Representation, Information Flow, and Model Integration

Christopher Menzel
Richard J. Mayer
Leslie K. Sanders
Knowledge Based Systems, Inc.
2726 Longmire
College Station, TX  77845

In this paper, we begin by exploring the general nature of modeling 
and information model integration.  We take such integration to 
involve at least two aspects:  (i) efficient consistency 
maintenance across models, and (ii) accurate and efficient 
propagation of the information carried by one or more models to 
other models that manage related information, albeit perhaps of a 
very different kind.  Call these two aspects of model integration 
the problem of (\it intermodel consistency} and the problem of {\it 
information propagation), respectively.  Models in general are 
effective research and development tools because they enable 
researchers and analysts to distill those features of the object or 
domain being modeled that are essential to their purposes.  They 
can then conduct their studies unencumbered by the extraneous 
"noise" of a real world setting.  A good modeling {\it method} 
codifies general procedures that embody good practice in the 
development and use of models of a certain kind, and provides tools 
that facilitate their construction.  An {\it information} modeling 
method, in particular, supplies procedures and tools for modeling a 
specific kind of information present in a given domain--employee 
information, product design information, manufacturing information, 
etc.--by drawing out certain types of entities and patterns of 
organization in that domain to the exclusion of others.  Call the 
general pattern of information targeted by a method its {\it 
information type}.  To express information of that type, a modeling 
method supplies a representational medium--usually some kind of 
graphical language--tailored specifically to that type.  That is, 
the method establishes semantic connections between the basic 
elements of the medium and the basic elements of its information 
type.  These basic connections in turn form the basis of more 
elaborate semantic connections between complex representations and 
correspondingly complex pieces of information.  Models, in 
particular, are such complex representations, and hence can be said 
to {\it express}, or {\it display}, information of the appropriate
type in virtue of the semantic connections established by the 
method.

These general observations suggest in very broad (and rather vague)
terms the steps that will likely have to be involved in any
effective approach to model integration:

1.  Clarify the general structure of the information type 
targeted by each modeling method used in an enterprise, e.g., 
isolate the basic kinds of elements involved in that type of 
information and characterize how those elements combine to generate 
information;

2.  Draw explicit structural connections between the various 
information types targeted within the enterprise, e.g., identify or 
otherwise relate basic kinds of elements in different information 
types;

3.  Isolate domain specific connections between and constraints 
on the various things referenced by actual models in the 
enterprise, e.g., identify information bearing links between things 
of one kind and those of another;

4.  Guided by those structural and domain specific connections, 
develop methods for intermodel consistency checking and information 
propagation, and in particular methods for determining how the 
information carried by a model generated by one method is to be 
transferred to and--as far as possible--expressed in a model 
generated by (in general) another.

In our paper, we use this general integration architecture to
explore the three dimensions of the model integration problem space
suggested in the Call for Participation.  We argue that, in the
context of that problem space, this architecture reveals in a
general way the need for a "Global Representational Medium" for
expressing information of any given type, along the lines pursued by
the Knowledge Interchange Format (KIF) of the DARPA sponsored
Knowledge Sharing Effort, the PDES Semantic Unification Metamodel
(SUMM), the ANSI Information Resources Dictionary System (IRDS),
and KBSI's Air Force sponsored Neutral Information Representation
Scheme (NIRS).

Consideration of the necessary expressive power for such a medium
leads us to a discussion of some of the intrinsic challenges of
integration.  A particularly important observation in this regard is
that integration is not in general mere translation.  Models are
very much like utterances in natural language in that, often, much
of the information they carry is implicit; much of the information
carried by a model is not overtly expressed as the interpretation of
some explicit representation in the model according some set of
semantical rules, but is instead determined by the model's
context--notably, the eccentricities of the particular domain it is
modeling, and its place among other, related system models.  On this
way of conceiving things, translation is just one form of the more
general phenomenon of the flow of information across models. In
addition to developing robust translation mechanisms based solely on
the formal syntax and semantics of various methods, then, a strong
challenge to effective integration is the development of
representational tools and methods for isolating and expressing the
often more subtle, context-dependent links that determine
information flow.

We close with a somewhat more formal rendering of integration and
related notions based upon recent work in situation theory.