Our PHYSSYS ontology intends to capture widely applicable concepts and background theories in physical systems engineering, an area which has stimulated quite a significant effort in ontology development in various application directions [Top & Akkermans, 1994, Neches et al., 1991, Alberts, 1993, Gruber & Olsen, 1994, van der Vet et al., 1995, Bernaras & Laresgoiti, 1996, Benjamin et al., 1996]. A feature of our approach is the postulated existence of different conceptual viewpoints on the domain objects and reasoning about them, that is, a grouping of properties of the domain objects into separate `natural' categories. This has the advantage that it leads to a corresponding strict separation of ontologies (describing these properties) as a basic structuring principle. The ontological viewpoints selected here -- technical system components, physical processes, mathematical descriptions -- have been adopted from [Top & Akkermans, 1994], but are much further worked out and operationalized in the present work. It is gratifying that the EngMath ontology for engineering mathematics [Gruber & Olsen, 1994] could be reused here and integrated into a wider physical systems ontology.
We do not at all want to imply that these are the only possible or relevant viewpoints. In our simulation experiment of a hospital heating system we have seen that the determination of exogenous model and design parameters often proceeds on the basis of geometric and material properties. This suggests to extend the set of ontologies. We are currently developing a geometry view, which has to be compatible with the already available topology theory, and can be linked to the relevant part of the EngMath ontology by means of an ontology projection. Spatial ontologies are discussed e.g. in [Cohn et al., 1995]. Also, work on separate ontologies for material properties is ongoing, such as the Plinius ontology [van der Vet et al., 1995]. In tasks like analyzing the environmental impact of alternative designs for engineering systems, modelling the material properties is a key issue.
In other work, we also find the idea of using different conceptual viewpoints to modularize ontologies, but usually in a much looser way. For example in the Ontolingua library, also a thermal systems ontology has been specified [Neches et al., 1991]. Basically, it contains a number of thermal components (such as `boiler'), for which then mathematical equations are given that specify the component behaviour. In our approach, such models are not part of the ontology, but are found in the OLMECO library. The PHYSSYS ontology rather specifies what such models should look like in general. Thus, our ontology expresses meta-level knowledge concerning modelling and simulation. This conforms to the view in [Schreiber et al., 1994, Schreiber et al., 1995], where ontologies are viewed as meta-level specifications of a set of possible domain theories or models.
This, by the way, does have practical consequences, since in our case there are clearly more constraints than in the KSE library on how components can or cannot be connected at the component level, and what implications follow for the assembly of mathematical models. Closest to our approach is probably the YMIR ontology of [Alberts, 1993], which is also based upon general systems theory. The systems part is essentially the same as the PHYSSYS component ontology, although it is not constructed out of smaller ontologies about mereology and topology. YMIR pays more attention than both PHYSSYS and EngMath to possible abstraction steps from larger to simpler models at the mathematical level. A major difference is the absence in YMIR of a process ontology. Like in the KSE physical systems library, physical processes are in fact equated with their mathematical descriptions. This is also a typical situation in AI qualitative reasoning frameworks that are device- and constraint-oriented, cf., [Kuipers, 1994] and references therein.
We have not made this choice for fundamental reasons: (i) it is common
in knowledge acquisition to encounter forms of conceptual or
qualitative reasoning (also by experts) about physical processes
without mathematics; (ii) in general the relationship between physical
processes and mathematical descriptions is n-to-n. Both our
ontology and the OLMECO library cater for this, leading to more
flexible modelling. Thus, the process ontology is a salient feature of
PHYSSYS. Forbus' qualitative process theory [Forbus, 1984]
and the associated modelling framework [Falkenhainer & Forbus, 1991] are
much in the same spirit, but there are important differences in the
underlying ontology. In contrast, the PHYSSYS process ontology
is formally built in a compositional way on a set of primitive
physical mechanisms, that in addition satisfy generic ontologies
concerning mereology, topology and (network) systems theory (see
Section 2). All this is left much more open (as well
as much more informal) in the ontology underlying QPT, resulting in
less commitment and less guidance. One side of the coin is that QPT
allows to specify processes according to, say, medieval, Aristotelian
or commonsense physics. This is not possible in PHYSSYS as it
commits to modern physical science
.
The other side of the same coin is that, due to this lack of
commitment, it is much easier in QPT to come up with nonsensical
process models. Here, PHYSSYS provides more physics knowledge
and guidance -- that is, according to current scientific standards.
A unique, strongly unifying, characteristic of PHYSSYS is that it formally specifies and exploits the analogies between different fields in physics. This makes it much more widely applicable and reusable than first selecting a physics subdomain (e.g. thermodynamics) and restricting the ontology to this subdomain as is usually done. This again has practical consequences: many modern engineering systems are inherently multidisciplinary --mechatronic systems but also heating systems are good examples-- and restricting ontologies in such situations to the standard physics subdomains not only hampers reusability but also usability.
In some ontologies for technical domains, see [Bernaras & Laresgoiti, 1996, Benjamin et al., 1996] and other ontologies from the KACTUS project, where we do find a separate notion of physical processes or phenomena, it resembles a function-oriented abstraction of our process notion. The reason is that it depends on the task how much detail one needs. In our type of tasks, control- and design-oriented prediction, process detail is required for making adequate modelling decisions. In other tasks, such as electrical network diagnosis and service recovery as in [Bernaras & Laresgoiti, 1996], (dys)function abstractions of underlying processes are sufficient to do the job.
