to be addressed in fundamental terms. We stress in particular the importance of further research into acousto-seismic approaches to surveillance. This technology is poorly understood at present from the system viewpoint, as we have said, yet it offers great potential for fulfilling a critical need for our military forces. 204 THREAT LOCATION WHEN THREAT LOCATION WHEN DEFENSE NOTIFIED OBSERVABLE EMITTED2 B A FEBA v ~/ > USMC FORWARD ASSET SENSOR PROCESS INTERPRET DISSEMINATE Figure 1. Determinants of Warning Time. 1000 '-SUPERSONIC SOUND 300. -. SUBSONIC S MIN / / 100 ) -Lw SIN 20 MI / / / 30 / 1/ 3 10 30 100 150 RA2NGE (km) FROM THiREAT TO DEFENDED RUNIT Figure 2. Maximum Warning Time Equals Threat Flight Time. 205 04 Uj ii (SW ct) IX ,I I Xl { I WO sos WWOE uol uzu~,o ~~2C: .4 '" A ao (Lu) 0'dV-S. . 206 1000 PATH LOSS o2 I . (dB/KM) 100- 0 5 10 15 RANGE (KM) Figure 5. Passive IR. "EARS" SMART I0 I% REPEATERS / I : I - LOCAL THREAT PROCESSING REGIONAL FINALHHREAT PROCESSING PROCESSING "EYES" USERS CUEING HIGH QUALITY DATA Figure 6. A System Architecture. 207 REFERENCES 1. Kramer, L.C. and N.R. Sandell, Jr., "Over-the-Horizon Detection Concepts," Technical Report TR-112, ALPHATECH, Inc., Burlington, MA., Dec. 15, 1980. 2. Skolnik, M.I., ed., Radar Handbook, McGraw-Hill, New York, 1980. 3. Sundaram, G.S., "REMBASS: The Army's New Battlefield Sensor System," International Defense Review, 4, 1980. 208 APPLICATION OF A. I. METHODOLOGIES TO OCEAN SURVEILLANCE PROBLEM Leonard S. Gross MichaeZ S. Murphy Charles L. Morefield VERA C, INC. Z0975 Torreyana Road Suite 300 San Diego, CA 92121 209 APPLICATION OF A.I. METHODOLOGIES TO OCEAN SURVEILLANCE PROBLEM Dr. Leonard S. Gross Dr. Michael S. MurphyDr. Charles L. Morefield VERAC, Incorporated ABSTRACT This paper will describe a data fusion system which combines artificial intelligence metnods with formal Bayesian techniques. Traditional approaches to the multisensor, multitarget ocean surveillance tracking problem have produced only limited success. Computational limitations, lack of flexibility and responsiveness,limited user understanding of program processing, and the inability of an analyst to guide the system are typical shortcomings of these systems. Previous applications of A.I. have approached some of these difficulties by use of "expert system" technology. In these cases, correlation is invariably based on simple mathematics or heuristics which generally do not take full advantage of the significant body of formal mathematical techniques available in the area of decision theory. In principle, these mathematical techniques provide a firm basis for forming correlation decisions in a multisensor, multitargetsituation. In practice, however, the ocean surveillance environment is such that important parameters required for the Bayesian methodology vary widely with sensor mix or surveillance scenario in a manner that is often poorly understood or even totally unknown. Thus, dependance on such techniques alone often results in an inflexible system which is unresponsive to the realities of an ocean surveillance problem. For this reason, the approach we have taken in this work is to overlay an A.I.-based control structure on a set of Bayesian theoretic functional elements. In this way we hope to successfully tune the individual Bayesian techniques to the changing surveillance situation and to manage the ambiguities that necessarily arise when analytical decision procedures are applied to a real surveillance problem. In particular, our research has identified several A.I. constructs which, when combined with Bayesian decision methods, can result in improved data fusion system capabilities. These constructs include the following: 1) A functional organization such as that found in A.I. systems like the Hearsay II speech understanding program can make the system both understandable to the analyst and highly flexible. This flexibility is most evident in the system's ability to incorporate diverse expert knowledge, encompassing 210 both quantitative and heuristic information, over a range of timelines, specific to an variety of sensors or scenarios, and responsive to a high degree to analyst review and editing requirements. 2) Production rule techniques are useful in dealing with the fuzzy control rules necessitated by the sensor mix and target scenarios inherent in the ocean surveillance environment. These techniques permit easy specification, review and moaification of the rules that guide the Bayesian processing and tune it for operation over a spectrum of situations, ranging from the well-known and usual to the poorly understood and unexpected. 