In this paper, we discuss how to use FDHECNN's (fuzzy degraded hyperellipsoidal composite neural networks) to extract fuzzy rules for function approximation. The FDHECNN can perform function approximation in the same manner as networks based on Gaussion potential functions, by linear combination of local functions. Furthermore, the output functions of the hidden nodes in the FDHECNN's offer more flexibility than Gaussion potential functions do. A special scheme is developed to find a set of good initial weights in order to speed up the convergence problem. Results of simulations of a system identification demonstrates that the feasibility and robustness of the proposed fuzzy neural networks.
Relation:
1994 International Computer Symposium Conference Proceeding Volume 2 of 2,頁1246-1250