3
Integrative Architectures for Modeling the Individual Combatant

We have argued in the introduction that in order to make human behavior representation more realistic in the setting of military simulations, the developers will have to rely more on behavioral and organizational theory. In Chapters 3 through 11 we present scientific and applied developments that can contribute to better models. Chapters 4 through 8 review component human capacities, such as situation awareness, decision making, planning, and multitasking, that can contribute to improved models.

However, to model the dismounted soldier, and for many other purposes where the behavior of intact individuals is required, an integrative model that subsumes all or most of the contributors to human performance capacities and limitations is needed. Although such quantitative integrative models have rarely been sought by psychologists in the past, there is now a growing body of relevant literature. In the few cases in which these integrative models have been applied to military simulations, they have focused on specific task domains. Future integrative model developers will likely be interested in other domains than the ones illustrated here. Because each of our examples embodies a software approach to integrative models, we are referring to these developments as integrative modeling architectures. They provide a framework with behavioral content that shows promise of providing a starting point for model developers who wish to apply it to their domains. Each has its own strengths and weaknesses and we do not explicitly recommend any one. The developers' choices depend entirely on their goals and objectives in integrative model development. In Chapter 13, we do not recommend converging on a single integrative architecture, although we do argue that it is important to adopt modular structures that will allow easier interoperability among developed models.



The National Academies | 500 Fifth St. N.W. | Washington, D.C. 20001
Copyright © National Academy of Sciences. All rights reserved.
Terms of Use and Privacy Statement



Below are the first 10 and last 10 pages of uncorrected machine-read text (when available) of this chapter, followed by the top 30 algorithmically extracted key phrases from the chapter as a whole.
Intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text on the opening pages of each chapter. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

Do not use for reproduction, copying, pasting, or reading; exclusively for search engines.

OCR for page 51
Modeling Human and Organizational Behavior: Application to Military Simulations 3 Integrative Architectures for Modeling the Individual Combatant We have argued in the introduction that in order to make human behavior representation more realistic in the setting of military simulations, the developers will have to rely more on behavioral and organizational theory. In Chapters 3 through 11 we present scientific and applied developments that can contribute to better models. Chapters 4 through 8 review component human capacities, such as situation awareness, decision making, planning, and multitasking, that can contribute to improved models. However, to model the dismounted soldier, and for many other purposes where the behavior of intact individuals is required, an integrative model that subsumes all or most of the contributors to human performance capacities and limitations is needed. Although such quantitative integrative models have rarely been sought by psychologists in the past, there is now a growing body of relevant literature. In the few cases in which these integrative models have been applied to military simulations, they have focused on specific task domains. Future integrative model developers will likely be interested in other domains than the ones illustrated here. Because each of our examples embodies a software approach to integrative models, we are referring to these developments as integrative modeling architectures. They provide a framework with behavioral content that shows promise of providing a starting point for model developers who wish to apply it to their domains. Each has its own strengths and weaknesses and we do not explicitly recommend any one. The developers' choices depend entirely on their goals and objectives in integrative model development. In Chapter 13, we do not recommend converging on a single integrative architecture, although we do argue that it is important to adopt modular structures that will allow easier interoperability among developed models.

OCR for page 51
Modeling Human and Organizational Behavior: Application to Military Simulations The chapter begins with a general introduction to integrative architectures that review their various components; this discussion is couched in terms of a stage model of information processing. Next we review 10 such architectures, describing the purpose, assumptions, architecture and functionality, operation, current implementation, support environment, validation, and applicability of each. The following section compares these architectures across a number of dimensions. This is followed by a brief discussion of hybrid architectures as a possible research path. The final section presents conclusions and goals in the area of integrative architectures. GENERAL INTRODUCTION TO INTEGRATIVE ARCHITECTURES A general assumption underlying most if not all of the integrative architectures and tools reviewed in this chapter is that the human can be viewed as an information processor, or an information input/output system. In particular, most of the models examined are specific instantiations of a modified stage model of human information processing. The modified stage model is based on the classic stage model of human information processing (e.g., Broadbent, 1958). The example given here is adapted from Wickens (1992:17). Such a stage model is by no means the only representation of human information processing as a whole, but it is satisfactory for our purposes of introducing the major elements of the architectures to be discussed. In the modified stage model, sensing and perception models transform representations of external stimulus energy into internal representations that can be operated on by cognitive processes (see Figure 3.1). Memory consists of two components. Working memory holds information temporarily for cognitive processing (see below). Long-term memory is the functional component responsible for holding large amounts of information for long periods of time. (See also Chapter 5.) Cognition encompasses a wide range of information processing functions. Situation awareness refers to the modeled individual combatant's state of knowledge about the environment, including such aspects as terrain, the combatant's own position, the position and status of friendly and hostile forces, and so on (see also Chapter 7). Situation assessment is the process of achieving that state of knowledge. A mental model is the representation in short- and long-term memory of information obtained from the environment. Multitasking models the process of managing multiple, concurrent tasks (see also Chapter 4). Learning models the process of altering knowledge, factual or procedural (see also Chapter 5). Decision making models the process of generating and selecting alternatives (see also Chapter 6). Motor behavior, broadly speaking, models the functions performed by the neuromuscular system to carry out the physical actions selected by the above-mentioned processes. Planning, decision making, and other "invisible" cognitive

