Detection and Repair of Poorly Converged Optimization Runs

Abstract

The use of iteratively reweighted least squares (IRLS) for detecting design points where optimizations converge poorly is demonstrated. Because optimization error tends to he one-sided with poor results producing overweight designs, a nonsymmetrical version of IRLS (NIRLS) that accounts for the asymmetry in optimization errors is also developed. A parameterized Rosenbrock function problem, where a programming error caused poor optimizations, is used to demonstrate the techniques. The identified outliers were repaired by more accurate optimizations to improve response surface approximations. Structural optimization of a high-speed civil transport (HSCT) produced inaccurate wing structural weight. Optimization studies with various sets of convergence criteria showed that the optimization errors were due to incomplete convergence. The IRLS technique could identify many of the points with very large optimization errors, but the NIRLS techniques were much more reliable in this task. However, results of HSCT configuration optimizations using wing weight response surface models show that it may be preferable to identify and repair only a subset of the poorly converged optima.