Situation and Threat Refinement Approach for Combating the Asymmetric Threat

Transcription

1 Approved for public release; distribution is unlimited Situation and Threat Refinement Approach for Combating the Asymmetric Threat Angela M. Pawlowski, Sergio Gigli, and Frank J. Vetesi Lockheed Martin Advanced Technology Laboratories Camden, NJ [apawlows, sgigli, 1 ABSTRACT In order to combat the present and future asymmetric threats to national and international security, information fusion developments must progress beyond current Level 1 (Object Refinement) paradigms. By focusing on the challenges of Continuous Intelligence Preparation of the Battlespace (CIPB), Lockheed Martin Advanced Technology Laboratories (ATL) has begun to elicit an infrastructure and enabling technologies for information fusion at Level 2 (Situation Refinement) and Level 3 (Threat Refinement). Our approach to Level 2 is to perform spatial and temporal processing on tracks produced by Level 1 multi-sensor, multi-target fusion supplemented with intelligence information from both structured data sources such as databases, and from unstructured data sources such as text documents. The output of Level 2 is referred to as a factlet. Factlet creation is driven both by the evolution of events in the battlespace and by the top-down information needs of the CIPB analysts working at Level 3. Our approach to Level 3 is to drive existing and newly formulated models of threat behavior, viewed from multiple perspectives, such as political, economic, and tactical, with factlets derived in Level 2, to support the determination of possible enemy courses of action. We leverage existing algorithmic lessons learned from Level 1 Classification Fusion, where each sensor represented track classification from a different, complementary perspective, for factlet or evidence fusion. In addition, an approach to automated threat information discovery is discussed, which is initiated by the need for supporting evidence to further refine the Level 3 inferencing processes. Our approach to Level 2 Situation Refinement and Level 3 Threat Refinement will be demonstrated via an Air Force CIPB scenario. 2 INTRODUCTION There is a critical need to provide today s military decision makers with actionable information. This is information that the decision makers can use to make more effective targeting decisions, to plan friendly courses of action (COAs), to mitigate the impact of unexpected adversary actions, and to direct sensing systems to better observe adversary behavior. A key requirement to generate this type of actionable information is the ability to predict the adversary s most likely future COAs. In today s asymmetric warfare, such prediction is especially difficult. Lockheed Martin Advanced Technology Laboratories (ATL) is engaged in internal research and development work to assess adversary COAs and thus enable the generation of actionable information for decision makers. ATL is leveraging this work into an Air Force Research Laboratories (AFRL) Information Institute Research Project, Adversary Intent Inferencing for Predictive Battlespace Awareness [1]. 2.1 OVERVIEW Section 3 of this paper presents a discussion of the operational issues facing military decision makers. Currently, intelligence analysts perform adversary COA prediction manually. This manual process is becoming increasingly difficult to perform in an accurate and timely fashion. Section 4 discusses ATL s technical approach to designing and implementing a COA decision aid (DA) for intelligence analysts. -1-

2 Specifically, ATL has addressed the COA analysis problem as a data fusion problem. Our technical approach leverages ATL s experience in Level 1 data fusion gained as developers of the Rotorcraft Pilot s Associate (RPA) Data Fusion System [2, 3], which was flight demonstrated aboard an Apache AH/64-D in August 1999 and ATL s Extendable Mobile Agent Architecture (EMAA), developed via 14 DARPA and DoD intelligent agent programs and more than $10M of R&D funding [4, 5]. In Section 5, we present an experimental scenario, based on the events leading up to the Battle of Khafji during the Gulf War, to illustrate the functioning of our DA. 3 OPERATIONAL ISSUES 3.1 OBJECTIVE General Jumper, Chief of Staff of the US Air Force is a key proponent of a major new Air Force initiative known as Predictive Battlespace Awareness (PBA). At the center of the PBA concept is the fact that, if one can predict what the adversary is likely to do next, friendly forces can be more effective in: Targeting, by ensuring that strike platforms are in the correct position and of the correct type to be able to best attack the adversary. Planning friendly COAs to mitigate the impact of unexpected adversary actions. Tasking of intelligence, surveillance, and reconnaissance (ISR) collection assets (usually sensor systems) to cover areas where the adversary is expected to be next. A key enabler for PBA is the ability to predict adversary COAs. 3.2 CURRENT ADVERSARY COA PREDICTION Today, the prediction of adversary COAs is done almost exclusively by intelligence analysts. In the case of the Air Force, these intelligence analysts work in the Fusion Cell of the Combat Intelligence Division of the Air Operations Center [6]. Intelligence analysts work towards a prediction of adversary COA by using a four-step process called Intelligence Preparation of the Battlespace (IPB). The four steps of the IPB process, as shown in Figure 1, are: (1) define the battlespace, (2) describe the battlespace effects, (3) evaluate the adversary, and (4) determine the adversary COA. Clearly, the output of Step 4 of IPB is the adversary COA. AF AF Definition of of Predictive Battlespace Awareness (PBA) IPB IPB TSA ISR Step Step 1 Step 2 Step 3 Step 4 Define Battlespace Describe Battlespace Effects Evaluate Adversary Determine Adversary COA Figure 1 Intelligence Preparation of the Battlespace (IPB) -2-

