Investigating the Use of Geometric Semantic Operators in Vectorial Genetic Programming

Abstract

Vectorial Genetic Programming (VE_GP) is a new GP approach for panel data forecasting. Besides permitting the use of vectors as terminal symbols to represent time series and including aggregation functions to extract time series features, it introduces the possibility of evolving the window of aggregation. The local aggregation of data allows the identification of meaningful patterns overcoming the drawback of considering always the previous history of a series of data. In this work, we investigate the use of geometric semantic operators (GSOs) in VE_GP, comparing its performance with traditional GP with GSOs. Experiments are conducted on two real panel data forecasting problems, one allowing the aggregation on moving windows, one not. Results show that classical VE_GP is the best approach in both cases in terms of predictive accuracy, suggesting that GSOs are not able to evolve efficiently individuals when time series are involved. We discuss the possible reasons of this behaviour, to understand how we could design valuable GSOs for time series in the future.