Fuzzy production rule refinement using multilayer perceptrons

Abstract

One of the problems encountered in using fuzzy production rules (FPRs) to represent fuzzy knowledge is the difficulty of assigning suitable threshold values of propositions and certainty factors to FPRs. It is a time consuming and an iterative task which involves painstaking consultation with domain experts to set suitable threshold values and certainty factors of FPRs. Furthermore, these value assignments may be conducted on a trial and error basis and sometimes on an ad hoc basis. In this paper, we propose an initial step to solve this problem with the help of a multilayer perceptron neural network. The initial step is by mapping fuzzy Petri nets, which is a representation of FPRs, into an enhanced multilayer perceptron. The backpropagation (BP) learning algorithm is modified to adopt this change in network structure. Traditional BP and two different sigma-pi functions are used and compared to find the appropriate function for this enhanced network. Furthermore, dummy nodes are also introduced to accommodate such changes. Test data is used to train the multilayer perceptron in order to help adjusting the threshold values and certainty factors of the FPRs. >