A Novel Fuzzy BP Learning Algorithm for Four-layer Regular Fuzzy Neural Networks 1
Puyin Liu, Wenqiang Yang · 2005
Abstract–The general fuzzy numbers are approximately represented as polygonal fuzzy numbers, which can be determined by finite nested closed interval. Based on interval arith-metic the input–output (I/O) relationship of a four-layer feedforward regular fuzzy neural network (FNN) is analyzed systematically. By introducing semi-jump function ‘Lor ’ a group of partial derivative formulas are established for the error function of the four-layer regu-lar FNN, in which the maximum operator ‘∨’ and minimum operator ‘∧ ’ are included. A BP learning algorithm for fuzzy weights of the reg-ular FNN is developed. To speed the conver-gence of the algorithm the learning constant is updated in each iteration step. Our exper-imental results show that the novel fuzzy BP algorithm can train a regular FNN efficiently to realize a family of fuzzy inference rules ap-proximately, and to finish a uncomplete fuzzy inference rule table to demonstrate the FNN trained by our learning scheme having strong generalization capability. Keywards–Polygonal line operator; Fuzzy arithmetic; Regular fuzzy neural network;