An Approach to Integrating Expert Systems with Connectionist Networks

Article type: Research Article

Authors: Gutknecht, Matthias^; | Pfeifer, Rolf

Affiliations: AI Lab, Institut für Informatik der Universität Zürich, Winterthurerstrasse 190, CH-8057 Zürich, Switzerland, e-mail: [email protected], phone: +41-1-257 43 23, fax: +41-257 43 43

Note: [] This paper is an extended version of “Experiments with a Hybrid Architecture: Integrating Expert Systems with Connectionist Networks”, which has been presented at Avignon 90 – the Tenth International Workshop on Expert Systems & their Applications, Avignon, France, 1990. It was awarded by INGENIA at the Avignon'90 Conference Expert Systems and their Applications, with the prize for the best paper for a student researcher.

Abstract: Conventional expert systems as well as connectionist systems suffer from several problems if they are used as isolated paradigms. However, in combination many of the advantages of each can be realized in one system. An example of such a “hybrid” system for the domain of high-school physics problems is presented. It is demonstrated how this system incrementally incorporates experience by compiling rule-based knowledge into a connectionist network. The problems of interfacing a symbolic with a connectionist system and the fundamental issues underlying incremental learning are discussed and illustrated by simulation experiments. It is also shown that a hybrid system not only leads to an increase in performance, but that the approach has a number of essential benefits as a research strategy, both for expert systems and connectionist networks. The results are encouraging and warrant application to a real-life domain.

DOI: 10.3233/AIC-1990-3303

Journal: AI Communications, vol. 3, no. 3, pp. 116-127, 1990