Genetic Algorithms

In the rest of this monograph I will deal exclusively with a special technique from the field of computational intelligence research, namely genetic algorithms. Genetic algorithms were developed by Holland [87] in 1975 as a tool to find solutions of optimization problems in poorly understood large spaces. They are based on the genetic processes of biological organisms, especially on the principle of natural selection that has become famous as “survival of the fittest” since the publishing of The Origin of Species by Charles Darwin [42]. Although, this slogan seems to be slightly tautological in the natural environment, where fitness is defined as the ability to survive, it makes good sense in the world of optimization problems, where the fitness of a string is given as the value of the function to be optimized at the argument encoded by the string. GAs proved to be quite successful in finding good solutions to such complex problems as the travelling salesman problem, the knapsack problem, large scheduling problems, graph partitioning problems, but also for engineering problems like the design of bridge structures or the optimal use of power plants. The most famous application is perhaps one of the first large applications done by Goldberg. By using classifier systems based on GAs he was able to generate a control system that managed the flow in the large gas pipeline from Texas to Chicago (Goldberg [72]). In 1989 Goldberg wrote a seminal book dealing with genetic algorithms, that is up to now probably the most widely spread book in this field of research (Goldberg [74]). Besides the principle of natural selection, GAs imitate not only the spreading of genetic material in a population but also the generation of new genetic material by mutations. Other features like the duality of genes or the existence of dominant and recessive genes are typically neglected in artificial systems. It is, however, up to now, by no means clear which processes in nature are essential for the process of evolution and which are not, so that the choice of the features that are transferred to GAs is due mainly to implementation problems. It is quite important to realize that an artificial GA is only a very simple model of the natural evolution, and that natural evolution is basically not an optimization technique, but it just evolves. I think, therefore, that the occurrence of some process in natural evolution is per se no argument that an operator modelling this process will improve the learning behavior of a GA. The connection between the natural and artificial evolution is not close enough, so as to guarantee that any argument valid in one world will also be valid in the other.