OCR for page 51
Modeling Human and Organizational Behavior: Application to Military Simulations FIGURE 3.1 Modified stage model. NOTE: Tasks shown are derived from the vignette presented in Chapter 2. behaviors are ultimately manifested in observable behaviors that must be simulated with varying degrees of realism, depending on actual applications; aspects of this realism include response delays, speed/accuracy tradeoffs, and anthropometric considerations, to name but a few. Each of the architectures reviewed incorporates these components to some extent. What is common to the architectures is not only their inclusion of submodels of human behavior (sensing, perception, cognition, and so on), but also the integration of the submodels into a large and coherent framework. It would be possible to bring a set of specific submodels of human behavior together in an ad hoc manner, with little thought to how they interact and the emergent properties that result. But such a model would not be an integrative architecture. On the contrary, each of these integrative architectures in some way

OCR for page 51
Modeling Human and Organizational Behavior: Application to Military Simulations reflects or instantiates a unified theory of human behavior (at least in the minds of its developers) in which related submodels interact through common representations of intermediate information processing results, and in which consistent conceptual representations and similar tools are used throughout. REVIEW OF INTEGRATIVE ARCHITECTURES This section reviews 11 integrative architectures (presented in alphabetical order): Adaptive control of thought (ACT-R) COGnition as a NEtwork of Tasks (COGNET) Executive-process interactive control (EPIC) Human operator simulator (HOS) Micro Saint Man machine integrated design and analysis system (MIDAS) MIDAS redesign Neural networks Operator model architecture (OMAR) Situation awareness model for pilot-in-the-loop evaluation (SAMPLE) Soar The following aspects of each architecture are addressed: Its purpose and use Its general underlying assumptions Its architecture and functionality Its operation Features of its current implementation Its support environment The extent to which it has been validated The panel's assessment of its applicability for military simulations It should be noted that the discussion of these models is based on documentation available to the panel at the time of writing. Most of the architectures are still in development and are likely to change—perhaps in very fundamental ways. The discussion here is intended to serve as a starting point for understanding the structure, function, and potential usefulness of the architectures. The organizations responsible for their development should be contacted for more detailed and timely information. Adaptive Control of Thought (ACT-R)1 ACT-R is a "hybrid" cognitive architecture that aspires to provide an integrated 1   This section draws heavily from Anderson (1993).

OCR for page 51
Modeling Human and Organizational Behavior: Application to Military Simulations account of many aspects of human cognition (see also the later section on hybrid architectures). It is a successor to previous ACT production-system theories (Anderson, 1976, 1983), with emphasis on activation-based processing as the mechanism for relating a production system to a declarative memory (see also Chapter 5). Purpose and Use ACT-R as originally developed (Anderson, 1993) was a model of higher-level cognition. That model has been applied to modeling domains such as Tower of Hanoi, mathematical problem solving in the classroom, navigation in a computer maze, computer programming, human memory, learning, and other tasks. Recently, a theory of vision and motor movement was added to the basic cognitive capability (Byrne and Anderson, 1998), so theoretically sound interaction with an external environment can now be implemented. Assumptions In general, ACT-R adheres to the assumptions inherent in the modified stage model (Figure 3.1), with the minor exception that all processors, including the motor processors, communicate through the contents of working memory (not directly from cognition). ACT-R assumes that there are two types of knowledge—declarative and procedural—and that these are architecturally distinct. Declarative knowledge is represented in terms of chunks (Miller, 1956; Servan-Schreiber, 1991), which are schema-like structures consisting of an isa pointer specifying their category and some number of additional pointers encoding their contents. Procedural knowledge is represented in production rules. ACT-R's pattern-matching facility allows partial matches between the conditions of productions and chunks in declarative memory (Anderson et al., 1996). Both declarative and procedural knowledge exist permanently in long-term memory. Working memory is that portion of declarative knowledge that is currently active. Thus, the limitation on working memory capacity in ACT-R concerns access to declarative knowledge, not the capacity of declarative knowledge. ACT-R assumes several learning mechanisms. New declarative chunks can be learned from the outside world or as the result of problem solving. Associations between declarative memory elements can be tuned through experience. New productions can be learned through analogy to old procedural knowledge. Production strengths change through experience. The visual attention model embedded within ACT-R assumes a synthesis of the spotlight metaphor of Posner (1980), the feature-synthesis model of Treisman (Treisman and Sato, 1990), and the attentional model of Wolfe (1994). Details of the ACT-R architecture have been strongly guided by the rational analysis of Anderson (1990). As a consequence of that rational analysis, ACT-R

