TIME SERIES PREDICTION USING EXTREME LEARNING BASED MODELS

Abstract

Implementation of soft-computing techniques in many areas has been largely studied recently. The self-adaptiveness and data-driven characteristics of these methods makes them suitable to solve complex problems where direct solutions are not easily obtained using available analytical methods. To train these models, various iterative methods have been implemented. The reported problems of these iterative methods such as overfitting, time-consuming, associated with the abundance of different versions motivate the implementation of much simpler and efficient algorithms. In this work, we study the implementation of random approaches for prediction and regression problems. These models are much simpler and provide reasonable results. The performance of the models is proposed to be improved and stabilized with the application of expansion through orthogonal functions. To evaluate the models, benchmark examples in nonlinear systems identification, classification and regression problems are also given. Two random models are studied, the single layer feedforward network model (SFLN) and the neuro-fuzzy model ELANFIS. It’s found that with the implementation of expansion through orthogonal functions, more specifically, Chebyshev polynomials, the performance of the models can be improved in some cases. The proposed models are called Functional Link Extreme ANFIS (FL-ELANFIS) and Improved Random vector functional Link (I-RVFL) for the neuro-fuzzy and SFLN based models respectively.