The Recombination Operator, its Correlation to the Fitness Landscape and Search Performance

Greg Hornby Masters Thesis Fall 1996

Abstract

A common misconception in evolutionary algorithms (EAs) is that one recombination operator is universally better than another. In fact, a recombination operator will only get better performance on a function if it incorporates some knowledge about that function -- called tuning it to the function's fitness landscape. In this thesis we identify three ways in which a recombination operator can be tuned to a real-valued landscape: distance, directionality, and distributional bias. We empirically show that a directionally tuned recombination operator gives better search performance than an untuned operator. We also show that a recombination operator that is tuned to one landscape can be mis-tuned to a similar landscape. In addition we find several surprises that contradict our initial intuition but yield to further analysis. For example one interesting observation is a decrease in the number of individuals on the global optimum. We show this to be caused by the attractive pull of a larger group of individuals on a peak with a larger basin of attraction.