3 In practice, because new intelligence information is constantly becoming available in the battlespace, IPB is a continuous process that iterates multiple times during the course of a conflict. Enemy COA analysis is a critical development in Continuous Intelligence Preparation of the Battlespace (CIPB). 3.3 CHALLENGES OF CURRENT ADVERSARY COAs PREDICTION We believe, in today s battlespace and with today s asymmetric threats present, CIPB is becoming an increasingly difficult process for humans to perform both accurately and in a timely fashion because: Human analysts are limited in the volume of data that they can organize, store, and process. Human analysts are limited in the speed at which they can process data. It is difficult for human analysts to perform a complete evaluation of an adversary from all possible perspectives, such as political, economic, and cultural. Data must be combined to arrive at a composite picture of the adversary. A doctrine is an organized pattern of behavior common in military operations. Doctrine for today s asymmetric adversaries are minimal or nonexistent. Without doctrine, adversary COA prediction becomes more difficult for human analysts; with faster moving adversaries, the analyst has less time to reach a prediction. 4 TECHNICAL ISSUES A key operational objective for military decision makers is to have advance knowledge of an adversary s COA. ATL is developing decision aids (DAs) to assist human intelligence analysts with the CIPB process. 4.1 DATA FUSION ATL s approach to resolving the prediction of adversary COAs is to treat the process as a data fusion problem. Data fusion is the complex problem of bringing together data from multiple sources in many different formats to create one overall integrated picture from which decision makers can act. An architecture for data fusion is provided by the Joint Directors of Laboratories (JDL) data fusion model [7]. Key features of the JDL model include: Information sources. The information sources may be of many different types including sensor systems, structured data sources such as databases, unstructured text documents and reports, and input from people. Level 2 processing Situation Refinement attempts to discover the relationships that exist between the objects (or entities) in the battlespace, events that occur within the battlespace, and the environment. Level 3 processing Threat Refinement projects the current situation in the battlespace forward into the future. One output from Level 3 processing is an assessment of the likely adversary COAs. ATL supports proposals to add an additional level, Level 5, to the JDL model. Level 5 processing, User Refinement, is concerned with taking feedback from the operator of the data fusion system, in our case, the military intelligence analysts, and using this feedback to control and improve the fusion performance. Our DA designs have been impacted by the JDL data fusion model. The JDL model does not, however, provide specific details as to how to implement the model in software. ATL has designed, and is currently implementing, a software prototype of a DA for intelligence analysts, providing Level 2, Level 3 and Level 5 capabilities (Figure 2). In addition, we have engineered the interfaces between our DA and ATLdeveloped data fusion systems performing Level 1 fusion. -3-

