Setting up Large - Scale Qualitative Modelst

A qualitative physics which captures the depth and breadth of an engineer's knowledge will be orders of magnitude larger than the models of today's qualitative physics . To build and use such models effectively requires explicit modeling assumptions to manage complexity . This, in turn, gives rise to the problem of selecting the right qualitative model for some purpose . This paper addresses these issues by describing a set of conventions for modeling assumptions . Simplifying assumptions decompose a domain into different grain sizes and perspectives which may be reasoned about separately . Operating assumptions reduce the complexity of qualitative simulation by focusing on particular behaviors of interest . We show how these assumptions can be directly represented in Qualitative Process theory, using a multi-grain, multi-slice model of a Navy propulsion plant for illustration . Importantly, we show that model selection can often be performed automatically via partial instantiation . We illustrate this technique with a simple explanation generation program that uses the propulsion plant model to answer questions about physical and functional characteristics of its operation . A long-range goal of qualitative physics is to develop systematic models that capture the breadth and depth of human reasoning about the physical world . Such models will be crucial for future intelligent computer-aided design and tutoring systems . Clearly, they will need to be orders of magnitude larger than today's models . Furthermore, they must capture phenomena at several levels of detail, and from a variety of perspectives . Building such models raises several new issues for qualitative modeling : 1 . Organization problem : How can we organize a model that captures phenomena at a variety of grain sizes and perspectives? 2 . Relevance problem : Generating all possible states becomes intractable as the size of system modeled grows . Even if we could generate them all, often we only care about a subset of the behavior . How can we use qualitative simulation in a more focused way to answer questions? 3 . Selection problem : As models get larger, complete instantiation becomes both undesirable and impossible . No one understands spilling a cup of coffee via quantum mechanics . Furthermore, some of the perspectives in a model will be mutually incompatible . How can an appropriate subset of a model be selected for reasoning, given a particular question? This paper addresses each of these issues . In particular, we claim the key idea in solving all of them is a set of conventions for explicitly representing modeling assumptions . We introduce explicit simplifying assumptions to solve the organization problem by providing "scoping", delimit ing when descriptions are and are not applicable . We introduce operating assumptions to describe `This paper appears in Proceedings AAAI-88 Brian Falkenhainer Kenneth D . Forbus Qualitative Reasoning Group Department of Computer Science University of Illinois at Urbana-Champaign 1304 W. Springfield Avenue, Urbana, Illinois 61801 standard behaviors or default conditions . We illustrate how, using these conventions, the selection problem can in some cases be solved automatically via partial instantiation . These conventions are illustrated using a multi-grain, multiple perspective high-level model of a Navy propulsion plant . We demonstrate our solution to the model selection problem by showing how, in the context of a tutoring system, the form of a question can be analyzed so that .the appropriate set of modeling assumptions can be automatically computed . In the next section we outline our perspective on qualitative modeling, showing the need for explicit modeling assumptions to control model instantiation and use . Section 3 gives a brief tour of the steam plant and its qualitative model which provides our motivating example . Section 4 describes our conventions for modeling assumptions, and Section 5 shows how they are used to organize the steam plant model. Section 6 describes our algorithm for automatically computing a minimal set of simplifying assumptions for a given query. Finally, we discuss directions for future research . 2 The Modeling Process We call the system or situation being modeled the scenario, and its qualitative model the scenario model. The simplest way to build a scenario model is to create a model of that specific scenario for a particular purpose. While useful systems may be built this way, it is also easy to generate ad hoc models of dubious merit, where the model must be thrown away whenever the scenario or purpose changes slightly . An indirect route is more robust build first a general-purpose domain model, which describes a class of related phenomena or systems . Ideally, a scenario model can be built by instantiating and composing descriptions from the domain model . Developing a domain model requires more initial work, but it simplifies generating models for a range of scenarios . Furthermore, ad hoc aspects of models are more likely to be discovered if the same descriptions are re-used in a variety of settings . So far, we have stated the commonplace view of modeling in qualitative physics. Qualitative process theory f31 organizes domain models around processes, which can be automatically instantiated to form scenario models. Device-centered ontologies [1,141 provide catalogs of devices, which can be composed to build scenario models . (Kuiper's QSIM ',81 does not provide any abstraction or organizing structure for domain models itself, but one could imagine using it with either ontology .) Unfortunately, as we have attempted to build more realistic models, we have discovered that this view is inadequate . This view breaks down in two ways for complex domain models . First, higher fidelity models are simply bigger than lower fidelity models . Representing fluids in detail, for instance, requires geometric information about the placement of portals, descriptions of head at every distinguishable place, models of fluid resistance in pipes, and so forth. We have built such models, (which turn out to be several times larger than than the models in '31), and even on.simple situations they swamp our machines . Only part of the problem is technological . Even if our computers ran infinitely fast, for most purposes we simply don't need or want such detailed answers . When we do need the details, it is typically about a very narrow range of behaviors . This scaling problem becomes even more acute when faced with modeling the kind of propulsion plant studied in STEAMER 5j, which used a numerical model that contained hundreds of parameters . The stock AI answer is "hierarchy", but how should it be done? The second breakdown comes from the use of multiple perspectives . In some cases, a feed tank is best viewed as an infinite capacity liquid source . In other cases, it should be viewed as a container rte>5. ',_f0~X. . . . Feed Pump Figure 1 : Simplified model of a navy steam-powered propulsion plant . which may be emptied (perhaps with dire consequences) . One cannot consistently use both views at once . One solution would be to create multiple, distinct models, one for each perspective and purpose . Doing so would significantly raise the difficulty of the selection problem, and make knowledge acquisition and maintenance nearly impossible . We must find ways for incompatible perspectives to peacefully coexist in a single domain model . These issues have been addressed before in qualitative physics, albeit partially and informally. de Kleer and Brown, for instance, describe class-wide assumptions, which roughly correspond to our use of simplifying assumptions . However, this notion has never been formalized nor explicitly used as part of their programs or models 71 . So far, the device ontology in qualitative physics has inherited a limitation from System Dynamics 1101 upon which it is based : the process of mapping from the "real-world" scenario to a device model lies outside the theory . Qualitative Process theory was designed with such problems in mind . The descriptions of the domain model are automatically instantiated by a QP interpreter, thus in theory providing the means for modeling assumptions to be explicitly represented. This paper describes a set of conventions for exploiting this power . 3 A steam plant model Since steam plants are not everyday systems, we begin with a brief description of steam propulsion plants, and the highlights of our model . Figure 1 shows an abstract view of propulsion plants adapted directly from Navy training materials 191 . The primary components operate in the following fashion : " Boiler assembly . The boiler assembly takes in distilled water and fuel and produces superheated steam . Most surface ships use several boilers, but this can be ignored . The heat is supplied in most ships by an oil-buring furnace . The steam produced by the boiler is fed through the superheater, which increases its temperature in order to get more work out of it . Turbines . The superheated steam then enters the turbines, which produce work (by driving the ship's propellers), resulting in the temperature, pressure, and kinetic energy of the steam dropping . Condenser assembly . The steam exhausts from the turbine to the condenser, where it is cooled by circulating sea water and condensed again into liquid .