A Two-Level Genetic Algorithm for Large Optimization Problems

Abstract

Many local two-level algorithms have been proposed for accelerating the electromagnetic optimization by stochastic algorithms. These algorithms use a combination of a coarse and a fine model in the optimization procedure. Despite the good results, the global convergence properties represent an important drawback of these approaches. A global two-level algorithm had been proposed to deal with the convergence problems, but the requirement to refine the global surrogate model in each step can demand high computational time. This paper introduces a global two-level genetic algorithm that uses single predefined coarse and fine surrogate models, which are defined as an artificial neural network nonlinear regression of a preliminary set of finite element simulations. The benchmark test problem, Hartmann 6, and the problem dealing with the eight-parameter design of superconducting magnetic energy storage have been analyzed..