4 Evidence Fusion Threat COA Generation Plausible COAs Level 2 Level 3 Master Threat Model Threat Perspective.. Threat Perspective A Model Z Model Evidence Needs Evaluation Calls to Factlet Functions Factlet Analysis Functions Virtual Battlespace Virtual Battlespace Factlets (Evidence) Factlets Level 5 COAs (New Models or Mods) Smart Agent Generation Engine (SAGE) DBs Non-Traditional Data Sources Level 1 Fused Tracks Figure 2 ATL s Decision Aid for Intelligence Analysts 4.2 OBSERVATIONS, DATA SOURCES, AND EMAA AGENTS This section introduces three components that form the core of ATL s decision aids. Firstly, we define observations as data containers that transport data into and around the DA. An observation may originate from a battlefield sensor or from an intelligence report. Secondly, we discuss the data sources that provide data for our DA to reason about. Finally, we present details of ATL s EMAA agents. These agents are a key enabling technology for discovering and retrieving data from the data sources Observations We have defined a format for a data container, referred to as an observation, that transports data within our DA. For example, a report from a field operative on the sighting of a tribal meeting, will be converted to the observation format for presentation to our DA. Each observation may have one or more of the following descriptors: The source of the data in the observation allows us to maintain the pedigree of the information. The time at which the observation was generated. The spatial location of the observation. The name of the object about which the observation is made. Objects can range broadly in extent and in type. Thus a howitzer is an object, as is a bridge, a tribal clan, or the leadership of a country. The attributes of the object in question. Attributes are grouped by perspectives such as damage range (how far can an object project damage?), economic significance and allegiance (with whom or what is the object in question allied?). The confidence measure in the observation can be expressed as actual probabilities; measures of belief, disbelief, and ignorance; fuzzy assessments of confidence such as very probable, somewhat probable, and unlikely; and simply true or false Data Sources ATL s DAs access information from a wide range of distributed, heterogeneous data sources, categorized into traditional and non-traditional sources. Traditional sources will typically be structured -4-

5 intelligence databases shown in Figure 2 as DBs and databases containing fused tracks derived by Level 1 processing. Non-traditional data sources refer to large unstructured text streams such as documents and Internet websites. These sources may well contain a wealth of information pertinent to the adversary and to the battlespace in general. In general, the DA must react to traditional information that is continually accumulating in the structured data stores. The DA must also be proactive about retrieving data from non-traditional, unstructured data sources EMAA Agents The mechanisms to populate and transport observations throughout the DA architecture are based on ATL s EMAA agents. EMAA agents are very small threads of mobile software, and are readily composable. A composable agent is an agent that is constructed from a series of tasks. Examples of tasks include moving from one computer to another or accessing data from a data source. These tasks are assembled in a sequence to form the itinerary of an EMAA agent. Control logic may also be embedded between tasks in an agent s itinerary to allow the agent to make intelligent, autonomous choices as to which particular subset of tasks to follow in a given situation. ATL s DA makes extensive use of EMAA composable agents as data discovery, retrieval, and monitoring tools. The tasks these agents perform closely mirror those executed by human analysts performing intelligence data retrieval. To support the accumulation of threat evidence, ATL has developed the Smart Agent Generation Engine (SAGE) to automate the construction of itineraries for data retrieval and discovery agents to support both Level 2 and Level 3 processing as shown in Figure 2 above. SAGE contains libraries of data retrieval, discovery, and processing steps that it can sequence and combine to construct unique itineraries for agents to follow. SAGE constructs these itineraries based on requests for information from the DA. Specific implementations of SAGE to support both the Level 2 and the Level 3 DA are discussed below. 4.3 JDL LEVEL 2 SITUATION REFINEMENT DA The objectives of JDL Level 2 Situation Refinement include making inferences about the measurements and observations that are available and establishing relationships between entities, events, and the environment. This section discusses ATL s concept of the functional objectives of JDL Level 2 in a DA that is being developed at ATL (Figure 2) Virtual Battlespace The Virtual Battlespace is the central Level 2 DA space in which to store objects. When the DA first encounters a new object, it instantiates it in the Virtual Battlespace, locating it, if appropriate in a spatial grid. In addition, the DA is responsible for correctly organizing observations in the Virtual Battlespace, both spatially and temporally, by grouping and queuing these observations with the objects to which they refer. The Virtual Battlespace has areas with accurate spatial dimensions to allow the placement of objects such as tanks, cities and power stations and it has areas without spatial dimensions to place objects such as the adversary leadership. The Virtual Battlespace is built in two stages. During the first, initialization phase, the Virtual Battlespace is populated with observations generated before the commencement of hostilities. These observations tend to be more strategic in nature addressing among other issues natural objects such as terrain, lakes, and forests; man-made objects such as cities, ports, bridges, roads, and railroads; and military objects such as troop camps, bunkers, airfields, and air defense sites. Also among the strategically oriented observations are those referring to political objects such as a country s leadership and ethnic groups; economic objects such as oil fields, refineries, and manufacturing centers; and religious objects such as -5-

