Floating Data Window Movement Influence to Genetic Programming Algorithm Efficiency

Abstract

Presented paper deals with problem of large data series modeling by genetic programming algorithm. The need of repeated evaluation constraints size of training data set in standard Genetic Programming Algorithms (GPAs) because it causes unacceptable number of fitness function evaluations. Thus, the paper discusses possibility of floating data window use and brings results of tests on large training data vector containing 1 million rows. Used floating window is small and for each cycle of GPA it changes its position. This movement allows to incorporate information contained in large number of samples without the need to evaluate all data points contained in training data in each GPA cycle. Behaviors of this evaluation concept are demonstrated on symbolic regression of Lorenz attractor system equations from precomputed training data set calculated from original difference equations. As expected, presented results points that the algorithm is more efficient than evaluating of whole data set in each cycle of GPA.