Neuro-fuzzy(-genetic) system for synthesizing rule-based knowledge from data

Abstract

Systems that synthesize “knowledge” from data have been under intensive investigation over the last several years. These systems are inherently associated with databases and provide tools for “making sense of data” or, more specifically, for revealing valid, useful and understandable patterns in data [36]. The patterns, which represent the knowledge ”encoded” in data, can be described in different ways depending on the theoretical tools applied to the considered problem. One of the most commonly used structures for knowledge representation are IF-THEN rules. Their main advantages are high readability and modularity. Knowledge represented in this way is also highly modifiable (the rules can be easily added to and deleted from the system). Among the main theoretical tools for the generation of IF-THEN rules from data are rough-set based approaches [218–220, 256, 257], decision tree methods enabling rule extraction, e.g., [36, 233], rule induction systems [37], and different neuro-fuzzy techniques, e.g., [18, 84–113, 116, 117, 119, 121, 126, 143, 183, 208, 242]. The first three theoretical tools are mainly used to generate rules for classification tasks (see also Chapter 8), whereas neuro-fuzzy methods can be applied to both classification and continuous-function approximation problems. The latter include modelling systems with continuous outputs as well as designing controllers for continuous systems.