Multiobjective Evolutionary Neural Networks for Time Series Forecasting

To improve the overall generalization ability for the set of solutions attained by the multiobjective evolutionary optimizer, a simple algorithm to filter possible outliers, which tend to deteriorate the overall performance.

Abstract

This paper will investigate the application of multiobjective evolu-tionary neural networks in time series forecasting. The proposed algorithmic model considers training and validation accuracy as the objectives to be optimized simultaneously, so as to balance the accuracy and generalization of the evolved neural networks. To improve the overall generalization ability for the set of solutions attained by the multiobjective evolutionary optimizer, a simple algorithm to filter possible outliers, which tend to deteriorate the overall performance, is proposed also. Performance comparison with other existing evolutionary neural networks in several time series problems demonstrates the practicality and viability of the proposed time series forecasting model.