Oxford University, Computing Laboratory

At the beginning of the last decade has become obvious that there are at least four distinct intelligent methodologies in the area of Artificial Intelligence: traditional knowledge based systems, neural networks, fuzzy systems and genetic algorithms. Other intelligent paradigms such as rough sets and multi-agent systems have also emerged in the field.

Neural networks have been criticized for their "black box" nature given by their lack of transparency in human understandable terms. On the other hand the main disadvantage of traditional symbolic knowledge based systems and fuzzy systems is their inability in learning and updating the knowledge base. Hybrid Intelligent Systems combine two or more intelligent approaches in order to overcome the difficulties of each approach and to solve in a better way the difficult and complex real-world problems.

The integration of connectionist modules with implicit knowledge representation, based on neural networks, with symbolic modules with explicit representation of knowledge requires methods for rule extraction from neural networks, insertion and refinement. Rule refinement is the process of extracting of a refined set of rules from a neural network that was initialised using crude domain knowledge in the form of an initial set of rules. The term rule generation is used to designate both the rule extraction and rule refinement/insertion techniques.

The talk will present few methods for rule extraction from neural networks, for generating both crisp rules as well as fuzzy rules using neural network learning algorithms. The talk will focus on how to extract rules from black-box neural networks with no structure that facilitates the rule extraction process, which is the case with the well-known neuro-fuzzy networks from the literature. In the proposed approach genetic algorithms are used for optimizing the extracted set of rules.

Additionally the talk will discuss how to combine in a hybrid architecture various modules based on different representation of knowledge.

DATE: May 31, 2002, Friday @ 15:45

PLACE: EA-409