Stepping stones and hidden haystacks: when a genetic algorithm defeats a hillclimber

Abstract

Following intuitive notions on gross aspects of how a GA behaves, we are able to demonstrate how to construct functions on which a GA will greatly outperform a hillclimber. This augments related work on long path problems, and gene switch cost functions, which describe similarly 'GA appropriate' landscapes but on rather less intuitively clear grounds. Although artificial, the construction of these problems relies on certain gross landscape features that may be a priori estimated in the case of many real problems, incrementing the collection of descriptive tools with which to assess potential amenability to evolutionary search. We argue in particular that a specific notion of hillclimbing behaviour can with certain merits, and with certain qualifications, be included in this collection.