3) Search techniques developed in speech understanding and game theory are useful in guiding the generation, evaluation andselection criteria for investigating competing surveillance hypotheses. The diversity of search strategies embodied in these techniques permits flexible management of ambiguities in the various plausible surveillance pictures, and provides a mechanism for tuning the search tactics to those specific scenarios where they are most efficient. The system we have developed to exploit these techniques is illustratea in the figure. We have termed the system the Intelligent Correlation Agent (INCA), and have derived its architecture from that of the HEARSAY II software system. In this design, development of a surveillance product is accomplished by a cooperating set of KnowledgeSources (KS) consisting of production rules and procedural code. These KS's communicate via a global data structure called a blackboard. There are essentially three tiers of knowledge sources. The lowest level "atomic" are concerned with detailed mathematical or symbolic correlation processes such as track scoring or filtering. The next level performs an observation/evaluation of the correlation modules to determine their success is propagating data up the blackboard data structure. Finally a control tier, using the conclusions of evaluation modules and the entire blackboard (if necessary), can make global decisions on system operation so that the user inputted goal state is most quickly achieved. Goal satisfaction occurs in either a forward or backwara chaining mode. 211 "ATOMIC' CORRELATIONSUMMARY EXTRACTION KNOWLEDGEBLACKBOARD K.S.'s BLACKBOARD SOURCES TIER II ~ TIER IITIER ITIER III OTT PROD OTT OTT SURVEILLANCE t t FlHAL SUMMARY PRODUCTRP HYPOTHESES COMPLETE SUMIMARY HYPOTHESES EY GROUP TRACK SUMMARY GROUP TRACK TRACK SUMMARY TRACK (Feasble Extensions) TRACK SEGMENTS T R ACKSSEGMENT ~--- SEGMETSEGMENTS (Unexamined & Old SE ,_SUMMARY INCA) DATA STRUCTURE CONTROL MONITOR K.S.'s TIER I DBM DBM EXTRACTOR .CONTROLLER DBM INCA System Structure 212 REFERENCES (1) Pattern Directed Inference Systems, Waterman and Hayes-Roth, ed., Academic Press, 1978. (2) Organization of the Hearsay II Speech Understanding System, Lesser, et.a., IEEE Transactions on Acoustics, Speech and Signal Processing, Vol. ASSP-23, Number 1, February 1975. (3) Increasing Tree Search Efficiency for Constant Satisfaction Problems, Haralic and Elliot, Artificial Intelligence 14, (1980), pp. 263-313. (4) A Parallel Tree Search Method, Nakagawa and Saka, IJCAI 1979, pp. 628. (5) Application of A.I. Methods to Surveillance Data Fusion (U), Final Report(S), VERAC document control number 81:S-0037, Gross, Morefield, and Murphy. 213 214 A PLATFORM-TRACK ASSOCIATION PRODUCTION SUBSYSTEM Robin DiZZard NAVAL OCEAN SYSTEMS CENTER San Diego, CaZifornia 215 A PLATFORM-TRACK ASSOCIATION PRODUCTION SUBSYSTEM Robin A Dillard Naval Ocean Systems Center INTRODUCTION This paper describes a method of extending the capability of a production system applied to Tactical Situation Assessment (TSA) by adding a "package" of system-logic rules. (A "production" is an if-then-rule implemented in a "production system," a system also having a data base and a number of control mechanisms.) The implementation of these rules within such a production system was termed a Platform-Track Association Production Subsystem (PTAPS) [ 1 ]. The function of PTAPS is to perform much of the logical reasoning such as process-of-elimination reasoning, needed to match tracks to specific platforms. Proof-of-concept experiments with PTAPS rules were conducted in a modified version of STAMMER, a System for Tactical Assessment of Multisource Messages, Even Radar [2], [3]. STAMMER was developed to serve as a demonstration of the applicability of rule-based inference techniques to the TSA problem. A small, fast skeleton version of STAMMER was created for PTAPS experiments by stripping the original of its confidence mechanisms, explanation functions, and graphics interface. The experiments involed two basic scenarios: one concerned with the identification of submarines, and the other with the identification of members of a Soviet task group with the help of satellite reconnaissance data. The submarine scenario was later successfully run in STAMMER2 [4] and in Rosie-l, the first version of Rosie (A Rule-Oriented System for Implementing Expertise) implemented by The Rand Corporation [ 5]. It was concluded from these experiments with different production system structures that PTAPS rules should work in any system in which con- ventional TSA rules will work [6] . Since then, many of the rules were implemented in the second version of Rosie, under development by the Rand Corporation. This effort has been just one phase of a larger research effort to develop automated data-fusion techniques using artificial intelligence tools. The work has been performed at NOSC under the Command Control & Communications Systems Theory project, sponsored by NAVELEX Code 613. PTAPS OVERVIEW Many of the PTAPS rules have the sole function of-building into the data base an "intermediate framework" of membership files which permit, via other rules, chains of reasoning not otherwise possible. This framework includes many kinds of "track files" and "platform files." To become a member of some track file or platform file, a track or platform must satisfy the conditions of a certain membership rule, and a member is removed by another rule when the original conditions are no longer all satisified. (The term "member" has a special meaning in Rosie, so the concept of listing an element in a file was substituted for membership.) Of particular importance are "OR-files." The members of the OR-file of a platform are those tracks which have not been ruled out as the track of that platform. A platform is a member of a track's OR-file if that track has not been ruled out as a track of that platform. The platform-OR-file of an emission has, as members, platforms which have not been ruled out as the emitting platform. 