For several years, researchers have argued that there are significant advantages when intelligent systems are constructed so that the procedural knowledge used to solve problems is made explicit. In the 19805, a number of researchers experimented with single, abstract problem—solving methods (such as "heuristic classification" or "propose and revise") that not only provided a unifying architecture for a class of knowledge-based systems at runtime, but that also defined precise "knowledge'roles" that could facilitate the process of knowledge acquisition: Building a new knowledge-based system became a matter of providing domain knowledge to satisfy each of the "knowledge roles" associated with the method. The explicitness of these knowledge roles made it possible to build both automated knowledge-acquisition tools that could create systems that incorporated these unitary problem—solving methods, and knowledge-acquisition-tool generators that could create interactive knowledge—acquisition tools that were specialized for different classes of application areas.

Although monolithic problem-solving methods such as "propose and revise" can be shared among developers and reused to generate new applications, they have significant limitations. They often do not fit the problem requirements of the application tasks for which they are employed, and lack flexibility at runtime in adapting their control strategies to unexpected situations. In recent years, investigators have begun to experiment with sharable and reusable problem-solving methods that may be composed from smaller grained building blocks, and that may provide more versatility in system development and operation.
Furthermore, several research groups have begun to investigate how domain models also might be made reusable and sharable. These investigations have resulted in various proposals for formal ontologies that make explicit the conceptualizations on which domain models are based, and for providing the means to define different views on such domain models.

This Dagstuhl seminar brought together investigators and practitioners from Europe, the United States, and Japan, all of whom were involved in development of intelligent systems from reusable components. Intensive discussions made it evident that, despite continued progress by the research community, terms used to describe problem-solving methods and the domain knowledge on which those methods operate are not always used in a consistent manner, thus sometimes making it difficult to compare approaches. An important benefit of the seminar was to understand alternative frameworks in more detail.

Although the use of formal languages to represent problem-solving behavior in a declarative fashion offers the best opportunity to clarify semantics, seminar participants were keenly aware of the difficulty of formalizing knowledge that is inherently procedural in nature. Much discussion focussed on the limitations both of formal descriptions of problem—solving methods, and of situations in which problem-solving components have only operational semantics.

Additional discussion concentrated on the relationships between problem-solving methods and domain knowledge. Seminar participants debated approaches to representing as explicit ontologies both domain knowledge and the data on which problem-solving methods operate.

Seminar participants demonstrated the importance ofdevelopingmodularized, component-based architectures for intelligent systems. There was a belief that researchers were developing a more shared perspective, but that differences in approach clearly remain. However, now that the development and definition of reusable problem-solving methods has become a maturing field, there is an urgent need for empirical results to validate many of the assumptions that investigators have made. The difficulty of carrying out many of the necessary empirical studies was well appreciated by seminar participants.

Creative Commons BY 3.0 DE