6 mosques and holy sites. During the second execution phase, observations will typically be tactical observations such as the movement of infantry divisions or small bands of soldiers within the battlespace. The Virtual Battlespace also ensures that observations correctly update the state of objects. For example, if an observation arrives that food supplies to a training camp have been cut off, then the occupants of the camp should have their morale and effectiveness descriptors reduced Factlet Analysis Functions The Factlet Analysis Functions execute across the extent of the Virtual Battlespace, reasoning across the objects present, and within each analysis perspective, to generate both measured and inferred items of evidence, the factlets. These Functions are concerned with establishing the relationships between objects in the Virtual Battlespace. For example, the Motion Analysis Function considers the movement patterns of groups (established by the Aggregate Analysis Function) of military objects such as armored personnel carriers. The Motion Analysis Function may conclude that the current movement pattern indicates a probing behavior on the part of the adversary, rather than a full scale attack. This inference becomes a factlet. Factlets are statements or evidence about the situation in the battlespace and they form the main input to the Level 3 DA. An example of a Level 2 Factlet Analysis Function is the Economic Center of Gravity Analysis Function, which executes across the entire Virtual Battlespace and identifies objects that are likely to have economic influence on the organization. This organization could be a country or a terrorist group. A power station is likely to be a major economic center of gravity as is a cocoa plantation. On the other hand, an un-navigable river, with no hydroelectric facilities, is unlikely to be a major economic contributor to its host organization. The output of this Function is a factlet describing the economic centers of gravity for each of the red (adversary), gray (non-combatant), and blue (friendly) forces present in the battlespace Level 2 SAGE In order to form factlets, the Factlet Analysis Functions determine the information they will need for their computations. In many cases, this information will flow into Level 2 from measurements made in Level 1. In other cases, software agents will be needed to retrieve the needed data from additional information sources. SAGE, at Level 2, takes a request for data and translates it into an agent s itinerary. As an example, the Level 2 Allegiance Analysis Function assesses the likely allegiance of objects within the battlespace. Although the names of the leaders of a collection of factions within the battlespace are known, the allegiances of these factions may have not yet been stored in the allegiance perspective of the faction objects. The Allegiance Analysis Function will task the Level 2 SAGE to discover, retrieve, and process any allegiance information that it can locate on the various factions. Level 2 SAGE will use the names of the leaders of the factions as search keys within both traditional, structured and non-traditional, unstructured data sources. In addition, SAGE understands rules of association (by leveraging DAML ontologies), namely that if Joe is associated with Pierre and Pierre is associated with Organization A, then there may be some link between Joe and Organization A. 4.4 JDL LEVEL 3 THREAT REFINEMENT DA The objectives of JDL Level 3 Threat Refinement include establishing possible adversary COAs. This section discusses ATL s concept of the functional objectives of JDL Level 3 in a DA that is being developed at ATL and is shown in Figure 2 above. -6-

7 4.4.1 Threat Perspective Models The Level 3 DA contains a number of threat perspective models. Collectively, these models form Level 3 s central repository of threat intent inference information. Each model is concerned with reasoning about possible adversary COAs from a single analysis perspective. Analysis perspectives include political, economic, social and ideological. Thus, an adversary s decision to invade a neighboring country will mostly likely be analyzed in different ways depending on the analysis perspective. From a political perspective, we may conclude that the adversary will not invade as it expects a strong, military intervention from United Nations forces. From an economic perspective, however, the possibility to acquire a route to the sea may be sufficient motivation for the adversary to invade. Each threat perspective model is represented by a hierarchical graph structure. The leaf nodes within the graph correspond to factlets passed in by the Evidence Needs Evaluation component described below. The graphs provide a structure for collecting factlets under a specific perspective and then reasoning as to the likelihood of one or more adversary COAs. For example, under the ideology threat perspective model, factlets contributing to the willingness of the adversary to sacrifice forces for its beliefs, will be reasoned about in determining the likelihood of a retreat COA. The computational mechanisms to perform the fusion of factlet evidence within each graph are provided by the Evidence Fusion component discussed below Master Threat Model The Level 3 Master Threat Model (see Figure 2) is responsible for reasoning about adversary COAs by combining evidence from the entire range of analysis perspectives. The Master Threat Model is thus highly interconnected with all the Threat Perspective Models. In addition the Master Threat Model can receive factlets directly from the Evidence Needs Evaluation component. The Master Threat Model uses reasoning already performed under each analysis perspective, to further corroborate or refute hypotheses about likely adversary COAs. Considering the country invasion example introduced above, the Master Threat Model will attempt to combine the conclusions reached under the political (no invasion due to possibility of United Nation intervention) and economic (access to sea strategically valuable) perspectives to arrive at a likelihood of invasion, relative to the many other COAs open to the adversary. As with the Threat Perspective Models, the Master Threat Model relies on the Evidence Fusion component to provide the computational mechanisms to fuse evidence across the multiple analysis perspectives Evidence Needs Evaluation The Evidence Needs Evaluation component of the Level 3 DA is responsible for supplying the Threat Perspective Models and Master Threat Model (discussed above) with evidence. Evidence may be generated by calls to the Level 2 Factlet Analysis Functions and by data discovery performed by Level 3 SAGE (discussed below). Certain high priority factlets will be automatically generated by Level 2 and the Evidence Needs Evaluation component will merely route these factlets to the appropriate perspective and master models. Given that the perspective and threat models can reason over a large range of factlets, it is often the case that at any one moment, the perspective and master models lack a complete set of factlets over which to reason. At such times, the Evidence Needs Evaluation component examines the perspective and master models to generate a prioritized list of factlet needs. This list is assembled by considering the factlets, which if available, would contribute most to corroborating or refuting the current set hypotheses in the perspective and master models as to the likely adversary COAs. Once the Evidence Needs Evaluation component has assembled the prioritized list of evidence needs, it makes calls to the Level 2 Factlet Analysis Functions or to Level 3 SAGE to retrieve the missing evidence. -7-

