Overfitting avoidance in genetic programming of polynomials

This paper proposes several techniques for avoiding overfitting in the genetic programming (GP) of polynomials. The model specification flexibility is increased by: (1) a polynomial block reformulation, which reduces the statistical bias, and, (2) complexity tuning using local ridge regression and regularized weight subset selection, which reduce the statistical variance. Another contribution is the designed fitness function for search navigation towards highly predictive models. Experimental results on time-series forecasting show that these techniques help GP to find accurate, less complex and better forecasting polynomials than traditional Koza-style GP (J.R. Koza, 1992) and the previous Stroganoff system (H. Iba et al., 1994, 2001)