An experimental study on dynamic random variation of population size

This paper presents an empirical comparative study of evolutionary algorithms with the purpose of determining if, for an evolutionary algorithm (EA), there exists any intrinsic advantage in using a dynamical population size control strategy over an initial arbitrarily setting of the population size, even without any explicitly defined control strategy. A brief survey of previous work on population size parameter control is presented, covering both static and dynamical methods. A classification framework for dynamical control of population size in EAs is proposed. Several strategies are proposed, characterized and applied to well-known binary and numerical test functions, both uni- and multi-modal, and with different degrees of complexity. For the case of the simple generational genetic algorithm, different dynamical strategies and the fixed population size are compared, in terms of the best absolute fitness and the improvement capability. Results indicate that, when no previous information exists, choosing a dynamic random variation control strategy for the population size is a reasonable choice, outperforming blind choices for the fixed settings.