4 Discussion of the Metaphor of Hill Climbing

Lineage selection changes the evolutionary dynamics of diversity according to edit distance and phenotypic entropy. The measures of diversity, when increased, are expected to decrease the chance that genetic programming will become stuck in local optima. However, only on the Ant problem did fitness improve. Lineage selection increases diversity by shifting the focus of selection away from the best fit individuals to the fit individuals from different lineages. Note that normal selection pressure is returned after lineages are lost. Why does the Ant problem benefit from reduced selection pressure and added diversity, and why does improving diversity and reducing selection negatively affect fitness on the Binomial-3 and Parity problems? The metaphor of genetic programming performing a type of hill-climbing search is now examined to help understand the results.

In standard genetic programming, the convergence of the population to similar programs leads recombination to be characterised as a type of hill-climber. Thus, one may think of the beginning of a genetic programming run as a short, parallel search period until convergence occurs. At that point, recombination coupled with selection pressure (or elitism) and a converged population behaves like a hill-climber on a single program. If this is considered as a metaphor for standard genetic programming search, then what changes to the algorithm might weaken or strengthen performance?