DESIGN AND CONTROL STRATEGIES FOR CONNECTIONIST EXPERT SYSTEMS

Abstract

Connectionist expert systems research has made remarkable progress in knowledge based system field this decade. This thesis is a result of research on the design and the control strategies for connectionist expert system. It consists of theory, methodology and practical application. It is mainly based on the Three-valued Logic and the Neural-logic Networks theory, a branch of neural networks studied in the National University of Singapore, which incorporates both pattern processing capability of multi-layer perceptron and logical reasoning functions of Boolean logic inference network in one constructive learnable network framework. Its distinctive feature of having order pair valued weights makes it capable and powerful in handling both perceptual knowledge and rational knowledge. In the light of nowadays expert system technology, an extensive comparison between general conventional expert systems and connectionist expert systems are made to give an overview of their features and limitations which lead to the discussion of design methods and control strategies used in different part of the connectionist expert system. For knowledge representation, the author introduces a new classification of knowledge types, which makes clearer in analyzing knowledge types and selecting knowledge base model from connectionist or conventional side. Furthermore, the Factor Space approach, which provides a formal measure to analyze and represent knowledge, forms a new way to bridge expert knowledge embodied in the past cases into training data for neural networks training. In the kernel part of this thesis, twenty strategies concerning many aspects of connectionist expert systems are summarized from rational analyses and practical experience. These strategies involve knowledge analyses and representation, connectionist model selection, knowledge base design and control, determination of quantity and quality of training set, learning technique and inference mechanism. Many control strategies are recommended in company with a concrete approach that makes the analysis from general to practical and vice versa. Some advanced control methods are suggested for handling fuzziness and metalevel control, such as Fuzzy Perceptron, Similarity Relations and Metanet Approach. In practical implementation, most of the strategies have been utilized in the development of an avionic unit fault diagnostic consulting expert system. This system, called Inertial Navigation System Interactive Diagnostic Expert (INSIDE), absorbed the quintessence both from connectionist and conventional expert systems. Not only can it learn the expert knowledge from past successful cases but also provides detail troubleshooting flowcharts for test and calibration. It has been practically used for assisting maintenance teams in the Avionic Department of Singapore Airlines. This project has been documented in /SS Technical Report. Its theory and techniques have been published in technical papers*, which were accepted by and presented at a number of conferences, including JJCNN'90** and JJCNN'91. Two developers including the author were awarded the Silver Award in 1990 Tan Kah Kee Young Inventors Competition. Another research project is trying to integrate the fuzzy perceptron with fuzzy rule base system into a heterogeneous fuzzy expert system for decision making on Stock Selection. This research project is still in progress.

*[Pan, Lui, Wang, Tch 19911, [Tan, Pan, Lui, Tch 1990a,b], [fch, Lui, Tan, Pan 1990!, [TR 1990J. **IJCNN: International Joint Conference on Neural Networks co-organized by International Neural Networks Society (INNS) and IEEE.