Population control in evolutionary algorithms
Jason E. Cook · 2010
Traditional evolutionary algorithms (EAs) are powerful robust problem solvers that have several fixed parameters which require prior specification. Having to determine good values for any of these parameters can be problematic, as the performance of EAs is generally very sensitive to these parameters, requiring expert knowledge to set optimally without extensive use of trial and error. Parameter control is a promising approach to achieving this automation and has the added potential of increasing EA performance based on both theoretical and empirical evidence that the optimal values of EA strategy parameters change during the course of executing an evolutionary run. While many methods of parameter control have been published that focus on removing the population size parameter (μ), most of these methods have undesirable side effects for doing so. This thesis starts by providing evidence for the benefits of making μ a dynamic parameter and then introduces two novel methods for removing the need to preset μ. These methods are then compared, explaining the strengths and weaknesses of each. The benefit of employing a dynamic value for μ is demonstrated on two test problems through the use of a meta-EA, and the first novel method is shown to be useful on several binary test problems while the second performs well on a real valued test problem. A condensed version of this thesis has been accepted for publication in the proceedings of the Genetic and Evolutionary Computation Conference 2010 [7].