Genetic Algorithms (GAs) are increasingly used for such purposes as deriving programs [Koza, 1992] and producing designs for robots [Cliff et al. 1993]. According to the building-block hypothesis and schema analysis of Holland  the GA is an efficient search method. However, empirical work has shown that in some cases the method is outperformed by simpler processes such as random-permutation hill climbing [Forrest and Mitchell, 1996] and [Lang, 1995]. The present paper reexamines Holland's framework (as formulated by Goldberg ) and finds that such in-practice failures are predictable given the implications of the schema analysis. The high efficiency of the GA method is commonly attributed to its `implicit parallelism', i.e., its ability to develop candidate solutions in parallel, without focussing on any particular solution at any one time. However, this efficiency is hard to realise because there is a deep contradiction between the building-block hypothesis and the schema theorem.
Download compressed postscript file