Using Genetic Algorithms with Small Populations
Colin R. Reeves · 1993
Most reported (serial) implementations of genetic algorithms have assumed population sizes of at least 30, and often very much larger. The general question of population sizing has been considered from several aspects, with somewhat conflicting conclusions, and then only for populations of binary strings. In this paper we consider applications where it is important to use as small a population as possible, where the number of fitness evaluations is limited, and where nonbinary representations are important. First, we approach the question of specifying a minimum size by asking what characteristics of an initial population are likely to lead to poor performance, and calculate population sizes which will almost certainly avoid such features, on the assumption that initial populations are chosen in a random fashion. It will be shown that rather small populations suffice for binary strings, but that populations need to be considerably larger for the more general case where string alleles a...