Using biologically inspired computing to effectively improve prediction models

Abstract

The complexity of certain problems causes that classical methods for finding exact solutions have some limitations. In this paper we propose an incremental heterogeneous ensemble model for time series prediction where biologically inspired algorithms offer a suitable alternative. Ensemble learning techniques are advantageously used for improving performance of various prediction methods. The quality of this kind of machine learning approaches depends on proper combination of used methods. The influence of each of the used method can change on the fly and is determined by proper choice of its weights. Finding optimal weights in prediction methods represents typical optimization problem with objective function reflecting error minimization, where biologically inspired algorithms can be used. In the proposed paper, we study several biologically inspired algorithms in the process of weights optimization. We investigate and compare ensembles using base models and ensembles op- timized by biologically inspired algorithms. We demonstrate that the ensemble learning prediction models optimized by biolog- ically inspired algorithms outperformed the base prediction methods. We present performance and accuracy results of proposed ensemble models that were evaluated on power load datasets with concept drifts.