Introducing performance landscapes and a generic framework for evolutionary search algorithms

Abstract

Introduces a generic framework for evolutionary search algorithms (ESAs) and shows how we can use it to imagine the space of all possible ESAs. The idea of the performance landscape for a given search problem is then introduced in relation to this space of an possible ESAs. The author has previously (1999) argued that search problems can be usefully classified by the most effective ESA on that problem, which translates into the highest peak of the performance landscape. The reason for introducing performance landscapes is to get a better feel for the implications of the no-free lunch theorem, together with the notion of real-world problems. These are discussed in relation to performance landscapes. Finally, using the generic framework, a simple generic ESA is described and then used to glean a view of the performance landscape of some NK fitness landscapes (S. Kauffman, 1993), with N=20 and K=2, 5 or 10, and a small real-world problem. With these results, a simple classification of each search problem is made.