A neuro-fuzzy architecture incorporating complex fuzzy logic.

摘要

Complex fuzzy sets have recently been a topic of interest in the fuzzy systems community. However, to date, no practical application of this concept has yet been proposed. The goal of this thesis is to create a time series forecasting system, which will be the first practical application of the complex fuzzy sets. We have constructed a neuro-fuzzy architecture, named ANCFIS, inducing complex fuzzy rules from time-series. The challenge of this architecture is how to update the parameters of complex fuzzy sets. We have developed a novel derivative-free optimization technique to overcome this problem: the Variable Neighborhood Chaotic Simulated Annealing (VNCSA) algorithm, and compare VNCSA against two existing alternatives: a chaotic simulated annealing technique, and Ant Colony Optimization algorithm. Our comparisons are carried out over one synthetic dataset and five real-world datasets. We found that the VNCSA algorithm leads to the best tracking error in the ANCFIS architecture, for all six datasets.