Embracing premature convergence: the Hypergamous Parallel GeneticAlgorithm

摘要

Traditional genetic algorithm (GA) approaches have emphasizedrecombination as a dominant heuristic but have typically includedmutation as a critical contributor to effective global search dynamics.Mutation has been widely regarded as an important “insurancepolicy” to avoid the well-known problem of premature convergencein the GA population. This paper introduces an alternative geneticalgorithm that uses only recombination as its global search operator.Rather than attempting to avoid problematic premature convergence, theHypergamous Parallel Genetic Algorithm (HPGA) embraces prematureconvergence as exponential-like convergence to some (probabilistic)local optimum. Local optima discovered in multiple subpopulations aretransferred to “melting pots” in which new rounds ofrecombination are used to search for more global optima. In lieu ofmutation, multiple randomly initialized subpopulations serve as theultimate source of diversity within the HPGA population