Fitness Distance Correlation And Problem Difficulty For Genetic Programming

This work is a first step in the attempt to verify whether (and in which cases) fitness distance correlation can be a good tool for classifying problems on the basis of their difficulty for genetic programming. By analogy with the studies that have already been done on genetic algorithms, we define some notions of distance between genotypes. Then we choose one of these distances to calculate the fitness distance correlation coefficient and we use it to study the difficulty of some problems. First, we do this for a syntactically limited language. Then we extend the study to standard genetic programming. For the functions used here i.e., traps and royal trees, the results confirm that fitness distance correlation is a good predictor of genetic programming difficulty.