Acceleration of differential evolution for aerodynamic design
Timothy P. Rogalsky · Mspace (University of Manitoba) · 2004
It has been demonstrated that Differential Evolution (DE) is a robust optimizer for aerodynamic desþ of fan blade profiles, but it can require 50,000 flow calculations to converge to a solution.This is feasible with only the simplest aerodynamic model.Accelerated convergence is required for the design algorithm to be more useful.This thesis presents, as benchmarks, convergence rates for three design cases using Bezier parameterization of airfoils and optimizing with DE.These benchmark rates are accelerated in two ways.First, an improved solution space is provided by Bezier-PARSEC airfoil parameterization.To compare their representation abilities, Bezier and Bezier-PARSEC parameterizations are used to reproduce 63 airfoils.Second, DE is modified in three different ways to provide improved convergence characteristics with the new parameterization: 1) A new selection operator is introduced, variable birthrate, which can bias the search toward the most promising regions of the solution space, 2)DE is hybridized with Downhill Simplex, a local search method, and 3) DE is accelerated by an algorithm modeling the biological immune system.The most successful strategy is Hybridized Immune Accelerated DE (HIADE).Using the BP 3333 parameterization, it converges within 10,000 flow calculations, four to ten times faster than the benchmarks.