8 4.4.4 Evidence Fusion The Evidence Fusion component of the Level 3 DA provides the computational mechanisms to fuse evidence in three areas of Level 3. Firstly, factlets grouped and routed to leaf nodes within the perspective and master models by the Evidence Needs Evaluation component must be combined. This stage of evidence fusion has to deal with conflicting evidence about the same concepts. For example, one data source may suggest that the morale of an enemy troop division is particularly low, while another source may suggest the opposite is true. Evidence Fusion takes these morale-related factlets and resolves conflicts based on, among other factors, the pedigree of the data sources. Secondly, Evidence Fusion combines evidence within each Threat Perspective Model. Evidence Fusion leverages evidence combination technology, known as the Modified Bayesian Evidence Combination (MBEC) algorithm, developed under ATL s Rotorcraft Pilot s Associate (RPA) [3]. This technology combines evidence in an efficient way, and with no need to specify a priori or conditional probability tables as required by Bayesian networks, which are typically difficult to obtain in this domain. Thirdly, Evidence Fusion combines evidence from the multiple Threat Perspective Models into the Master Threat Model. Evidence Fusion once again leverages the RPA MBEC algorithm to fuse this evidence Threat COA Generation The primary output from our Level 3 DA is the adversary s likely COAs. The Threat COA Generation component is responsible for compiling the list of the most likely adversary COAs. The Threat COA Generation component selects and ranks plausible adversary COAs suggested by the individual Threat Perspective Models and by the Master Threat Model. This selection and ranking process is based on the Level 1 track classification hypothesis generation paradigm developed by ATL under the RPA program. This approach examines the various hypotheses within the models, and from their likelihood values and from the structure of the graphs within each model, computes an optimal ranking of the hypotheses for selection and output to the intelligence analyst Level 3 SAGE The Level 3 DA contains many models that reason across information from multiple analysis perspectives. In cases where these models have incomplete information, Level 3 SAGE will be tasked to either retrieve additional information directly from the data sources or to make calls to the Level 2 Factlet Analysis Functions to generate additional factlets. 4.5 JDL LEVEL 5 HUMAN/COMPUTER COLLABORATION ATL has designed its data fusion system specifically as a decision aid for intelligence analysts, thereby explicitly recognizing the crucial role that the human/computer collaboration plays in assessing adversary COAs. The following functionality is provided by our Level 5 Human/Computer Collaboration system: The intelligence analyst is able to directly augment or override the factlets generated by the Level 2 Factlet Analysis Functions. The analyst can supply new COAs, new Threat Perspective Models, and can make modifications to existing Threat Perspective and Master Models. The DAs provide the analyst with plausible adversary COAs, along with detailed explanations of the processing steps involved. The DAs permit the analyst to call upon SAGE and COA explanation facilities on-demand. -8-

