Potholes on the Royal Road

TL;DR

This paper investigates how evolutionary algorithms navigate fitness landscapes, aiming to identify functions where simple genetic algorithms outperform hillclimbers, thereby testing foundational theories like the building-block hypothesis.

Contribution

It introduces a heuristic for discovering functions that favor genetic algorithms and compares their performance on pothole functions, extending previous research on fitness landscape analysis.

Findings

Genetic algorithms outperform hillclimbers on certain pothole functions

The heuristic effectively identifies functions suited for genetic algorithms

Results challenge or support the building-block hypothesis

Abstract

It is still unclear how an evolutionary algorithm (EA) searches a fitness landscape, and on what fitness landscapes a particular EA will do well. The validity of the building-block hypothesis, a major tenet of traditional genetic algorithm theory, remains controversial despite its continued use to justify claims about EAs. This paper outlines a research program to begin to answer some of these open questions, by extending the work done in the royal road project. The short-term goal is to find a simple class of functions which the simple genetic algorithm optimizes better than other optimization methods, such as hillclimbers. A dialectical heuristic for searching for such a class is introduced. As an example of using the heuristic, the simple genetic algorithm is compared with a set of hillclimbers on a simple subset of the hyperplane-defined functions, the pothole functions.