Fuzzy-Neural Network For Decision Control Systems

Abstract

In this work, a proposed Fuzzy-Neural Network for Decision Control System (FNNDCS) is presented. It was built through four distinctive phases: knowledge representation phase, merged rules formatting phase, mapping phase, and training phase. The first phase begins by extracting the knowledge from experts in the form of initial Expert Rules Set (ERS). If extracted knowledge is linguistic, then the solution will proceed directly through the suggested phases. Else, if it is numerical, then one of proposed three-fuzzification realization (3-FRs) is used to fuzzity those numerical data into ERS format. FNNDCS design depends on construction a new set of truth linguistic terms that represented mathematically by fuzzy numbers. Three classes of fuzzy numbers are presented: Non-Linear Degradation Fuzzy Numbers (NLDFNs), Linear Degradation Fuzzy Numbers (LDFNs), and Constant Degradation Fuzzy Number (CDFN). In the second phase, ERS is reformulated into a new set of fuzzy rules, which is called "Domain Merged Rules" (DMRs) based on the previously mentioned fuzzy numbers and by using two algorithms: Grouping algorithm and Entries-Merge algorithm.

In the third phase, Structure Construction Rules (SCR) algorithm is used for mapping DMRs into initial connectionist architecture, which is the proposed Fuzzy Neural Network (FNN-1 ). In the last phase, Structure Connectivity Learning Rules (SCLR) algorithm is used for training the network. It is possible to prune the resultant network by using Rules-Eliminating algorithm.

Five case studies for different fields of application include control, classification, prediction, decision-making, and curve fitting are selected from literature, and FNNDCS has proved to efficiently address and solve each one of them. Moreover, the results of are compared favorably against the published results in terms of structure, time of learning, performance, and solution efficiency.