OCR for page 51
Modeling Human and Organizational Behavior: Application to Military Simulations is a production system tuned to perform adaptively given the statistical structure of the environment. Architecture and Functionality ACT-R's declarative and procedural knowledge work together as follows. All production rules in ACT-R have the basic character of responding to some goal (encoded as a special type of declarative memory element), retrieving information from declarative memory, and possibly taking some action or setting a subgoal. In ACT-R, cognition proceeds forward step by step through the firing of such production rules. All production instantiations are matched against declarative memory in parallel, but the ACT-R architecture requires only one of these candidate productions to fire on every cycle. To choose between competing productions, ACT-R has a conflict-resolution system that explicitly tries to minimize computational cost while still firing the production rule most likely to lead to the best result. The time needed to match a production depends on an intricate relationship among the strength of the production, the complexity of its conditions, and the level of activation of the matching declarative memory elements. In addition to taking different times to match, productions also differ in their contribution to the success of the overall task. This value is a relationship among the probability that the production will lead to the goal state, the cost of achieving the goal by this means, and the value of the goal itself. The architecture chooses a single production by satisfying: when the expected cost of continuing the match process exceeds the expected value of the next retrieved production, the instantiation process is halted, and the production that matches the highest value is chosen. To implement this conflict-resolution scheme, ACT-R models require many numerical parameters. Although most early ACT-R models set these parameters by matching to the data they were trying to explain, more recent models have been able to use stable initial estimates (Anderson and Lebiere, forthcoming). As mentioned above, ACT-R includes several learning mechanisms. Declarative knowledge structures can be created as the encoding of external events (e.g., reading from a screen) or created in the action side of a production. The base-level activation of a declarative knowledge element can also be learned automatically. Associative learning can automatically adjust the strength of association between declarative memory elements. ACT-R learns new procedural knowledge (productions) through inductive inferences from existing procedural knowledge and worked examples. Finally, production rules are tuned through 2   In some ways, the ACT-R implementation is intended to be a tool for exploring architectural assumptions. One method to support such exploration is the ability to turn each learning mechanism on or off independently. ACT-R is implemented in Lisp as an invitation to change the architecture.

OCR for page 51
Modeling Human and Organizational Behavior: Application to Military Simulations learning strengths and updating of the estimates of success-probability and cost parameters. All of these learning mechanisms can be turned on and off depending on the needs of the model. 2 Operation Most existing ACT-R models stand alone; all of the action is cognitive, while perception and motor behavior are finessed. However, some models have been built that interact with an external world implemented in Macintosh Common Lisp or HyperCard™. To our knowledge, no ACT-R models currently interact with systems not implemented by the modelers themselves (e.g., with a commercially available system such as Flight Simulator™ or a military system). Initial knowledge, both declarative and procedural, is hand-coded by the human modeler, along with initial numerical parameters for the strength of productions, the cost and probability of success of productions with respect to a goal, the base-level activation of declarative knowledge structures, and the like. (Default values for these parameters are also available, but not assumed to work for every new task modeled.) The output of an ACT-R model is the trace of productions that fire, the way they changed working memory, and the details of what declarative knowledge was used by those productions. If learning is turned on, additional outputs include final parameter settings, new declarative memory elements, and new productions; what the model learns is highly inspectable. Current Implementation The currently supported versions of ACT-R are ACT-R 3.0 and ACT-R 4.0. ACT-R 3.0 is an efficient reimplementation of the system distributed with Rules of the Mind (Anderson, 1993), while ACT-R 4.0 implements a successor theory described in Atomic Components of Thought (Anderson and Lebiere, 1998). Since both systems are written in Common Lisp, they are easily extensible and can run without modification on any Common Lisp implementation for Macintosh, UNIX, and DOS/Windows platforms. ACT-R models can run up to 100 times as fast as real time on current desktop computers, depending on the complexity of the task. Support Environment In addition to the fully functional and portable implementation of the ACT-R system, a number of tools are available. There is a graphical environment for the development of ACT-R models, including a structured editor; inspecting, tracing, and debugging tools; and built-in tutoring support for beginners. A perceptual/motor layer extending ACT-R's theory of cognition to perception and action is also available. This system, called ACT-R/PM, consists of a number of modules for visual and auditory perception, motor action, and speech production,

