Friedrich Gebhardt

Reasoning with complex cases

Confronted with a problem, humans often remember previous problem solving episodes or solutions and try to modify them for their current situation (see figure 2.1). Case-based reasoning is an attempt to support this approach with a computer. A quick tour through case-based reasoning in section 2.1 gives a first impression of the main concepts and some central problems encountered in this field. All the notions and procedures encountered there will be reconsidered in more detail in sections 2.2 to 2.8. The chapter closes with a summary characterizing complex cases.

Survey on structure-based case retrieval

The main components of case-based reasoning are case retrieval and case reuse. While case retrieval mostly uses attribute comparisons, many other possibilities exist. The case similarity concepts described in the literature that are based on more elaborate structural properties are classified here into five groups: restricted geometric relationships; graphs; semantic nets; model-based similarities; hierarchically structured similarities. Some general topics conclude this survey on structure-based case retrieval methods and systems.

Identifying clusters in spatial area data

A task peculiar to spatial data is to find spatial clusters. Three alternative tests will be given here for the case that a region is divided into a number of districts and each district is assigned a numerical value, i. e. the value for a variable of interest. A cluster of districts with high (or low) values cannot just be based on the adjacency of districts since then a random assignment would already produce too many and too large clusters. Thus the cluster definition is based on district triplets, essentially three districts meeting at a common corner. Two tests are designed for a division of the area into marked and unmarked districts. The third, more elaborate test uses the given values directly.

Cluster tests for geographical areas with binary data

Abstract Consider an area subdivided into non-overlapping districts, e.g. a state divided into counties, and assume that some districts are marked for having some distinguishing property. Then the question arises whether the marked districts are distributed randomly or exhibit some spatial clustering. This question is pursued here to some extent theoretically, in particular using regular tesselations of the plane (hexagons in a honeycomb), and to some extent by simulations using these regular constellations as well as real situations, in particular 171 counties in six Bundeslander of Germany. Connectivity regions (maximal regions of marked districts) turn out to be a bad choice in general. Instead, clusters are formed from triplets of marked districts. Approximate statistical tests are developed. They are simple enough to be used for data mining where potentially a large number of tests has to be performed.

Virtual Building Site

FABEL prototype 3.0 is a research system to support engineers and architects in building design. We introduce and motivate the metaphor of a virtual building site, which has guided the development of this system. In the virtual building site a common data model of the artefact is designed. The user interface is based on a graphic editor which handles design objects of all kinds and at several levels of abstraction and even organizes the application of the tools. There are tools to supply the designer with useful cases, to create pieces of design by adaptation of similar cases or by knowledge-based refinement, and there are tools to assess parts of the design. We describe some of the tools to illustrate the multi-method approach of FABEL.

Reuse of a Single Case: Adaptation

This chapter proposes a framework for describing systems that reuse a single source case. It applies it to the systems selected and extracts a first set of guidelines. The major issue in the design of such systems is the approach to adaptation, because it influences many other decisions. Adaptation remains an issue with case-based reasoning systems that reuse multiple source cases. But there, strategic questions of how to handle more than one source case move into foreground. Therefore, this chapter focuses on case adaptation and the next one on strategies.

Source Cases — The Third Factor

The conclusions from the micro level of single case adaptation in section 13.3 and from the macro level of multiple case reuse in section 14.3 indicate that both aspects are related to a third factor in the design of a case reuse system: the size and content of the source cases and the relations between them. This chapter is devoted to this still neglected yet central issue. The systems need not be revisited, because source cases have been discussed together with the strategies in section 14.2. A last group of guidelines will be given to relate the choice of source cases to the approach to adaptation and to the strategy.

Model-Based Similarities

The case retrieval systems in the preceding chapters employ more or less domain-independent methods. Some other systems utilize deep knowledge of the special application; they are based on some kind of domain model. Retrieval tends to be more complicated than just comparing two cases; the notion of similarity becomes blurred.

Similarities Between Attributes

Retrieval systems have been used for decades and brought to some perfection for document retrieval. No wonder they are being employed for cases, too, even probably in the majority of applications.

Designing Integrated CBR Systems

Most existing CBR systems focus on one particular issue, e.g. a retrieval or a reuse method. Therefore no unusual questions of system architecture arise. If however many methods are to be combined, this will lead to problems. FABEL PROTOTYPE is the only known CBR system combining a bunch of quite diverse methods. Therefore, this part of the book on integrated systems is predominantly devoted to FABEL PROTOTYPE. In the present chapter, we deal with questions of knowledge acquisition for complex CBR systems and with general problems of CBR architectures and we illustrate the FABEL PROTOTYPE solution in a non-technical overview.