Identification of time series models using sparse Takagi–Sugeno fuzzy systems with reduced structure

Abstract

AbstractSimplifying fuzzy models, including those for predicting time series, is an important issue in terms of their interpretation and implementation. This simplification can involve both the number of inference rules (i.e., structure) and the number of parameters. This paper proposes novel hybrid methods for time series prediction that utilize Takagi–Sugeno fuzzy systems with reduced structure. The fuzzy sets are obtained using a global optimization algorithm (particle swarm optimization, simulated annealing, genetic algorithm, or pattern search). The polynomials are determined by elastic net regression, which is a sparse regression. The simplification is based on reducing the number of polynomial parameters in the then-part by using sparse regression and removing unnecessary rules by using labels. A new quality criterion is proposed to express a compromise between the model accuracy and its simplification. The experimental results show that the proposed methods can improve a fuzzy model while simplifying its structure.