Control of Parallel Population Dynamics by Social-Like Behavior of GA-Individuals

A frequently observed difficulty in the application of genetic algorithms to the domain of optimization ariscs from premature convergence. In order to preserve genotype diversity we develop a new model of auto-adaptive behavior for individuals. In this model a population member is an active individual that assumes social-like behavior patterns. Different individuals living in the same population can assume different patterns. By moving in a hierarchy of "social states" individuals change their behavior. Changes of social state are controlled by arguments of plausibility. These arguments are implemented as a rule set for a massively-parallel genetic algorithm. Computational experiments on 12 large-scale job shop benchmark problems show that the results of the new approach dominate the ordinary genetic algorithm significantly.