AMAZING-GEOMETRY-OF-GENETIC-SPACE-OR-ARE-GENETIC-ALGORITHMS-CONVERGENT？.pdfVIP

AMAZING GEOMETRY OF GENETIC SPACE OR ARE GENETIC ALGORITHMS CONVERGENT? Marek W. Gutowski Institute of Physics, Polish Academy of Sciences 02–668 Warszawa, Al. Lotnik´ow 32/46, Poland e-mail: gutow@.pl Abstract There is no proof yet of convergence of Genetic Algorithms. We do not supply it too. Instead, we present some thoughts and arguments to convince the Reader, that Genetic Algorithms are essentially bound for success. For this purpose, we consider only the crossover operators, single- or multiple-point, together with selection procedure. We also give a proof that the soft selection is superior to other selection schemes. Keywords genotype space; crossover operators; soft selection; convergence; stopping rules I. Introduction We are using Genetic Algorithms (GA’s) for solving hard global optimization problems for at least three reasons: • they are easy to implement in many computer languages, • they are applicable to problems, which cannot be easily, if at all, specified analytically as a set of closedform formulas, • we believe, that their inherent intelligence will automatically, i.e. with almost no programmer’s effort, find the way to solve, ”sufficiently well”, our difficult problems. The very idea of GA’s, to simply mimic the Nature [1], belongs mostly to the sphere of intuition, and is almost lacking a solid mathematical background. Indeed, numerical values of many important ”tuning parameters” (mutation rate, probability of selection for reproduction, etc.) are largely selected on the base of experience of other people solving probl