9 5 SCENARIO 5.1 OVERVIEW In order to demonstrate the processing flows of our DA and the potential impact such a DA might have in an operational environment, ATL identified the need for a scenario based on real world military entities and events. Exercising the DA against a rich, yet plausible scenario would also better demonstrate the capabilities that it could provide to support military intelligence analysts. Our selection criteria gave preference to high pedigree, public domain sources that provided rich descriptions of recent military engagements and resulted in the selection of a scenario based on the Battle of Khafji, which took place during the Gulf War between January 29, 1991 and January 31, 1991 [8, 9, 10, 11]. In the three weeks prior to the Battle of Khafji, Iraq encountered heavy and sustained losses to its forces and infrastructure. With Iraqi air defenses eliminated, the Coalition exercised its air superiority by focusing most of its air resources on SCUD batteries and Republican Guard divisions north of Khafji, Saudi Arabia. The shift of military priorities left the southern theater lightly defended, and with little coverage by the Coalition JSTARS sensor platform. Most Iraqi engagements in this area consisted of simple probing attacks that were quickly and easily repelled by Coalition forces. On January 29, 1991, Iraq decided to directly engage the Coalition forces by crossing the Saudi Arabian border with one armored and two mechanized divisions. The Iraqi forces broke through a lightly defended border crossing and continued down the main coastal road towards Khafji and Al Mishab. With Khafji deserted, the town was quickly captured by Iraqi forces on January 30, Due to pressure from the Saudi contingent of the Coalition, a plan to re-take Khafji was quickly formulated and ultimately executed successfully. Though no confirming data is available, it is speculated that the Iraqi offensive into Saudi Arabia was initiated in order to either (a) capture Coalition prisoners or (b) thwart a possible invasion by attacking the suspected Coalition staging area at the port of Al Mishab to the south of Khafji. One of the pivotal events in the days leading up to the Battle of Khafji was the change in Iraqi COAs from simple, light cross border probing incursions, to the massing and movement of a large scale force across the Saudi border. ATL s focus was to analyze whether our DAs could have anticipated this large Iraqi offensive. 5.2 DECISION AID EXECUTION This section presents a chronological account of the execution of ATL s decision aid on the Battle of Khafji scenario from the Gulf War. A variant of this scenario has also been used by ATL on the AFRL Information Institute Research Project, Adversary Intent Inferencing for Predictive Battlespace Awareness January 22, 1991: Massing On January 22, 1991, the DA has access to data from the US JSTARS platform and from other Coalition sensors that vehicles are massing in Kuwait. The DA also knows the location of various cultural features in the battlespace, such as the Kuwait Saudi border, and the location of the friendly, Coalition forces. This data is represented in the Virtual Battlespace. The Level 2 Factlet Analysis Functions then proceed with the Aggregate and Allegiance Analysis Functions and determine that the force is large and has allegiance to Iraq. These factlets are collected by the Level 3 Evidence Needs Evaluation component and routed to the Threat Perspective Models and to the Master Threat Model. The tactical perspective analysis asks whether this massing activity is taking place in proximity to any relevant battlespace features. SAGE -9-

10 accesses the intelligence databases and determines that the massing is in close proximity to both the Coalition s Marine outposts (shown in Figure 3 as numbered stars) and to the Kuwait Saudi border Marine Outpost Large force 5 Near outposts 4 Oil Field JSTARS reports 320 vehicles 2 Allegiance to Iraq Near border Iraqi Forces US Forces January 22 29, 1991: Moving Figure 3 January 22, 1991: Massing Between January 22 and 29, 1991, the DA receives observations from the JSTARS platform that the previously identified Iraqi vehicles are moving towards the Kuwait Saudi border (Figure 4). The Level 2 DA generates motion factlets, detailing that a large force is moving, with consistent velocity and direction, towards the border. The Level 3 Master Threat Model identifies a number of possible adversary COAs including a border attack, cross-border probing activity, or a training mission. SAGE is dispatched to look for additional evidence of jamming activity on behalf of the adversary - evidence that might corroborate a border attack. 6 Marine Outpost 5 4 Oil Field JSTARS reports Large force moving show heavy Consistent velocity traffic Look for evidence of attack, 2 probing or training Iraqi Forces US Forces Figure 4 January 22 29, 1991: Moving -10-

