Parameter adaptation for GP forecasting applications

Abstract

Genetic Programming (GP) has been applied to time series forecasting often with favorable results. However, for forecasting tasks several open issues concerning parameter settings exist. Many real-world forecasting tasks are dynamic in nature and, thus, static parameter settings may lead to inferior performance. This paper presents the results of recent studies investigating non-static parameter settings that are controlled by feedback from the GP search process. Specifically, non-static settings for population size and training data size are explored. © Springer-Verlag Berlin Heidelberg 2007.