On rule pruning using fuzzy neural networks

Shann and Fu (SF) proposed a fuzzy neural network (FNN) for rule pruning in a fuzzy controller. In this paper we first analyze the FNN of SF and discu ss some of its limitations. SF attempted to eliminate redundant rules inter preting some of the connection weights as certainty factors of rules. In th eir strategy the connection weights are unrestricted in sign and hence thei r interpretation as certainty factors introduces some inconsistencies into the scheme. We propose a modification of this FNN, which eliminates these i nconsistencies. Moreover, we also propose a pruning scheme which, unlike th e scheme of SF, always produces a compatible rule set. Superiority of the m odified FNN is established using the inverted pendulum problem. (C) 1999 El sevier Science B.V. All rights reserved.