OCR for page 51
Modeling Human and Organizational Behavior: Application to Military Simulations which can be added in modular fashion to the basic ACT-R system. Both ACT-R and ACT-R/PM are currently available only for the Macintosh, but there are plans to port them to the Windows platform or to some platform-independent format, such as CLIM. The basic system and the additional tools are fully documented, and manuals are available both in MS Word and over the World Wide Web. The Web manuals are integrated using a concept-based system called Interbook, with a tutorial guiding beginners through the theory and practice of ACT-R modeling in 10 lessons. The Web-based tutorial is used every year for teaching ACT-R modeling to students and researchers as part of classes at Carnegie Mellon University (CMU) and other universities, as well as summer schools at CMU and in Europe. The yearly summer school at CMU is coupled with a workshop in which ACT-R researchers can present their work and discuss future developments. The ACT-R community also uses an electronic mailing list to announce new software releases and papers and discuss related issues. Finally, the ACT-R Web site (http://act.psy.cmu.edu) acts as a centralized source of information, allowing users to download the software, access the Web-based tutorial and manuals, consult papers, search the mailing list archive, exchange models, and even run ACT-R models over the Web. The latter capacity is provided by the ACT-R-on-the-Web server, which can run any number of independent ACT-R models in parallel, allowing even beginners to run ACT-R models over the Web without downloading or installing ACT-R. Validation ACT-R has been evaluated extensively as a cognitive architecture against human behavior and learning in a wide variety of tasks. Although a comprehensive list of ACT-R models and their validation is beyond the scope of this chapter, the most complete sources of validation data and references to archival publications are Anderson's series of books on the successive versions of ACT (Anderson, 1983, 1990, 1993; Anderson and Lebiere, 1998). The latter reference also contains a detailed comparison of four cognitive architectures: ACT-R, executive-process interactive control (EPIC), Soar, and CAPS (a less well-known neural-cognitive architecture not reviewed here). Applicability for Military Simulations The vast majority of ACT-R models have been for relatively small problem solving or memory tasks. However, there is nothing in principle that prevents ACT-R from being applicable to military simulations. There may be some problems associated with scaling up the current implementation to extremely large tasks that require extensive knowledge (only because the architecture has not been pushed in this manner), but any efficiency problems could presumably be

OCR for page 51
Modeling Human and Organizational Behavior: Application to Military Simulations solved with optimized matchers and other algorithmic or software engineering techniques such as those that have been applied to Soar. The intent of some currently funded military research contracts (through the Office of Naval Research) is to use ACT-R to model the tasks of the tactical action officer in submarines (Gray et al., 1997) and of radar operators on Aegis-like military ships (Marshall, 1995), but the work is too preliminary at this time to be reported in the archival literature. Although ACT-R's applicability to military simulations is as yet underdeveloped, ACT-R is highly applicable to military training in cognitive skill. ACT-R-based intelligent tutors (called cognitive tutors) are being used to teach high-school-level mathematics and computer programming in many schools around the world, including the DoD schools for the children of military personnel in Germany. These cognitive tutors reliably increase the math SAT scores of students by one standard deviation. Any cognitive skill the military currently teaches, such as the operation of a dedicated tactical workstation, could be built into a cognitive tutor for delivery anywhere (e.g., on-board training for Navy personnel). COGnition as a NEtwork of Tasks (COGNET) COGNET is a framework for creating and exercising models of human operators engaged in primarily cognitive (as opposed to psychomotor) tasks (Zachary et al., 1992; Zachary et al., 1996). Purpose and Use COGNET's primary use is for developing user models for intelligent interfaces. It has also been used to model surrogate operators and opponents in submarine warfare simulators. Assumptions COGNET allows the creation of models of cognitive behavior and is not designed for modeling psychomotor behavior. The most important assumption behind COGNET is that humans perform multiple tasks in parallel. These tasks compete for the human's attention, but ultimately combine to solve an overall information processing problem. COGNET is based on a theory of weak task concurrence, in which there are at any one time several tasks in various states of completion, though only one of these tasks is executing. That is, COGNET assumes serial processing with rapid attention switching, which gives the overall appearance of true parallelism. The basis for the management of multiple, competing tasks in COGNET is a pandemonium metaphor of cognitive processes composed of ''shrieking demons,"

