Authors:
Wanqing Zhao
1
;
Kang Li
1
;
George W. Irwin
1
and
Qun Niu
2
Affiliations:
1
Queen's University Belfast, United Kingdom
;
2
Shanghai University, China
Keyword(s):
Fuzzy neural systems, Interpretable model, Differential evolution, Weighted fast recursive algorithm, ANFIS.
Related
Ontology
Subjects/Areas/Topics:
Artificial Intelligence
;
Computational Intelligence
;
Fuzzy Systems
;
Fuzzy Systems Design, Modeling and Control
;
Learning and Adaptive Fuzzy Systems
;
Neuro-Fuzzy Systems
;
Soft Computing
;
Soft Computing and Intelligent Agents
;
System Identification and Fault Detection
Abstract:
Many learning methods have been proposed for Takagi-Sugeno-Kang fuzzy neural modelling. However, despite achieving good global performance, the local models obtained often exhibit eccentric behaviour which is hard to interpret. The problem here is to find a set of input space partitions and, hence, to identify the corresponding local models which can be easily understood in terms of system behaviour. A new hybrid approach for the construction of a locally optimized, functional-link-based fuzzy neural model is proposed in this paper. Unlike the usual linear polynomial models used for the rule consequent, the functional link artificial neural network (FLANN) is employed here to achieve a nonlinear mapping from the original model input space. Our hybrid learning method employs a modified differential evolution method to give the best fuzzy partitions along with the weighted fast recursive algorithm for the identification of each local FLANN. Results from a motorcycle crash dataset are i
ncluded to illustrate the interpretability of the resultant model structure and the efficiency of the new learning technique.
(More)