A Multiple Model Approach to Time-Series Prediction Using an Online Sequential Learning Algorithm

• Published in: IEEE transactions on systems, man, and cybernetics. Systems Vol. 49; no. 5; pp. 976 - 990
• Main Authors: George, Koshy, Mutalik, Prabhanjan
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.05.2019; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Adaptation models; Algorithms; Approximation algorithms; Artificial neural networks; Brain modeling; Computational modeling; Distance learning; Machine learning; Mathematical model; Mathematical models; Neural networks; Performance prediction; Prediction algorithms; Prediction models; Predictive models; recurrent neural networks; supervised learning; Switching theory; time-series analysis
• ISSN: 2168-2216; 2168-2232; 2168-2232
• DOI: 10.1109/TSMC.2017.2712184

Time-series prediction is important in diverse fields. Traditionally, methods for time-series prediction were based on fixed linear models because of mathematical tractability. Researchers turned their attention to artificial neural networks due to their better approximation capability. In this paper, we use feedforward neural networks with a single hidden layer, and present a rather simple online sequential learning algorithm (OSLA) together with its proof. The convergence properties of this algorithm are those of the well-known recursive least squares algorithm. We demonstrate that the prediction performance is better than other OSLAs, and show that it is statistically different from them. In addition, we also present the multiple models, switching, and tuning methodology that enhances the prediction performance of the learning algorithm.