Abstract: This part shows, in short, what can be understood from a ANN trained as, e.g., in the previous parts. Understanding what is written in a black-box ANN is also called Knowledge Eliciting.
This part is structured into 2 sections.
The first section deals with the theoretical aspects of rule extraction (knowledge eliciting) from ANNs. Starting with the definition of rule extraction, this section then explains the inverse problem: rule implementation, or the transformation of a rule base into a ANN. With this platform, the rule extraction mechanism is easier understood. Next, a few popular rule extraction mechanisms are presented. These algorithms are then discussed, classified, and finally rule quality criteria emerge.
The second section presents the specific application implemented and simulated on a computer. This problem is the second part of the main application of the thesis: the extracted knowledge has to be used in a classroom for teaching purposes. This being the problem setting, the quality criteria have to change. Starting from a typical knowledge based ANN, a new tool is built, a sunsymoblic knowledge extraction tool for time-series prediction. Implementation details are given, then some evaluations are made and limitations are discussed. Experts are called to judge upon the system, some results are displayed, conclusions are drawn.