OCR for page 51
Modeling Human and Organizational Behavior: Application to Military Simulations proposed by Selfridge (1959). In this metaphor, a task competing for attention is a demon whose shrieks vary in loudness depending on the problem context. The louder a demon shrieks, the more likely it is to get attention. At any given time, the demon shrieking loudest is the focus of attention and is permitted to execute. Architecture and Functionality The COGNET architecture consists of a problem context, a perception process, tasks, a trigger evaluation process, an attention focus manager, a task execution process, and an action effector. It is a layered system in which an outer shell serves as the interface between the COGNET model and other components of the larger simulation. There is no explicit environment representation in COGNET. Rather, COGNET interfaces with an external environment representation through its shell. The problem context is a multipanel blackboard that serves as a common problem representation and a means of communication and coordination among tasks (see below). It provides problem context information to tasks. Its panels represent different parts of the information processing problem and may be divided into areas corresponding to different levels of abstraction of these problem parts. Perception is modeled in COGNET using a perception process consisting of perceptual demons. These software modules recognize perceptual events in the simulated environment and post information about them (e.g., messages and hypotheses) on the blackboard. Tasks in COGNET are independent problem solving agents. Each task has a set of trigger conditions. When those conditions are satisfied, the task is activated and eligible to execute. An activated task competes for attention based on the priority of its associated goal (i.e., the loudness of its shrieks in the "shrieking demon" metaphor). Its priority, in turn, is based on specific blackboard content. Task behaviors are defined by a procedural representation called COGNET executable language. COGNET executable language is a text-based language, but there is also a graphical form (graphical COGNET representation). COGNET executable language (and by extension, graphical COGNET representation) contains a set of primitives supporting information processing behavior. A goal is defined by a name and a set of conditions specifying the requirements for the goal to become active (relevant) and be satisfied. Goals may be composed of subgoals. Four types of COGNET operators support the information processing and control behaviors of COGNET models. System environment operators are used to activate a workstation function, select an object in the environment (e.g., on a display), enter information into an input device, and communicate. Directed perceptual operators obtain information from displays, controls, and other sources. Cognitive operators post and unpost and transform objects on the blackboard.

OCR for page 51
Modeling Human and Organizational Behavior: Application to Military Simulations Control operators suspend a task until the specified condition exists and subrogate (turn over control) to other tasks. COGNET executive language also includes conditionals and selection rules, which are used to express branching and iteration at the level of the lowest goals. Conditionals include IF and REPEAT constructs. Selection rules (IF condition THEN action) can be deterministic (the action is performed unconditionally when the condition is true) or probabilistic (the action is performed with some probability when the condition is true). The COGNET trigger evaluation process monitors the blackboard to determine which tasks have their trigger conditions satisfied. It activates triggered tasks so they can compete for attention. The attention focus manager monitors task priorities and controls the focus of attention by controlling task state. A task may be executing (that is, its COGNET executive language procedure is being performed) or not executing. A task that has begun executing but is preempted by a higher-priority task is said to be interrupted. Sufficient information about an interrupted task is stored so that the task can be resumed when priority conditions permit. Using this method, the attention focus manager starts, interrupts, and resumes tasks on the basis of priorities. As mentioned above, COGNET assumes serial behavior with rapid attention switching, so only the highest-priority task runs at any time. Tasks are executed by the task execution process , which is controlled by the attention focus manager. The action effector changes the environment. Operation The operation of a COGNET model can be described by the following general example, adapted from Zachary et al. (1992): Perceptual demon recognizes event, posts information on blackboard. New blackboard contents satisfy triggering condition for high-level goal of task B, and that task is activated and gains the focus of attention. B subrogates (explicitly turns control over) to task A for more localized or complementary analysis. A reads information from blackboard, makes an inference, and posts this new information on the blackboard. New information satisfies triggering condition for task D, but since it lacks sufficient priority, cannot take control. Instead, new information satisfies triggering condition for higher-priority task C, which takes over. C posts new information, then suspends itself to wait for its actions to take effect. Task D takes control and begins posting information to blackboard. etc.

OCR for page 51
Modeling Human and Organizational Behavior: Application to Military Simulations This page in the original is blank.

OCR for page 51
Modeling Human and Organizational Behavior: Application to Military Simulations   Multitasking Architecture Serial/Parallel Resource Representation Goal/Task Management Multiple Human Modeling Implementation Platform ACT-R Serial with switching (no interruptions) Amount of declarative memory activation None (one goal at a time) Potential through multiple ACT models Mac, PC (limited) COGNET Serial with switching and interruptions Limited focus on attention, parallel motor/perceptual processors Pandemonium model—priorities based on task context Through multiple COGNET models PC, SGI Sun EPIC Parallel Limited perceptual and motor processors, unlimited cognitive processor Priority-based task deferment No Lisp platforms HOS Serial with switching, plus parallel movement possible Speech, hand, foot, and cognitive channels Priority-based attention switching No PC Micro Saint based network tools Parallel with switching to serial resources limited Visual, auditory, cognitive, psychomotor workload Simple dynamic prioritization Yes PC, Mac, Unix platforms MIDAS Resource-limited parallel Visual, auditory, cognitive, motor resources Z-Scheduler uses time, resource constraints Limited SGI MIDAS Redesign Resource-limited parallel Visual, auditory, cognitive, motor resources Agenda Manager uses time and resource constraints, goal priorities Yes   Neural network based tools Contention scheduling via competition among components Units and connections allocated to a task Competitive systems of activation Potential through multiple networks MAC, PC, Unix platforms OMAR Serial with some parallelism for automatic tasks Perceptor, cognitive, effector resources Tasks compute own priorities Yes Sun, SGI, PC (future) SAMPLE Serial with interruptions Sensory, cognition, and action channels Priority based on base value + ''situational relevance" Yes PC, Unix Soar Serial with switching and interruptions Serial cognitive processor, limited perceptual and motor resources Preference-based attention allocation Yes Mac, PC, Unix platforms

