In an important paper, Hinton and Nowlan (1987) demonstrate the Baldwin effect in a simple Genetic Algorithm. The ability of the phenotype to adapt, coupled with the evolutionary process, allows behavioural goals to become over time genetically specified; this seems Lamarckian but is not. In that paper, as a subsidiary point, the slowness of fixation of the last few goals is commented on, and a later paper by Belew (1989) attempts an analysis. In this paper I show that genetic drift is the explanation for this slowness phenomenon. Using a diffusion equation approach, I give an analysis of genetic drift for genetic algorithms, where it is too often ignored. Critical relationships between mutation rate, population size, and forces of selection are given which decide whether genetic drift will be of significance or not.
Download compressed postscript file