11 :30 Hours, January 29, 1991: Jamming At 18:30 hours on January 29, 1991, the DA receives observations from COMINT data sources, that the adversary is jamming friendly force communications (Figure 5). SAGE, which had been dispatched to look for evidence of jamming is now able to return positive evidence to the Threat Perspective and Master Threat Models. The border attack COA is assigned greater plausibility. SAGE is dispatched to find further evidence of a border attack including a search for evidence of an Iraqi move across the border and friendly forces taking fire. 6 Marine Outpost 5 Iraqi Look for further forces evidence of Oil attack Field 4 2 COMINT 1 7 detects jamming by Iraqi forces 8 Iraqi Forces US Forces Figure 5 18:30 Hours, January 29, 1991: Jamming :00 Hours, January 29, 1991: Engagement At 19:00 hours on January 29, 1991, the DA receives a situation report indicating that the Coalition Marine outposts are under fire and a JSTARS report showing Iraqi forces crossing the border close to a major coastal road from Kuwait into Saudi Arabia (Figure 6). Level 2 Factlet Analysis Functions 6 Marine Outpost 5 4 Oil Field SITREP Marine outpost 2 under fire JSTARS reports show Iraqi 1 forces 7 in column 8 Iraqi Forces US Forces Force-on-force attack Iraqi forces on coastal road Figure 6 19:00 Hours, January 29, 1991: Engagement -11-

12 conclude that the troop movement is down the coastal road and the Level 3 Threat COA Generation component selects and ranks a full force-on-force attack as the most plausible adversary COA. 5.3 LEVELS 2 AND 3 DECISION AIDING IMPACT Actual coalition forces, developing threat COAs manually, were unaware of the Iraqi maneuver until they were actually attacked on January 29, ATL s Level 2 and 3 Decision Aiding would have prevented the Coalition s surprise, via inferred COAs, starting one week before the attack on January 22, CONCLUSIONS The conclusions derived from our current research and development include: A key operational objective for military decision makers is the availability of automated assessments of the adversary s most likely courses of action. Such predictions represent actionable information. Currently, predictions of adversary COAs are generated by intelligence analysts using processes such as Intelligence Preparation of the Battlespace. Human analysts are limited in the volume and complexity of data that they can process. ATL has begun to design and develop decision aids to help intelligence analysts assess adversary COAs in a more accurate and timely fashion. ATL s DAs resolve adversary COA assessment as a data fusion problem. ATL s DA leverages a number of technologies developed by ATL, including the Extendable Mobile Agent Architecture, link discovery and analysis, and the evidence combination algorithms developed for the Rotorcraft Pilot s Associate. A high level overview of the execution of the DAs was shown using an experimental scenario developed from the Battle of Khafji during the Gulf War. 7 REFERENCES [1] Bell, B., Santos, E., Brown, S., Making Adversary Decision Modeling Tractable with Intent Inference and Information Fusion, 11 th Conference on CGF-BR, May 2002, Orlando, Florida. [2] Malkoff, D. and Pawlowski, A., RPA Data Fusion, 9 th National Symposium on Sensor Fusion, Vol. 1, Infrared Information Analysis Center, pp , September [3] Hofmann, M., Multi-Sensor Track Classification in Rotorcraft Pilot s Associate Data Fusion, American Helicopter Society 53 rd Annual Forum, Virginia Beach, Virginia, April 29 May 1, [4] Lentini, R., Rao, G., and Thies, J., EMAA: An Extendable Mobile Agent Architecture, AAAI Workshop, Software Tools for Developing Agents, July [5] Hofmann, M., Chacon, D., Mayer, G., Whitebread, K., CAST Agents: Network-Centric Fires Unleashed, 2001 National Fire Control Symposium: Session: Automation of the Kill Chain, August 27-30, 2001, Lihue Hawaii (Island of Kauai). [6] Air Operations Center Standard Operating Procedure Twelfth Air Force Air Force Forces, [7] Hall, D. L., Llinas, J., Handbook of Multisensor Data Fusion, CRC Press, [8] The Battle of Khafji: An Overview and Preliminary Analysis, Titus, September 1996, Airpower Research Institute, Maxwell AFB. [9] The Battle of Khafji: Implications for Airpower, Rochelle, MAJ USAF, June 1997, School of Advanced Airpower Studies, Air University, Maxwell AFB. [10] The Battle of Khafji: An Assessment of Airpower, Palmer LTC USA, Scott LTC USAF, Toolan LTC USMC, April 1998, Air War College, Air University, Maxwell AFB. [11] Airpower and The Battle of Khafji: Setting the Record Straight, Newell MAJ USAF, June 1998, School of Advanced Airpower Studies, Air University, Maxwell AFB. -12-