OCR for page 51
Modeling Human and Organizational Behavior: Application to Military Simulations This page in the original is blank.

OCR for page 51
Modeling Human and Organizational Behavior: Application to Military Simulations Architecture Language Support Environment Validation Comments ACT-R Lisp Editors, debuggers Extensive at many levels ACT-R models focus on single, specific information processing tasks; has not yet been scaled up to complex multitasking situations or high-knowledge domains COGNET C++ Graphical editors, syntax/semantic checkers, debuggers, structural SPI to external processes Full model Used in low-fidelity submarine training simulations and high-fidelity AEGIS CIC training simulations to provide surrogate operators and adversaries EPIC Lisp Lisp programming environment tools Extensive for reaction time EPIC models focus on simple, dual-task situations; has not yet been scaled up to complex multitasking situations or high-knowledge domains HOS FORTRAN Editors (mostly text-based) Components (micro-models) Currently capable of scripted behaviors only Micro Saint based network tools C, C++ Editors, debuggers Some micro-models; at least one implementation Used extensively in military simulations MIDAS Lisp, C, C++ Graphical editors, graphical data displays Full model Currently capable of scripted behaviors only MIDAS Redesign C++ Similar to original MIDAS None In development Neural network based tools C, NETLAB Commercial products, GUIS Extensive for component tasks Focus on sensory/motor integration, have not yet been scaled up to complex multitasking situations or high-knowledge domains OMAR Lisp Compilers, editors, browsers, online animation tools, post-run analysis tools Components   SAMPLE C++ Editors Control tasks (OCM) Has been used in small-scale military simulations Soar C Editors, debuggers Extensive at multiple levels Has been used in military simulations (e.g., synthetic theater of war-Europe [STOW-E], STOW-97)

OCR for page 51
Modeling Human and Organizational Behavior: Application to Military Simulations This page in the original is blank.

OCR for page 51
Modeling Human and Organizational Behavior: Application to Military Simulations KATE, a highly-simplified stick figure mannequin that can execute OMAR tasks. Many of the architectures can be extended by building custom motor modules. Most of the architectures generate outputs in the form of behaviors of the human being represented (e.g., motions, gestures, utterances) that can or could be used as inputs to other simulation modules, such as tank or aircraft models. HOS (in its original version) and Micro Saint, however, generate performance metrics, such as task start and end times and task accuracy data. The former are more applicable for virtual simulations, which need to generate visible images of human and system behaviors, while the latter are more applicable to constructive simulations. In principle, however, it should be possible to modify the two systems to produce such behaviors; in fact, as discussed earlier, modified HOS micro-models have recently been added to COGNET for that very purpose. Knowledge Representation Declarative Representation of declarative knowledge ranges from simple variables to complex frames and schemas. Those applications in which complex factual knowledge structures must be represented explicitly would be better served by the more sophisticated techniques of such architectures as ACT-R and OMAR. Procedural All the architectures, except neural net architectures, have a means of representing procedural knowledge beyond that provided by the languages in which they are written. The production rule languages of ACT-R and Soar appear to be the most flexible and powerful, though the related complexity of such languages may not be warranted for those applications in which behavior is highly procedural (perhaps even to the point of being scripted). Higher-Level Cognitive Functions Learning Only three architectures offer learning: ACT-R, neural net architectures, and Soar. Of these, ACT-R provides the most flexibility. Situation Assessment In most of the architectures, situation assessment is overt, in the sense that