We now turn to aspects of what [van Heijst et al., 1996] calls generic ontologies, representing theories that are supposed to be valid across many fields. Devising a satisfactory top-level categorization of generic concepts (such as thing, object, state, event, etc.) has turned out to be extremely hard [Lenat & Guha, 1990, Sowa, 1995, Skuce, 1993, Benjamin et al., 1996]. On the other hand, the present work has indicated that it is practically feasible and useful to use and reuse generic theories such as mereology, topology and systems theory in domain ontology building. Hobbs [Hobbs, 1995] comes to a similar conclusion (he calls it `core theories') in the context of language understanding. These generic ontologies are abstract theories that define particular kinds of relations (part-of, connected-to, etc.) over abstract entities. A standard top-level concept taxonomy for such entities apparently is not a requirement for the reuse of generic ontologies. What happens is that these abstract entities are projected onto the relevant domain objects. After this, the way is open for further extension and specialization by adding axioms expressing more specific domain knowledge.
Concerning the contents of the generic ontologies that have been reused in PHYSSYS, we note that we have employed rather classical theories of mereology and topology. Alternatives are being discussed also in the ontology literature [Gerstl & Pribbenow, 1995, Guarino, 1995, Eschenbach & Heydrich, 1995]. One of the efforts in ontology research is to combine mereology and topology in one theory that expresses the part-of relation in terms of connectedness [Clarke, 1981]. In PHYSSYS we have followed an approach similar to what is described in [Eschenbach & Heydrich, 1995] where mereology is extended with topological relations. This is because we are not primarily interested in the philosophical question whether mereology and topology can be unified within a single theory. Rather, we want to reflect the engineering practice where components are thought to be decomposed first and connected later on (often as off-the-shelf components) as a step in configuration design. Furthermore, an ontology of mereology without topology imposes less ontological commitments. Our approach is based on incremental specification which yields more structure and is also easier to understand.
In this paper we have barely touched upon the issue of method ontologies: ontologies that are oriented towards problem solving methods, specifying information requirements that must be fulfilled by domain models such that inference methods can be executed [Gennari et al., 1994, Tu et al., 1995, van Heijst et al., 1996]. In the OLMECO library work both task and method were essentially fixed and outside the scope of the research, the task being prediction and the method the standard class of numerical simulation (ODE integration) algorithms. As a result, the PHYSSYS ontology library is currently mainly a collection of domain and generic ontologies, although extensions to method ontologies are very well possible. Especially at the mathematical level it is straightforward how to approach this, since different simulation methods entail different requirements on the form of the inputted mathematical model, whereas (computer) algebraic methods bring along yet other requirements. But also with regard to the process ontology, method-oriented extensions are conceivable, for example in causal reasoning and feedback analysis (whereby as a bonus, graph theory is typically invoked as a generic ontology). The method aspects will mostly entail extensions of the current ontology library, implying that many current definitions and theories are to a certain extent neutral with respect to problem solving methods. This suggests that it is possible to mitigate the so-called interaction problem [Bylander & Chandrasekaran, 1988, van Heijst et al., 1996].
The current state of the OLMECO library thus gives passive, but not active support to the modelling (sub)task. In recent work, on automated model revision for which a running KBS called 007 has been developed [Pos et al., 1996b, Pos et al., 1996a], model revision is actively carried out by the system on the basis of repair plans. Model revision itself is based upon a (considerable) extension of the Propose-and-Revise method [Marcus & McDermott, 1989]. What is interesting in the present context is that some of the repair plans are able to automatically adapt models, such that they conform to the requirements of given simulation methods. Thus, some repair plans function on the basis of knowledge about available method ontologies and about method-oriented revisions of domain models.
Gruber has listed in [Gruber, 1995] a number of design principles for ontologies -- clarity, coherence, extendability, minimal encoding bias, minimal ontological commitment. In general, these turn out to be valid design principles as far as the PHYSSYS ontology is concerned. The principle of `minimum ontological commitment' deserves however some further discussion. In [van Heijst et al., 1996] it is suggested to operationalize this principle as the minimization of the number of theory inclusions in the ontology. Guarino [Guarino & Giaretta, 1995] proposes a formalization of ontological commitment in a modal-logic style. Informally and roughly stated, statements of an ontological theory must be true in every possible world; ontological commitment comprises the set of possible worlds thus allowed by the ontological theory specification. The minimum commitment principle favours the weakest theory (maximum number of models) and tends to emphasize the danger of overcommitment by excluding allowable worlds. In our opinion, there are two practical dangers: excluding acceptable possible worlds, but also including undesired ones. Overcommitment leads to reduction of reuse and sharing, but undercommitment diminishes end-user guidance and support. For example in the component part of our ontology, an issue is whether or not one wants to have typed connections between components. No typing means less commitment, but it also implies that end users are not prevented from making undesirable system models whereby an electrical outlet of one component is directly connected to a mechanical plug of another component. Our experience in the PHYSSYS ontology is that the specification is a balancing act between over- and undercommitment.