Author:
Vitaliy Kolodyazhniy
Affiliation:
University of Salzburg, Austria
Keyword(s):
Multilayer perceptron, B-spline, Nonlinear synapse, Fuzzy rule, Deterministic initialization, Singular value decomposition.
Related
Ontology
Subjects/Areas/Topics:
Approximate Reasoning and Fuzzy Inference
;
Artificial Intelligence
;
Computational Intelligence
;
Fuzzy Systems
;
Neuro-Fuzzy Systems
;
Pattern Recognition: Fuzzy Clustering and Classifiers
;
Soft Computing
Abstract:
A multilayer spline-based fuzzy neural network (MS-FNN) is proposed. It is based on the concept of multilayer perceptron (MLP) with B-spline receptive field functions (Spline Net). In this paper, B-splines are considered in the framework of fuzzy set theory as membership functions such that the entire network can be represented in form of fuzzy rules. MS-FNN does not rely on tensor-product construction of basis functions. Instead, it is constructed as a multilayered superposition of univariate synaptic functions and thus avoids the curse of dimensionality similarly to MLP, yet with improved local properties. Additionally, a fully deterministic initialization procedure based on principal component analysis is proposed for MS-FNN, in contrast to the usual random initialization of multilayer networks. Excellent performance of MS-FNN with one and two hidden layers, different activation functions, and B-splines of different orders is demonstrated for time series prediction and classificat
ion problems.
(More)