OCR for page 51
Modeling Human and Organizational Behavior: Application to Military Simulations the programmer/modeler must explicitly code all its aspects. Some of the architectures, however, have the capability of inferring details of the situation from other information. This capability obviously reduces the programmer's burden, provided the mechanism is suitable for the application. Planning HOS and Micro Saint do not provide a planning function (though presumably the programmer can implement it). Several of the architectures (e.g., MIDAS) have the capability of instantiating general plans based on the specifics of a given situation. ACT-R and Soar models can create new plans. This capability is obviously valuable for some applications; however, for those applications not requiring planning (i.e., where procedures can be initiated in a prespecified way), the development and computational overhead may be undesirable. Decision Making HOS does not provide decision making capability, except to account for the time required for the operator to make a decision given certain characteristics of the decision problem. Most of the other architectures provide for knowledge-based decision making, with a few offering Bayesian techniques. Some form of decision making capability seems essential for most military simulations, and architectures such as ACT-R, OMAR, and Soar seem most capable in this regard. Multitasking All the architectures permit multitasking in some sense, though ACT-R enforces strict serial execution of tasks with no interruptions allowed. The degree of multitasking varies from that of HOS, which allows only a single cognitive task (with parallel, ballistic body motion possible) to EPIC, which permits any number of concurrent tasks provided resource requirements are not exceeded. From the standpoint of multitasking flexibility, then, architectures such as EPIC and MIDAS appear to offer the greatest potential. However, there is disagreement over the psychological validity of such systems, and architectures such as OMAR that limit the operator to only one task requiring conscious thought may be more realistic. It is clear, however, that interruptability is an important feature; all architectures, except ACT-R, have this feature. Multiple Human Modeling In principle, multiple copies of any of the architectures could be integrated to allow modeling of multiple operators. However, only about half of the architectures

OCR for page 51
Modeling Human and Organizational Behavior: Application to Military Simulations explicitly provide this capability. It is worth noting that this capability was a major driving force in the development of OMAR and the MIDAS redesign. Implementation Platform Most of the architectures have been implemented on one or only a few computer platforms. COGNET and OMAR are more flexible, and Micro Saint and Soar are available for almost any platform a user could require. Language For those users needing to modify the architectures to fit the needs of their applications, the language in which the architectures are written is of some concern. To many military simulation modelers who have done most of their work in FORTRAN and C, architectures such as ACT-R, EPIC, and OMAR (written in Lisp) may pose some (small) problems. Support Environment The software provided to assist the modeler in creating models in the various architectures ranges widely. At one extreme is EPIC, for which there exists very little such software; at the other extreme is OMAR, whose toolkit appears well thought out and tailored to the user not intimately familiar with OMAR's inner workings. To their credit, though, ACT-R and Soar both have a very large base of users, many of whom have developed additional tools to facilitate the use of the two systems. Validation Some of the architectures are too new for any substantive validation to have been done. In other cases (e.g., HOS), certain submodels have been extensively validated. Some of the architectures have been validated in their entirety ("full model" validation in Table 3.1), some of these extensively. At least one (Micro Saint) has instances that have received military accreditation. Unfortunately, most of the validation of "full models" has been based on subjective assessments of subject matter experts, not real human performance data. This last consideration dictates considerable caution on the part of users. HYBRID ARCHITECTURES: A POSSIBLE RESEARCH PATH As discussed above, the various architectures have different strengths and

OCR for page 51
Modeling Human and Organizational Behavior: Application to Military Simulations weaknesses with respect to their relevance for military simulations. To model all the details of the complex behavior involved in the tank platoon's hasty defense described in the vignette of Chapter 2, an architecture would have to encompass all the phenomena addressed in all the following chapters: attention and multitasking, memory and learning, decision making, situation assessment, and planning, all subject to behavioral moderators. No existing architectures approach this all-encompassing capability. This does not make these architectures useless—many have been demonstrated to be quite useful despite their limitations—but it does suggest areas of improvement for each. One research and development path is obvious: start with one of the existing architectures, and expand it in the areas in which it is weak. For example, an emerging ACT architecture, ACT-R/PM, adds EPIC-like perception and motor capabilities to the ACT-R architecture (Byrne and Anderson, 1998). Each of the reviewed architectures could be expanded in this way, as needed for task performance. This approach represents a time-honored path with almost guaranteed incremental payoff, but may eventually encounter boundaries as the architectures reach their limits of expressibility. Another research and development path might prove more fruitful: combine the strengths of two or more architectures to produce a hybrid that better encompasses human phenomena. A simple example, the combination of Soar's cognitive processor with EPIC's perceptual and motor processors, has already been mentioned. In contrast with ACT-R/PM's reimplementation of EPIC-like processors in the ACT architecture itself, neither Soar nor EPIC was rewritten, but communicate through a shared working memory. More fundamental combinations of architectures are the subject of an ongoing basic research program at the Office of Naval Research (Hybrid Architectures as Models of Human Learning), which supported the infancy of several hybrid architectures. To address the effects of environmental frequency in Soar, that architecture was combined with Echo (Thagard, 1989), a statistical technique for belief updating. Neural nets were augmented with a symbolic explanation-based learning system (Dietterich and Flann, 1997) to address the learning of long procedures when the effects of actions are widely separated from the actions, credit and blame are difficult to assign, and the right combination of moves should count more toward learning than the myriad of failures along the way. The CLARION architecture (Sun, 1995) integrates reactive routines, generic rules, learning, and decision making to develop versatile agents that learn in situated contexts and generalize resulting knowledge to different environments. Gordan (1995) extends Marshall's schema theory of human learning, supplementing its high-level planner with a low-level stimulus-response capability. Marshall (1995) is using a neural net for the identification component of schema theory and ACT-R for the elaboration, planning, and execution components. Cohen and Thompson (1995) use a localist/connectionist model to support rapid recognitional domain reasoning, a distributed

