Genetic programming using chebishev polynomials
Nikolay Nikolaev, Hitoshi Iba · 2001
This paper proposes a tree-structured representation for genetic programming (GP) using Chebishev polynomials as building blocks. They are incorporated in the leaves of tree-structured polynomial models. These trees are used in a version of the GP system STROGANOFF to avoid overfitting with the data when searching for polynomials. Search contro is organized with a statistical fitness function that favours accurate, predictive, and parsimonious polynomials. The improvement of the evolutionary search performance is studied by principal component analysis of the error variations of the elite individuals in the population. Empirical results show that the novel version outperforms STROGANOFF, and the traditional Kozastyle GP on processing benchmark and real-world time series.