A comparison of fuzzy identification methods on benchmark datasets

In this paper, we address the task of discrete-time modeling of nonlinear dynamic systems. We use Takagi-Sugeno fuzzy models built by LOLIMOT and SUHICLUST, as well as ensembles of LOLIMOT fuzzy models to accurately model nonlinear dynamic systems from input-output data. We evaluate these approaches on benchmark datasets for three laboratory processes. The measured data for the case studies are publicly available and are used for development, testing and benchmarking of system identification algorithms for nonlinear dynamic systems. Our experimental results show that SUHICLUST produces smaller models than LOLIMOT for two of the three datasets. In terms of error, ensembles of LOLIMOT models improve the predictive performance over that of a single LOLIMOT or SUHICLUST model.