Analysis of GP Improvement Techniques over the Real-World Inverse Problem of Ocean Color

Abstract

This paper is a follow-up of Maarten Keijzer's award-winning EUROGP'03 paper [Kei03], that suggests using Interval Arithmetic (IA) and Linear Scaling (LS) in Genetic Programming algorithms. The ideas exposed in this paper were so nice that it was decided to experiment with them on a real-world problem on which the LIL research team had some experience and results with: the Ocean Color Inverse Problem. After extensive testing of IA, LS as well as a progressive learning method using thresholds (T), results seem to show that functions evolved with GP algorithms that do not implement IA may output erroneous values outside the learning set, while LS and T methods produce solutions with a greater generalisation error. A simple and apparently harmless improvement over standard GP is also proposed, that consists in weighting operands of + and - operators. © Springer-Verlag 2004.