216 It is assumed that no active track entered into the PTAPS data base is a false track, eg the result of a radar's false alarm. Also, no two active tracks can be the track of a single platform. The amount of time after contact is lost that must elapse before an active track is made an inactive track should depend on the situation and be specified by rules. A track file is "complete" if the system holds an active track for every platform in that region (of that category, type, or class, if a subset file), even if none is identified. As a result of high-altitude surveillance, for example, the surface track file of a region may temporarily be complete. A platform file is complete if every platform in that region (of that category, type, or class, if a subset file) is a member of the file. Note that members of a complete platform file need not actually be in the region. Most platform files will be complete if the region is enclosed (eg the Persian Gulf, Red Sea, Mediterranean Sea) and the entrance/exit areas are continually monitored. Another concept used in combination with file completeness is that of correspondence between platform files and track files. For example, the track file containing tracks of sur- face platforms has as its corresponding file the platform file of surface platforms thought to be in the region. A typical rule involving these concepts is: The OR-file of a track is com- plete if the track is a member of a track-file whose corresponding platform-file is complete. The OR-file of a surface track is complete, for example, if the file of surface platforms possibly in the region is complete. Examples of other rules are given below. - If the OR-file of track t is complete and its only member is platform p, then t is the track of p. - If t is the track of p, and track t' is not t, then t' is an impossible- track of p. - If t' is an impossible-track of p, then remove p from the OR-file of t' and remove t' from the OR-file of p. These rules are written in widely different forms for the different production systems in which they have been implemented. The last rule above is implemented in STAMMER2 as follows. ORFILEREDUC (CONDITIONS ((MEMBER RPF *P) (ORFILE *P *ORF) (MEMBER *ORF *TR) (IMPOSTRACK *P *TR) (ORFILE *TR *FRO)) ACTIONS ((MEMBER *ORF *TR) (MEMBER *FRO *P)) 217 CONF -1.0 PRINFORM "A track and a platform are removed from each others' OR-files if the track is found to be an impossible-track of that platform.") The same rule is implemented in Rosie in this English-like form: For each platform, for each track such that the OR-file of that platform does list that track, if that track is an impossible-track of that platform, deny the OR-file of that platform does list that track and the OR-file of thattrack does list that platform and send (return, _____ _ that track, "and", that platform, "have been deleted from each other's", "OR-files because", that track, "was found to be an impossible-track", "of", that platform, return). Examples of rules involving emissions are the following. - For each platform p possibly in the region of an emission, if platforms of p's class or type carry the emitter type, then p is a member of the platform-OR-file of that emissioni - [completeness rule and various removal rules] - If the platform-OR-file of an emission is complete and has only one member, then that platform is the emitting-platform of that emission. - If platform p is the emitting-platform oif an emission and track t is incon- sistent in bearing with that emission, then t is an impossible-track of p. Some of the rules needed to support the chains of logical reasoning in PTAPS are also individually useful in an unextended system, and some of these require routine but ex- tensive geometry calculations. Most of the latter were omitted from the experiments, and the data were obtained by other means. The geometry functions involved could be imple- mented without difficulty, but would increase execution time while not serving a purpose relative to the intent of the investigations. SCENARIOS The experiments involved two basic scenarios, one concerned with the identification of submarines and the other the identification of members of a Soviet task group with the help of satellite reconnaissance data. The two scenarios used in the recent experiments with Rosie are summarized below. Two-Sub Scenario: Only two submarines could be in the region - a Delta and an Echo II; and two subsurface tracks are reported. The acoustic signature of one track shows that it cannot be a Delta; therefore, it must be the Echo II, and the other track must be the Delta. UNREP Scenario: Recent positions on all major surface tracks are obtained from a satellite radar map, for a region corresponding to a segment of a radar swath. The positions