Advanced Correlation Analysis of Operators for the Traveling Salesman Problem

An extension to the correlation analysis of Manderick et al. and Mathias and Whitley used to study TSP operators is introduced. Part correlation and partial correlation further refine the relationship between an operator's preservation of genetic information and its overall fitness. The fitness landscapes of ten genetic operators are studied in isolation and in conjunction with the 2-opt local improvement technique in a GENITOR environment. Five variables (parent fitness, child fitness, child failure rate, insertion rate and increased tour length due to failure) are considered in the analyses. Results suggest that certain operators (Edge Family, MPX and Schleuter crossover (SCHX)) maintain higher correlations over several interesting variable relations, especially the parent to child (PC) relation, than do other genetic operators.