In this paper we describe an experiment to study the problem solving behavior of a group of knowledge engineers. The subjects are knowledge engineers trained in the Generic Task framework [4]. The study has two aims: (1) to evaluate the degree of consistency among a set of engineers trained in the same high-level framework in order to assure the presence of a consistent methodology within such a group, and (2) to develop methods for studying knowledge engineering activity, which can also be applied to practitioners trained in other paradigms of knowledge engineering. Since such an analysis is exactly the domain of knowledge engineering, we use a knowledge level framework to model the knowledge engineering task. The use of the Design Model of the Generic Task theory [5] as an analysis framework for the knowledge engineers' problem solving process is motivated and its application demonstrated by in-depth analyses of solutions produced by our subjects. The results of our empirical study and its interpretations as well as methodological questions are discussed. It is concluded that the analysis of the knowledge engineers' task with the Generic Task Design Model provides interesting insights, but it also needs to be refined and complemented with more empirical evidence.

We describe the knowledge acquisition process that is used in MedFrame/CADIAG-IV, a medical computer consultation system. Fuzzy medical knowledge is used to model the vagueness and the uncertainty of medical concepts and fuzzy logic reasoning mechanisms provide the basic inference engines. Knowledge acquisition procedures and computer tools have been implemented in order to make the tasks of (a) defining medical concepts, (b) providing appropriate interpretations for patient data, and (c) constructing inferential knowledge easier and more accessible. This paper explains how the knowledge acquisition tasks are supported both by special representations and by a stepwise knowledge acquisition process.

A medical diagnostic and therapeutic consultation system, called MEDFRAME/CADIAG-4, is developed to support diagnostic and therapeutic decision making in various subdomains of internal medicine.' To encompass some of the limitation of its predecessor systems, MEDFRAME/CADIAG-4 has been completely redesigned to (a) account for today's demands for client/server-based, platformindependent systems, (b) use flexible object-oriented knowledge-modeling techniques, and (c) provide interoperability with medical terminology and knowledge servers. The object structure of MEDFRAMEICADIAG-4's knowledge base was analyzed using the Syntropy method, a second-generation method for objectoriented analysis and design.2 This method allows for an implementation-independent, but still semantically meaningful representation (and checking) of the knowledge types and relationships that are needed to model a complex knowledge-based system. However, a common shortcoming of such design methods is the lack of appropriate tools to actually implement these structures in a usable computer system. Thus, we had to translate the final object model manually into an augmented relational database system. Despite the expressive power of object-oriented representation models, a second gap remains when it comes to actually acquiring the specific knowledge instances that will be used for the execution of the resulting system. Although the object model clearly defines what can and should be acquired from either the domain expert (knowledge acquisition) or during the execution of the consultation system (run-time), it does not prescribe reasonable or even useful knowledge acquisition interfaces (neither does the database implementation). Due to the highly complex structure of object-oriented representations, many of our manually constructed user-interface prototypes were error-prone or incomplete at best. To overcome this shortcoming, we used some of the tools that are provided in the PRYllEGt-l system.3 One important and well-developed part is specifically designed to assist in the automatic generation of knowledge-acquisition tools from so-called domain ontologies. Because of their similar background (both approaches use object-oriented class hierarchies and are able to express associations and relations between objects/classes), a manual translation from some parts of the MEDFRAME/CADIAG-4 object-oriented Syntropy model into PRCYIlG&-Hl's domain ontology representation was straightforward. However, we had to omit some parts of our knowledge model (such as rules and fuzzy membership functions) because the current implementation of PROTGItlt-l can not produce adequate interfaces for them.

We propose a stepwise knowledge acquisition approach based on fuzzy set theory to support the development and refinement of medical knowledge bases. The definition of fuzzy relationships between medical entities allows to represent knowledge at different levels of precision. The definition of relationships is supported by the use of linguistic variables and a semi-automatic knowledge acquisition program.

Medical data, which are results of examinations performed on a patient, require an interpretation to allow an assessment of the patient's condition. As a result of this interpretative procedure, medical data are converted into symbolic descriptions of the patient. It is shown that context-specific fuzzy membership functions and fuzzy membership functions with two function parameters are helpful to acquire a symbolic description of most clinically relevant medical data.

The main goal of this paper is to explore the possibilities of exploiting psychological methods for the purpose of knowledge engineering. Hypotheses are presented why both the pure ''psychological'' and the pure ''engineering'' positions are not viable for building expert systems. A ''middle-out'' strategy is proposed that preserves the best of both worlds while minimizing the problems of each. This ''middle-out'' strategy consists of the application of so-called ''task-level frameworks''. However, these frameworks do not sufficiently support one of the most crucial tasks in the knowledge engineering process, namely the mapping of the actual expert behavior onto conceptual models. In this paper, a new method which makes this process easier and more reliable is described and a standardized several-step procedure for mapping expertise-in-action protocols onto a task-level framework is illustrated with a case study. It is concluded (a) that protocol analysis is a good starting point for developing tools to support the knowledge engineering process-if appropriate methods are available, and (b) that methods are only appropriate if they are ecological on the one hand and pragmatic on the other.