Comparison of two sets of first-order conditions as bases of interior-point Newton methods for optimization with simple bounds (Q700699)

The authors compare the behavior of two Newton interior-point methods derived from two different first-order necessary conditions for the same nonlinear optimization problem with simple bounds. One set of conditions was proposed by Coleman and Li (CL conditions); the other is the standard Karush-Kuhn-Tucker (KKT) set of conditions. In the paper, optimization problems of convex quadratic functions with simple bounds on variables are considered. In order to prevent the behaviour of sticking to boundaries the right-hand sides of the equations in the KKT and CL conditions are perturbed. For the KKT conditions, the perturbed formulations possesses better global convergence properties than the unperturbed ones. The perturbed CL conditions work well for the case of one-sided bounds, but do not work well at all for the case of general bounds. The authors discuss the difficulties encountered when trying to derive a perturbation of the CL conditions for problems with general bounds. On the sets of test problems used in the paper, the perturbed KKT algorithm is the more robust of the interior-point Newton algorithms.