of ownforce ships are known, leaving four unidentified tracks. From earlier surveillance, it is known that only four other ships could be in that region, an oiler, a destroyer, and a 218 cruiser, the latter three comprising a Soviet underway replenishment group. A signal which only the cruiser could emit is consistent in bearing with two tracks; and a visual sighting of the merchant eliminates one. It is then known which tracks correspond to the unrep group, so the lead-track of the group is eliminated as a track of the oiler. Every platform is then associated with its track. CONCLUSIONS Reference I discusses the additional kinds of rules and capabilities that must be included in an operational PTAPS and the problems involved in integrating PTAPS rules into an actual tactical situation assessment (TSA) system, and none of these conclusions has changed. The most difficult problem with compatibility with TSA rules may be the handling of confidence values. PTAPS does not use confidence values and must be constrained from operating on assertions that have less than a near-certainty confidence value. In discussions regarding con- fidence values, reference 1 describes how conclusions which would logically follow from different assumptions about particular tracks or platforms could be determined by PTAPS and assigned confidence values based on the confidence values of the initial data. Implementing this would not be an easy task. The logical reasoning that can be implemented with PTAPS rules is essential to the function of associating tracks with platforms. If the other reasoning functions of tactical situation assessment are to be performed in a production system, then probably the PTAPS function also should be performed within that system, so that the functions can be easily coordinated and can share the data base. A possible alternative would be to create a specialized problem-solving technique for platform-track association and interface it with the production system, but coordination and data base sharing would be more difficult. The next desirable step in continuing PTAPS investigations is to integrate experimentally PTAPS rules with TSA rules in a production system. 219 REFERENCES 1. Dillard RA. "Higher Order Logic for Platform Identification in a Production System," Technical Document 288, Naval Ocean Systems Center, October 17, 1979. 2. Bechtel RJ and Morris PH. "STAMMER: System for tactical assessment of multisource messages, even radar," Technical Document 252, Naval Ocean Systems Center, May, 1979. 3. McCall DC. "Tactical Situation Assessment Using a Rule-Based Inference System." In Proceedings of the Second MIT/ONR Workshop on Distributed Information and Decision Systems Motivated by Naval Command-Control-Communications (C3) Problems. Held at the Naval Postgraduate School, July, 1979. 4. McCall DC, Morris PH, Kibler DF, and Bechtel RJ. "STAMMER2: A Production System for Tactical Situation Assessment," Technical Document 298, Volumes 1 and 2, Naval Ocean Systems Center, October, 1979. 5. Waterman, DA, Anderson, RH, Hayes-Roth, F, Klahr, P, Martins, G, Rosenschein, SJ. "Design of a Rule-Oriented System for Implementing Expertise," N-1 158-1-ARPA, The Rand Corporation, Santa Monica, CA, May, 1979. 6. Dillard RA. "Experimental Tests of PTAPS Performance in Three Types of Production System Structures," Technical Document 385, Naval Ocean Systems Center, September 17, 1980. 220 APPENDIX FOURTH MIT/ONR WORKSHOP ON DISTRIBUTED INFORMATION AND DECISION SYSTEMS MOTIVATED BY COMMAND-CONTROL-COMMUNICATIONS (C ) PROBLEMS June 15, 1981 through June 26, 1981 San Diego, California List of Attendees Table of Contents Volumes I-IV 221 MIT/ONR WORKSHOP OF DISTRIBUTED INFORMATION AND DECISION SYSTEMS MOTIVATED BY COMMAND-CONTROL-COMMUNICATIONS (C 3 ) PROBLEMS JUNE 15, 1981 - JUNE 26, 1981 ATTENDEES David S. Alberts Vidyadhana Raj Avilla Special Asst. to Vice President Electronics Engineer & General Manager Naval Ocean Systems Center The MITRE Corporation Code 8241 1820 Dolley Madison Blvd. San Diego, CA 92152 McLean, VA 22102 Tel: (714) 225-6258 Tel: (703) 827-6528 Dennis J. Baker GlZen AZZgaier Research Staff Electronics Engineer Naval Research Laboratory Naval Ocean Systems Center Code 7558 Code 8242 Washington DC 20375 San Diego, CA 92152 Tel: (202) 767-2586 Tel: (714) 225-7777 AZan R. Barnum Ami Arbel Technical Director Senior Research Engineer Information Sciences Division Advanced Information & Decision Systems Rome Air Development Center 201 San Antonio Circle #201 Griffiss AFB, NY 13441 Mountain View, CA 94040 Tel: (315) 330-2204 Tel: (415) 941-3912 Jay K. Beam MichaeZ Athans Senior Engineer Professor of Electrical Engineering Johns Hopkins University & Computer Science Applied Physics Laboratory Laboratory for Information and Johns Hopkins Road Decision Systems Laurel, MD 20810 Massachusetts Institute of Technology Tel: (301) 953-7100 x3265 Room 35-406 Cambridge, MA 02139 Tel: (617) 253-6173 Robert BechteZ Scientist Naval Ocean Systems Center Daniel A. Atkinson Code 8242 Executive Analyst San Diego, CA 92152 CTEC, Inc. Tel: (714) 225-7778 7777 Leesburg Pike Falls Church, VA 22043 Tel: (703) 827-2769 222 ATTENDEES C3 CONFERENCE PAGE TWO Vitalius Benokraitis Alfred Brandstein Mathematician Systems Analysis Branch US ARMY Material Systems Analysis CDSA MCDEC USMC ATTN: DRXSY - AAG Quantico, VA 22134 Aberdeen Proving Ground, MD 21005 Tel: (703) 640-3236 Tel: (301) 278-3476 James V. Bronson Lieutenant Colonel USMC Patricia A. BiZllings ey Naval Ocean Systems Center Research Psychologist MCLNO Code 033 Navy Personnel R&D Center San Diego, CA 92152 Code 17 Tel: (714) 225-2383 San Diego, CA 92152 Tel: (714) 225-2081 Rudolph C. Brown, Sr. Westinghouse Electric Corporation WiZZiam B. Bohan P. 0. 