Next: Product and Production Up: Knowledge-Base Evolution for Product Previous: Knowledge-Base Evolution for Product

Introduction

It is a long held belief, that micro-worlds, such as the blocks world, sorting tasks or chess end games are the drosophila of Artificial Intelligence and Machine Learning research, where the fundamental successes are to be achieved and demonstrated. A quote by Amarel 1000 ##1##2[##1, ##2 ##1##2##2[p.258]Amarel83 highlights this view. "These toy problems provide an excellent paradigmatic task environment in which essential aspects of the representation problem can be studied...They are serving as drosophila of research in the general area of problem representations, and in the study of acquisition of problem solving skills".

Although there cannot be any doubt that many successes of Machine Learning have been achieved in these micro-worlds, the utilization of these achievements in complex real world domains (e.g. the industrial applications of Machine Learning) is much more difficult than had been originally anticipated. Buchanan 1000 ##1##2[##1, ##2 ##1##2##2[p.5]Buchanan89 for example reports, that except for simple classification systems, knowledge-based systems do not yet employ a learning component to construct parts of the knowledge bases from libraries of previously solved cases.

It has been pointed out only recently, that real world domains have quite different characteristics than the micro-words where new machine learning techniques are routinely demonstrated. Complexity, continuous innovations and documentation as well as incomplete and conflicting knowledge are the most eminent characteristics (cf. [\protect\citeauthoryearSchmalhofer et al.in press]). Because of the dynamic character of real world domains, the application of knowledge-based systems requires that the changes in the field can at least be traced (preferably predicted and discovered) by appropriately selected machine learning techniques. Such updating and revision processes are termed knowledge base evolution. Comparable to the human genome project which also requires additional resources, above and beyond the discovery of the genetic mechanisms with the drosophila, the ILP community must therefore also pay more attention to applications in complex real world domains.

In order to develop knowledge-base evolution techniques with respect to complex real world domains, we first analyzed the requirements of product and production planning with new materials by using the specific example of the manufacturing of bucket seats in the car industry. The results are summarized in Section 2 of this paper. Section 3 then describes a respective knowledge-base that is currently being developed by an iterative application of the CLASSIC methodology to knowledge engineering [\protect\citeauthoryearBrachman et al.1990]. Section 4 will then show how the knowledge evolution can be understood as theory revision [\protect\citeauthoryearRichards and Mooney1991], where the knowledge-base evolution system and the user cooperate in a way, similar to an apprenticeship learning system [\protect\citeauthoryearTecuci and Kodratoff1990].

Theory Revision has recently been proposed as a general framework, where Explanation-Based Learning (EBL) and Inductive Logic Programming (ILP) can be integrated [\protect\citeauthoryearMooney and Zelle1994]. For mastering the knowledge evolution requirements of the specific application, we can thus draw upon the basic research results from both EBL as well as ILP. Furthermore, exploration and verification processes will be distinguished. A continuous (interactive) improvement of a knowledge base during its entire life-time starting with the first formalizations (knowledge base seed) and still continuing along its practical use can thus be achieved (cf. [\protect\citeauthoryearMeyer1994]).

Expert knowledge from the application domain is used for constraining the exploration processes, so that an efficient implementation can be obtained. Expert knowledge will be employed to determine the representation bias (also known as `restricted hypothesis space bias') and search bias (also known as `preference bias') of induction [\protect\citeauthoryearRendell1986]. More specifically, domain knowledge is used to specify the representational bias and metaknowledge to determine the search bias. The paper will be concluded with a general discussion of the role of knowledge-base evolution for the quality of practical knowledge bases.



Next: Product and Production Up: Knowledge-Base Evolution for Product Previous: Knowledge-Base Evolution for Product


Harold Boley & Stefani Possner (possner@dfki.uni-kl.de)