AI: what simulationists really need to know

ion. In human problem solving, abstraction is an important technique for managing complexity. One characterization of human expertise is the ability to make the most appropriate abstraction in a particular domain, domain situation, and problem-solving situation. Qualitative representations of system/mechanism primitive behaviors, constraints, state, and behavior are one dimension in which abstraction applies (as opposed to eliminating specific components from a system/mechanism and the associated variables from the state). This abstraction dimension is in fact very useftd, as demonACM TransactIons on Modehng and Computer Slmulatlon, Vol. 2, No. 4, October 1992 What Simulationists Really Need to Know . 275 strated in human reasoning and programs that reason from such qualitative representations. For example, digital circuit simulation abstracts the actual voltages that exist in the circuit to logic values O, 1, and X. Completeness. In deriving behavior via simulation, qualitative simulation (e.g., QSIM) is complete in that all possible behaviors are represented in the envisionment (assuming that generation of such an envisionment is tractable). For numerical simulation approaches, the same claim cannot be made. It should be understood that qualitative distinctions of behaviors are dependent on the specification of the system/mechanism (e.g., introducing a landmark into a variable’s quantity space can result in qualitatively distinct behaviors not observed before the landmark was added). This, however, is the price of abstraction. Operating with Incomplete Knowledge of the Domain. The qualitative model specification and simulation techniques developed in the AI qualitative-reasoning community have emphasized the ability to proceed in the face of incomplete knowledge (theory or model) of the system/mechanism and any initial conditions. This is exhibited not only in qualitative variable and state values, but also in the expression of primitive behaviors used in a system/ mechanism description. For example, QSIM provides monotonic increasing (M + ) and decreasing (M — ) constraints, and Qualitative Process Theory expresses influences between variables. The ability to develop a model and simulate it in the presence of incomplete knowledge is important in that some initial information can be collected and subsequently used in problem solving and model refinement. If we consider qualitative and numerical models and simulation techniques as points or areas on an abstraction spectrum, the problem of developing, validating, and maintaining theories about the domains of interest (either for humans or for autonomous agents) can be viewed as building and validating a theory at some point on the spectrum, and then possibly modifying the theory in the direction most appropriate for the task at hand (i.e., more or less abstract). For a design activity, the modification must necessarily go to a very fine level of detail so that the associated mechanism can be constructed. A diagnosis or explanation capability, however, may not require such a fine-grain description. In fact, for explanation or prediction purposes, a more abstract description is often appropriate (e.g., cyclic, or remains with limits). Issues in model construction and selection are an active area of research in the qualitative-modeling and model-based reasoning communities. The integration of quantitative and qualitative information is also being investigated. One of the central issues in AI has been knowledge representation. Issues of expressive power, tractability and completeness of inference procedures, and conceptual integrity with respect to the problem domain have guided research. The problem domain governs ontological issues for the objects examined in the problem-solving process (e.g., components of a mechanism, observations such as medical data, physical processes) as well as objects/ concepts of the problem-solving process itself (e.g., design goals, explanations). These representation issues (domain objects, problem-solving process ACM TransactIons on Modeling and Computer Simulation, Vol. 2, No. 4, October 1992. 276 . David P. Miller et al. concepts) plus the goals of the problem-solving technique provide an understanding of the current AI approaches to and uses of simulation (e.g., qualitative simulation). Particular choices are sometimes motivated by models of human problem solving, not with the goal of accurately modeling human problem-solving activity, but with the goal of giving programs better problem-solving capabilities. The goal of (AI’s) qualitative-modeling research has been much discussed, ranging from the desire to faithfully model human cognition to the ability to build and utilize precise, accurate models of the real world. To repeat an earlier message, I [D. W. Franke] believe that the appropriate context is the pragmatic one, in which the particular simulation or modeling approach can best be judged by (1) its ability to solve a particular problem and (2) the ability for humans or other programs (autonomous agents) to evaluate and utilize the results of the modeling. Unfortunately, many claims of the form “Yes, we use AI techniques” are made for systems and products. We must be careful in evaluating such claims and must examine the problem-solving capabilities as well as any implementation approaches. 4. PAUL A. FISHWICK: Al AND SIMULATION; SOME LESSONS LEARNED The fields of AI and simulation are fairly large in terms of literature and interdisciplinary tendencies. Discussing, therefore, how the two relate to one another is a formidable task; however, we have learned many key points or “lessons” especially during the AI and simulation workshops, conferences, and panel sessions over the past several years. In this section, I [P. A. Fishwick] will discuss some things that I have learned during my time studying the benefits of AI and simulation to each other. These “lessons learned” are personal reflections that have been gathered from verbal and email conversations, workshops, and literature searches. 4.1 Code A// the Knowledge Perhaps the chief contribution of AI to all fields, including simulation, is the realization that knowledge which is nonequational, quantitative, or precise in nature can still be used for useful problem solving. The primary example of this type of AI research is found within “expert systems,” which, from a problem-solving viewpoint, are not unique because they represent expert knowledge per se. After all, continuous models for aircraft flight or discreteevent models for assembly lines are also reservoirs of knowledge—specifically, “expert knowledge” about the principles of flight and the operation of assembly lines. What, then, makes an expert system unique? Expert systems have been built in those areas where models have been either very weak or nonexistent such as in medical diagnosis; we do not have a simple set of equations that accept symptoms as inputs and produce a correct diagnosis as an output. Mycin [3] provides an excellent example of a program that contains the deepest knowledge available in the domain for which Mycin was coded: the selection of antimicrobial drugs given specific symptoms of bacteACM Transactions on Modeling and Computer Simulation, Vol 2, No. 4, October 1992 What Simulationists Really Need to Know . 277 rial infection. Because we do not have such equations for the automatic calculation of drugs given medical symptoms, AI technology has suggested to us that models based on predicate calculus (of which expert system knowledge is a special case) are indeed useful if we can feed in inputs and obtain reasonable outputs. The AI approach suggests that we code all the knowledge that is available to us for our simulation models and not only that knowledge which yields to traditional forms of systems analysis. The systems problem-solving process is highly iterative; we start with high-level models and progress to complex models. The high-level knowledge that is coded within expert systems is usually of a “decision-making” or diagnostic type. How does this affect the field of computer simulation? It suggests that we code decision-making and planning components within our simulations. As simulations become more complex, they will contain autonomous objects, and simulation of these autonomous objects (such as robots, autonomous vehicles, and humans) will require the fruits of AI research in areas such as expert systems and mental modeling. 4.2 Integrate Qualitative and Quantitative Knowledge One of the problems in the area of AI and simulation is that many researchers have thought of expert systems (and other AI models) as being completely different than models that exist in simulation. Also, the concept of simulation as being inherently numerical has led to some perceived differences between the AI and simulation modeling efforts. These differences have caused a split in the two communities. Since the AI community is primarily concerned with qualitative knowledge and the simulation community with quantitative, the most recent fertile area of research in both groups involves integrating quantitative and qualitative knowledge. The chief difference between the AI and simulation efforts—with regard to the study of qualitative/quantitative integration-resolves around the treatment of uncertainty. In simulation, the term “qualitative” [8, 9] has often been equated with abstraction in terms of the abstraction level associated with systems [22, 27] and system components such as time, state, and event [20]. For instance, while continuous simulation provides us with a model for obtaining continuously changing state values, discrete-event simulation fosters discrete changes in state and event where the values may be either quantitative or qualitative. Mixtures of these two different types of models fall under the general category of combined modeling. By partitioning state space [10], we can formally map