746 Operations Research Analyst MS - 434 Naval Ocean Systems Center Baltimore, MD 21203 Code 722 Tel: San Diego, CA 92152 Tel: (714) 225-7778 Thomas G. Bugenhagen Group Supervisor James Bond Applied Physics Laboratory Senior Scientist Johns Hopkins University Naval Ocean Systems Center Johns Hopkins Road Code 721 Laurel, MD 20810 San Diego, CA 92152 Tel: (301) 953-7100 Tel: (714) 225-2384 James R. CaZZan Paul L. Bongiovanni Research Psychologist Research Engineer Navy Personnel R&D Center Naval Underwater Systems Center Code 302 Code 3521 Bldg. 1171-2 San Diego, CA 92152 Newport, RI 02840 Tel: (714) 225-2081 Tel: (401) 841-4872 David Castanon Christopher Bowman Research Associate Member, Technical Staff Laboratory for Information and VERAC, Inc. Decision Systems 10975 Torreyana Road Massachusetts Institute of Suite 300 Technology San Diego, CA 92121 Room 35-331 Tel: (714) 457-5550 223 Cambridge, MA 02139 Tel: (617) 253-2125 ATTENDEES C3 CONFERENCE PAGE THREE S. I. Chou Robin Dillard Engineer Mathematician Naval Ocean Systems Center Naval Ocean Systems Center Code 713B Code 824 San Diego, CA 92152 San Diego, CA 92152 Tel: (714) 225-2391 Tel: (714) 225-7778 Gerald A. Clapp Elizabeth R. Ducot Physicist Research Staff Naval Ocean Systems Center Laboratory for Information and Code 8105 Decision Systems San Diego, CA 92152 Massachusetts Institute of Tel: (714) 225-2044 Technology Room 35-410 Cambridge, MA 02139 Douglas Cochran Tel: (617) 253-7277 Scientist Bolt Beranek & Newman Inc. Donald R. Edmonds 50 Moulton Street Group Leader Cambridge, MA 02138 MITRE Corporation Tel: (415) 968-9061 1820 Dolley Madison Blvd. McLean, VA 22102 Tel: (702) 827-6808 A. Brinton Cooper, III Chief, C3 Analysis US ARMY Material Systems Analysis Martin Einhorn ATTN: DRXSY-CC Scientist Aberdeen Proving Ground, MD 21005 Systems Development Corporation Tel: (301) 278-5478 4025 Hancock Street San Diego, CA 92110 Tel: (714) 225-1980 David E. Corman Engineer Jonhs Hopkins University Leon Ekchian Applied Physics Laboratory Graduate Student Johns Hopkins Road Laboratory for Information and Laurel, MD 20810 Decision Systems Tel: (301) 953-7100 x521 Massachusetts Institute of Technology Room 35-409 Wilbur B. Davenport, Jr. Cambrdige, MA 02139 Professor of Communications Sciences Tel: (617) 253-5992 & Engineering Laboratory for Information and Thomas Fortmann Decision Systems Senior Scientist Massachusetts Institute of Technology Bolt, Beranek & Newman, Inc. Room 35-214 50 Moulton Street Cambridge, MA 02139 Cambridge, MA 02138 Tel: (617) 253-2150 Tel: (617) 497-3521 224 ATTENDEES C3 CONFERENCE PAGE FOUR Clarence J. Funk Peter P. Groumpos Scientist Professor of Electrical Eng. Naval Ocean Systems Center Cleveland State University Code 7211, Bldg. 146 Cleveland, OH 44115 San Diego, CA 92152 Tel: (216) 687-2592 Tel: (714) 225-2386 George D. Halushynsky Mario Gerla Member of Senior Staff Professor of Electrical Engineering Johns Hopkins University & Computer Science Applied Physics Laboratory University of California Johns Hopkins Road Los Angeles Laurel, MD 20810 Boelter Hall 3732H Tel: (301) 953-7100 x2714 Los Angeles, CA 90024 Tel: (213) 825-4367 Scott Harmon Donald T. GiZes, Jr. Electronics Engineer Technical Group Naval Ocean Systems Center The MITRE Corproation Code 8321 1820 Dolley Madison Bldv. San Diego, CA 92152 McLean, VA 22102 Tel: (714) 225-2083 Tel: (703) 827-6311 David Haut Irwin R. Goodman Research Staff Scientist Naval Ocean Systems Center Naval Ocean Systems Center Code 722 Code 7232 San Diego, CA 92152 Bayside Bldg. 128 Room 122 Tel: (714) 225-2014 San Diego, CA 92152 Tel: (714) 225-2718 C. W. HeZstrom Professor of Electrical Eng, Frank Greitzer- & Computer Science Research Psychologist University of California, Navy Personnel R&D Center San Diego San Diego, CA 92152 La Jolla, CA 92093 Tel: (714) 225-2081 Tel: (714) 452-3816 Leonard S. Gross Ray L. Hershman Member of Technical Staff Research Psychologist VERAC, Inc. Navy Personnel R&D Center 10975 Torreyana Road Code P305 Suite 300 San Diego, CA 92152 San Diego, CA 92121 Tel: (714) 225-2081 Tel: (714) 457-5550 225 ATTENDEES C3 CONFERENCE PAGE FIVE Sam R. HoZZingsworth CarroZZ K. Johnson Senior Research Scientist Visiting Scientist Honeywell Systems & Research Center Naval Research Laboratory 2600 Ridgway Parkway Code 7510 Minneapolis, MN 55413 Washington DC 20375 Tel: (612) 378-4125 Tel: (202) 767-2110 Kuan-Tsae Huang Jesse KasZer Graduate Student Electronics Engineer Laboratory for Information and Naval Ocean Systems Center Decision Systems Code 9258, Bldg. 33 Massachusetts