Initial Population Influence on Hypervolume Convergence of NSGA-III
Johannes Glamsch, Tobias Rosnitschek, Frank Rieg · International Journal of Simulation Modelling · 2021
A common method for solving multi-objective optimization problems are evolutionary algorithms (EA), which are utilizing an iterative population-based approach and do not need prior information about the problem to be solved.These algorithms require a variety of control parameters, e. g. the three evolutionary operators (selection, crossover and mutation), a termination criterion and the population size, which are subject of many studies.In contrast to these a less considered factor is the initialization of the first population.This paper analyses the influence of different initialization methods besides the classic sampling with a pseudo-random number generator on the convergence behaviour of the algorithm NSGA-III.It can be shown that different sampling methods affect the convergence behaviour significantly, whereby some methods increase while others decrease the convergence speed.The results also show a strong dependency and interaction between the initialization method and the optimization problem.