OCR for page 51
Modeling Human and Organizational Behavior: Application to Military Simulations connectionist architecture for the learning of metacognitive behavior, and a symbolic subsystem for the control of inferencing that mediates between the domain and metacognitive systems. These preliminary architectural explorations are promising examples of an area ripe for research. Through extensive use of the reviewed architectures, researchers and developers in the field now know enough about their relative strengths and weaknesses to judiciously explore plausible combinations for achieving the greatest impact of human behavior models. CONCLUSIONS AND GOALS None of the architectures described in this chapter is precisely what is needed for military simulations. Collectively, however, they offer a foundation on which to build models that will be truly useful and practical for military simulations. Once thorough task analyses of the combatant's activities have been conducted to determine exactly what behaviors are to be modeled, these architectures offer frameworks and a wide variety of detailed models of specific perceptual, cognitive, and motor processes that can be used to represent these behaviors. Elements of the various architectures could be combined to yield representations with psychological validity, model fidelity, and computational efficiency. The problem with these architectures is not what they do not model. Among the architectures reviewed here, there is hardly an interesting and potentially useful phenomenon that is not considered in some way. The problem is that the architectures are not validated. Many of these architectures have submodels that are well validated and have reached a level of maturity suitable for application in military simulations. However, few of the architectures have been validated overall, and their emergent properties are not well understood. More experience with them is needed. A careful, measured expansion of their application in military simulations may just be the proper path to take. Short-Term Goals Make validation of integrative architectures a priority. Continued use of any integrative architecture or tool should be predicated on its validation. Early effort will be required to define performance measures and standards for use in the validation, as well as scenarios to be used as test cases. Data from exercises and human-in-the-loop simulations can be collected in preparation for validation of models to be developed, and human behavior models currently used in military simulations can be tested against these data. As a first cut at upgrading these architectures, gradually incorporate the concepts, theories, and tools presented in subsequent chapters of this report into existing simulations to incrementally improve existing models. One approach to this end would be to use a simplified stage model to augment current human

OCR for page 51
Modeling Human and Organizational Behavior: Application to Military Simulations behavior representations. Such an augmented representation should incorporate, at a minimum, a module representing human perceptual processes, a module representing motor processes, a means for representing multiple active tasks (e.g., planning, decision making, communicating, moving), and a mechanism for selecting a subset of active tasks for execution at any given time. The perceptual module could be based on simple detection and identification probability data derived from existing psychological data. Similarly, the motor process module could be based on simple movement models that yield time and accuracy estimates, or on time and accuracy probabilities derived from human performance data. All current human behavior representations have some means of representing tasks, and most of these could be modified to permit the existence of more than one task at a time. The mechanism for selecting among competing tasks could be based on a simple, static priority system with task priorities derived from task analyses. Such an approach would yield simple architectures with greater face validity than that of current architectures. Although it is unlikely that such architectures would be entirely satisfactory, they would give modelers more experience in developing and validating human behavior representations. Intermediate-Term Goals Continue validation into the intermediate term as more sophisticated integrative architectures are developed. This generation of architectures can be expected to draw on the integrative architectures reviewed in this chapter, as well as newer architectures that will emerge after this report is published. Continue the development of hybrid architectures, such as those described in this chapter, combining the best elements of existing and emerging integrative architectures. For such hybridization to proceed in a timely manner, it will also be necessary to conduct research and development activities to modularize existing architectures and yield interchangeable components. Apply these architectures in sustained and intensive development of human behavior representations that incorporate specific military tasks in selected domains, such as tank warfare. Compare the different modeling approaches by developing alternative architectures for a domain and comparing them against data from field exercises and human-in-the-loop simulations. Long-Term Goals Continue architecture validation. Continue to refine new architectures created in the intermediate term. In addition to continued efforts to improve the quality of existing modeling approaches, explore entirely new approaches that will result in architectures as yet unconceived.