Institute of Technology San Diego, CA 92152 Room 35-329 Tel: (714) 225-2752 Cambridge, MA 02139 Tel: (617) 253- Richard T. KeZZey Research Psychologist James R. Hughes Navy Personnel R&D Center Major, USMC Code 17 (Command Systems) Concepts, Doctrine, and Studies San Diego, CA 92152 Development Center Tel: (714) 225-2081 Marine Corps Development & Education Quantico, VA 22134 Tel: (703) 640-3235 David KZeinman Professor of Electrical Eng. & Computer Science University of Connecticut Kent S. HuZZ Box U-157 Commander, USN Storrs, CT 06268 Deputy Director, Tel: (203) 486-3066 Mathematical & Information Sciences Office of Naval Research Code 430B Robert C. KoZb 800 N. Quincy Head Tactical Command Arlington, VA 22217 & Control Division Tel: (202) 696-4319 Naval Ocean Systems Center Code 824 San Diego, CA 92152 Carolyn Hutchinson Tel: (714) 225-2753 Systems Engineer Comptek Research Inc. 10731 Treena Street MichaeZ Kovacich Suite 200 Systems Engineer San Diego, CA 92131 Comptek Research Inc. Tel: (714) 566-3831 Mare Island Department P.O. Box 2194 Vallejo, CA 94592 Tel: (707) 552-3538 226 ATTENDEES C3 CONFERENCE PAGE SIX Timothy Kraft Alexander H. Levis Systems Engineer Senior Research Scientist Comptek Research, Inc. Laboratory for Information and 10731 Treena Street Decision Systems Suite 200 Massachusetts Institute of San Diego, CA 92131 Technology Tel: (714) 566-3831 Room 35-410 Cambridge, MA 02139 Tel: (617) 253-7262 Manfred Kraft Diplom-Informatiker Victor O.-K. Li Hochschule der Bundeswehr Professor of Electrical Eng. Fachbereich Informatik & Systems Werner-Heissenbergweg 39 PHE 8014 Neubiberg, West Germany University of Southern California Tel: (0049) 6004-3351 Los Angeles, CA 90007 Tel: (213) 743-5543 Leslie Kramer Senior Engineer GZenn R. Linsenmayer ALPHATECH, Inc. Westinghouse Electric Corporation 3 New England Executive Park P. O. Box 746 - M.S. 434 Burlington, MA 01803 Baltimore, MD 21203 Tel: (617) 273-3388 Tel: (301) 765-2243 Ronald W. Larsen Pan-Tai Liu Division Head Professor of Mathematics Naval Ocean Systems Center University of Rhode Island Code 721 Kingston, RI 02881 San Diego, CA 92152 Tel: (401) 792-1000 Tel: (714) 225-2384 Robin Magonet-Neray JoeZ S. Lawson, Jr. Graduate Student Chief Scientist C31 Laboratory for Information and Naval Electronic Systems Command Decision Systems Washington DC 20360 Massachusetts Institute of Tel: (202) 692-6410 TechnologyRoom 35-403 Cambridge, MA 02139 Tel: 617) 253-2163Dan Leonard Electronics Engineer Naval Ocean Systems Center Kevin MatZoy Code 8105 SCICON Consultancy San Diego, CA 92152 Sanderson House Tel: (714) 225-7093 49, Berners StreetLondon W1P 4AQ, United Kingdom Tel: (01) 580-5599227 ATTENDEES C3 CONFERENCE PAGE SEVEN Dennis C. McCaZZ Charles L. Morefield Mathematician Board Chairman Naval Ocean Systems Center VERAC, Inc. Code 8242 10975 Torreyana Road San Diego, CA 92152 Suite 300 Tel: (714) 225-7778 San Diego, CA 92121 Tel: (714) 457-5550 Marvin Medina Scientist Peter Morgan Naval Ocean Systems Center SCICON Consultancy San Diego, CA 92152 49-57, Berners Street Tel: (714) 225-2772 London W1P 4AQ, United Kingdom Tel: (01) 580-5599 Michael Melich Head, Command Information John S. Morrison Systems Laboratory Captain, USAF Naval Research Laboratory TAFIG/IICJ Code 7577 Langeley AFB, VA 23665 Washington DC 20375 Tel: (804) 764-4975 Tel: (202) 767-3959 MichaeZ S. Murphy John MeZviZZlle Member of Technical Staff Naval Ocean Systems Center VERAC, Inc. Code 6322 10975 Torreyana Road San Diego, CA 92152 Suite 300 Tel: (714) 225-7459 San Diego, CA 92121 Tel: (714) 357-5550 Glenn E. MitzeZ Engineer Jim Pack Johns Hopkins University Naval Ocean Systems Center Applied Physics Laboratory Code 6322 Johns Hopkins Road San Diego, CA 92152 Laurel, MD 20810 Tel: (714) 225-7459 Tel: (301) 953-7100 x2638 Bruce Patyk MichaeZ H. Moore Naval Ocean Systems Center Senior Control System Engineer Code 9258, Bldg. 33 Systems Development Corporation San Diego, CA 92152 4025 Hancock Street Tel: (714) 225-2752 San Diego, CA 92037 Tel: (714) 225-1980 22G ATTENDEES C3 CONFERENCE PAGE EIGHT Roland Payne Barry L. Reichard Vice President Field Artillery Coordinator Advanced Information & Decision Systems US Army Ballistic Research 201 San Antonio Circle #286 Laboratory Mountain View, CA 94040 ATTN: DRDAR-BLB Tel: (415) 941-3912 Aberdeen Proving Ground, MD 21014 Tel: (301) 278-3467 Anastassios Perakis Graduate Student David RenneZs Ocean Engineering Professor of Computer Science Massachusetts Institute of Technology University of California, LA Room 5-426 3732 Boelter Hall Cambridge, MA 02139 Los Angeles, CA 90024 Tel: (617) 253-6762 Tel: (213) 825-2660 Lloyd S. Peters Thomas P. Rona Associate Director Staff Scientist Center for Defense Analysis Boeing Aerosapce Company SRI International MS 84-56 EJ352 P. O. Box 3999 333 Ravenswood Avenue Seattle, WA 98124 Menlo Park, CA 94025 Tel: (206) 773-2435 Tel: (415) 859-3650 Nils R. Sandell, Jr. HariZaos N. Psaraftis President & Treasurer Professor of Marine Systems ALPHATECH, Inc. Massachusetts Institute of Technology 3 New England Executive Park Room 5-213 Burlington, MA 01803 Cambridge, MA 02139 Tel: (617) 273-3388 Tel: (617) 253-7639 Daniel Schutzer Paul M. Reeves Technical Director Electronics Engineer Naval Intelligence Naval Ocean Systems Center Chief of Naval Operations Code 632 NOP 009T San Diego, CA 92152 Washington DC 20350 Tel: (714) 225-2365 Tel: (202) 697-3299 229 ATTENDEES C3 CONFERENCE PAGE NINE Adrian SegaZZ T. Tao Professor of Electrical Engineering Professor Technion IIT Naval Postgraduate School Haifa, Israel Code 62 TV Tel: (617) 253-2533 Monterey, CA 93940Tel: (408) 646-2393 or 2421 Prodip Sen H. Gregory Tornatore Polysystems Analysis Corporation Johns Hokpins University P. O. Box 846 Applied Physics Laboratory Huntington, NY 11743 Johns Hopkins Road Tel: (516) 427-9888 Laurel, MD 20810 Tel: (301) 953-7100 x2978 Harlan Sexton Naval Ocean Systems Center Edison Tse Code 6322 Professor of Engineering San Diego, CA 92152 Economic Systems Tel: (714) 225-2502 Stanford University Stanford, CA 94305 Tel: (415) 497-2300 Mark J. Shensa Naval Ocean Systems Center Code 6322 E. B. TurnstaZZll San Diego, CA 92152 Head, Ocean Surveillance Tel: (714) 225-2349 or 2501 Systems Department Naval Ocean Systems Center Code 72 J. R. Simpson San Diego, CA 92152 Office of Naval Research Tel: (714) 225-7900 800 N. Quincy Arlington, VA 22217 Tel: (202) 696-4321 Lena Valavani Research Scientist Laboratory for Information and Stuart H. Starr Decision Systems Director Systems Evaluation Massachusetts Institute of The Pentagon Technology DUSD (C31), OSD Room 35-437 Room 3E182 Cambridge, MA 02139 Washington DC 20301 Tel: (617) 253-2157 Tel: (202) 695-9229 230 ATTENDEES C3 CONFERENCE PAGE TEN ManieZ Vineberg Richard P. Wishner Electronics Engineer President Naval Ocean Systems Center Advanced Information & Decision Code 9258 Systems San Diego, CA 92152 201 San Anotnio Circle Tel: 714) 225-2752 Suite 286 Mountain View, CA 94040 Tel: (415) 941-3912 Joseph H. Wack Advisory Staff Westinghouse Electric Corporation Joseph G. WohZ P. O. Box 746 MS-237 V. P. Research & Development Baltimore, MD 21203 ALPHATECH, Inc. Tel: (301) 765-3098 3 New England Executive Park Burlington, MA 01803 Tel: (617) 273-3388 Jan D. Wald Senior Research Scientist Honeywell Inc. John M. Wosencraft Systems & Research Center Head of C3 Curriculum MN 17-2307 Naval Postgraduate School P. O. Box 312 Code 74 Minneapolis, MN 55440 Monterey, CA 93940 Tel: (612) 378-5018 Tel: (408) 646-2535 Bruce K. Walker Lofti A. Zadeh Professor of Systems Engineering Professor of Computer Science Case Western Reserve University University of California Cleveland, OH 44106 Berkeley, CA 94720 Tel: (216) 368-4053 Tel: (415) 526-2569 David White Advanced Technology, Inc. 2120 San Diego Avenue Suite 105 San Diego, CA 92110 Tel: (714) 981-9883 Jeffrey E. Wieseithier Naval Research Laboratory Code 7521 Washington DC 20375 Tel: (202) 767-2586 231 SURVEILLANCE AND TARGET TRACKING FOREWORD ..................................................... DATA DEPENDENT ISSUES IN SURVEILLANCE PRODUCT INTEGRATION Dr. Daniel A. Atkinson .......................................... MEMORY DETECTION MODELS FOR PHASE-RANDOM OCEAN ACOUSTIC FLUCTUATIONS Professor Harilaos N. Psaraftis, Mr. Anatassios Perakis, and Professor Peter N. Mikhahelvsky ................................. DETECTION TRESHOLDS FOR MULTI-TARGET TRACKING IN CLUTTER Dr. Thomas Fortmann,Professor Yaakov Bar-ShaZom, and Dr. Molly Scheffe ............................................... MULTISENSOR MULTITARGET TRACKING FOR INTERNETTED FIGHTERS Dr. Christopher L. Bowman ....................................... MARCY: A DATA CLUSTERING AND FUSION ALGORITHM FOR MULTI-TARGET TRACKING IN OCEAN SURVEILLANCE Dr. Michael H. Moore ............................................ AN APOSTERIORI APPROACH TO THE MULTISENSOR CORRELATION OF DISSIMILAR SOURCES Dr. Michael M. Kovacich ........................................ A UNIFIED VIEW OF MULTI-OBJECT TRACKING Drs. Krishna R. Pattipati, Nils R. SandellZZ, Jr., and Leslie C. Kramer ................................................ OVERVIEW OF SURVEILLANCE RESEARCH AT M.I.T. Professor Robert R. Tenney ..................................... A DIFFERENTIAL GAME APPROACH TO DETERMINE PASSIVE TRACKING MANEUVERS Dr. Paul L. Bongiovanni and Professor Pan-T. Liu ............... 232 DESCRIPTION OF AND RESULTS FROM A SURFACE OCEAN SURVEILLANCE SIMULATION Drs. Thomas G. Bugenhagen, Bruce Bundsen, and Lane B. Carpenter ............................................ AN OTH SURVEILLANCE CONCEPT Drs. Leslie C. Kramer and NiZs R. Sandell, Jr .................. APPLICATION OF AI METHODOLOGIES TO THE OCEAN SURVEILLANCE PROBLEM Drs. Leonard S. Gross, MichaeZ S. Murphy, and CharZes L. Morefield .......................................... A PLATFORM-TRACK ASSOCIATION PRODUCTION SUBSYSTEM Ms. Robin DiZZard ............................................. SYSTEM ARCHITECTURE AND EVALUATION FOREWORD ................................................ C I SYSTEMS EVALUATION PROGRAM Dr. Stuart H. Starr ......................................... C SYSTEM RESEARCH AND EVALUATION: A SURVEY AND ANALYSIS Dr. David S. AZberts. THE INTELLIGENCE ANALYST PROBLEM Dr. DanieZ Schutzer ......................................... DERIVATION OF AN INFORMATION PROCESSING SYSTEMS (C /MIS) --ARCHITECTURAL MODEL -- A MARINE CORPS PERSPECTIVE Lieutenant CoZoneZ James V. Bronson ......................... A CONCEPTUAL CONTROL MODEL FOR DISCUSSING COMBAT DIRECTION SYSTEM (C2) ARCHITECTURAL ISSUES Dr. Timothy Kraft and Mr. Thomas Murphy ..................... EVALUATING THE UTILITY OF JINTACCS MESSAGES Captain John S. Morrison .................................... FIRE SUPPORT CONTROL AT THE FIGHTING LEVEL Mr. Barry L. Reichard ....................................... A PRACTICAL APPLICATION OF MAU IN PROGRAM DEVELOPMENT Major James R. Hughes ........................................ HIERARCHICAL VALUE ASSESSMENT IN A TASK FORCE DECISION ENVIRONMENT Dr. Ami ArbeZ ........................................... 234 OVER-THE-HORIZON, DETECTION, CLASSIFICATION AND TARGETING (OTH/DC&T) SYSTEM CONCEPT SELECTION USING FUNCTIONAL FLOW DIAGiAMS Dr. GZenn E. MitzeZ ......................................... A SYSTEMS APPROACH TO COMMAND, CONTROL AND COMMUNICATIONS SYSTEM DESIGN Dr. Jay K. Beam and Mr. George D. Haluschynsky .............. MEASURES OF EFFECTIVENESS AND PERFORMANCE FOR YEAR 2000 TACTICAL C3 SYSTEMS Dr. Djimitri Wiggert ........................................ AN END USER FACILITY (EUF) FOR COMMAND, CONTROL, AND COMMUNICATIONS (C3) Drs. Jan D. Wald and Sam R. HoZZlingsworth ................... 235 COMMUNICATION, DATA BASES & DECISION SUPPORT FOREWORD ................................................ RELIABLE BROADCAST ALGORITHMS IN COMMUNICATIONS NETWORK Professor Adrian Segall ..................................... THE HF INTRA TASK FORCE COMMUNICATION NETWORK DESIGN STUDY Drs. Dennis Baker, Jeffrey E. Wieselthier, and Anthony Ephremides ..................................... FAIRNESS IN FLOW CONTROLLED NETWORKS Professors Mario Gerla and Mark Staskaukas .................. PERFORMANCE MODELS OF DISTRIBUTED DATABASE Professor Victor O.-K. Li ................................... ISSUES IN DATABASE MANAGEMENT SYSTEM COMMUNICATION Mr. Kuan-Tase Huang and Professor Wilbur B. Davenport, Jr... MEASUREMENT OF INTER-NODAL DATA BASE COMMONALITY Dr. David E. Corman ......................................... MULTITERMINAL RELIABILITY ANALYSIS OF DISTRIBUTED PROCESSING SYSTEMS Professors Aksenti Grnarov and Mario GerZa .................. FAULT TOLERANCE IMPLEMENTATION ISSUES USING CONTEMPORARY TECHNOLOGY Professor David Rennels ..................................... APPLICATION OF CURRENT AI TECHNOLOGIES TO C2 Dr. Robert Bechtal .......................................... 236 A PROTOCOL LEARNING SYSTEM FOR CAPTURING DECISION-MAKER LOGIC Dr. Robert Bechtal ........................................... ON USING THE AVAILABLE GENERAL-PURPOSE EXPERT-SYSTEMS PROGRAMS Dr. CarroZZ K. Johnson ........................................ 237 IV C THEORY FOREWORD. RATE OF CHANGE OF UNCERTAINTY AS AN INDICATOR OF COMMAND AND CONTROL EFFECTIVENESS Mr. Joseph G. WohZ ........................................... THE ROLE OF TIME IN A COMMAND CONTROL SYSTEM Dr. JoeZ S. Lawson, Jr ...................................... GAMES WITH UNCERTAIN MODELS Dr. David Castanon ........................................... INFORMATION PROCESSING IN MAN-MACHINE SYSTEMS Dr. Prodip Sen and Professor Rudolph F. Drenick .............. MODELING THE INTERACTING DECISION MAKER WITH BOUND RATIONALITY Mr. Kevin L. Boettcher and Dr. Alexander H. Levis ............ DECISION AIDING -- AN ANALYTIC AND EXPERIMENTAL STUDY IN A MULTI-TASK SELECTION PARADIGM Professor David L. KZeiman and Drs. Eric P. Soulsby, and Krishna R. Pattipati ............................... FUZZY PROBABILITIES AND THEIR ROLE IN DECISION ANALYSIS Professor Lotfi A. Zadeh .................................... COMMAND, CONTROL AND COMMUNICATIONS (C3 ) SYSTEMS MODEL AND MEASURES OF EFFECTIVENESS (MOE's) Drs. Scot Harmon and Robert Brandenburg ...................... THE EXPERT TEAM OF EXPERTS APPROACH TO C ORGANIZATIONS Professor MichaeZ Athans ..................................... 238 A CASE STUDY OF DISTRIBUTED DECISION MAKING Professor Robert R. Tenney ................................... ANALYSIS OF NAVAL COMMAND STRUCTURES Drs. John'R. Delaney, Nils R. Sandell, Jr., Leslie C. Kramer, and Professors Robert R. Tenney and MichaeZ Athans ........... MODELING OF AIR FORCE COMMAND AND CONTROL SYSTEMS Dr. Gregory S. Lauer, Professor Robert R. Tenney, and Dr. Nits R. SandeZZ, Jr ..................................... A FRAMEWORK FOR THE DESIGN OF SURVIVABLE DISTRIBUTED SYSTEM -- PART I: COMMUNICATION SYSTEMS Professors Marc Buchner and Victor MatuZa: presented by Professor Kenneth Loparo .................................. A FRAMEWORK FOR THE DESIGN OF SURVIVABLE DISTRIBUTED SYSTEMS -- PART II: CONTROL AND INFORMATION STRUCTURE Professors Kenneth Loparo, Bruce Walker and Bruce Griffiths .. CONTROL SYSTEM FUNCTIONALIZATION OF C SYSTEMS VIA TWO-LEVEL DYNAMICAL HIERARCHICAL SYSTEMS (DYHIS) Professor Peter P. Groumpos .................................. SEQUENTIAL LINEAR OPTIMIZATION & THE REDISTRIBUTION OF ASSETS Lt. CoZoneZ Anthony Feit and Professor John M. Wozencraft .... C AND WAR GAMES -- A NEW APPROACH Dr. AZfred G. Brandstein ..................................... 239