Two differential equation systems for inequality constrained optimization (Q2372013)

The inequality-constrained optimization problem \[ \min f(x)\text{ with respect to }g_i(x)\leq 0,\qquad i= 1,\dots, m \] is studied, where all real-valued functions are twice continuously differentiable. The idea is to construct a system of ordinary differential equations (ODEs), so that its equilibrium points coincide with the solutions to the inequality-constrained problem. This idea was used by \textit{Yu. G. Evtushenko} and \textit{V. G. Zhadan} [Stable barrier projection and barrier-Newton methods in nonlinear progrmming. Opt. Software 3, 237--256 (1994)] in the case of equality-constrained problems. The transformation proposed in the present paper is a new one. Then a nonlinear Lagrangian is defined by \[ F_\sigma(x, y)= f(x)+ \sigma\Sigma y^2_i(e^{Sg_i(x)}- 1),\quad S= \sigma^{-1}. \] Let \((x^*, \mu^*)\) be a Kuhn-Tucker point of the original optimization problem, necessary optimality conditions are derived using the Lagrangian. These points are asymptotically stable equilibrium points of two differential equations. The Euler discrete schemes with constant stepsizes for the two ODE-systems are given and their convergence theorems are demonstrated. An Euler discrete scheme with an Armijo line search has a quadratic convergence rate. An algorithm is given. The Runge-Kutta method for the second system has good stability and the Euler scheme with the Armijo line search rule has fast convergence. Numerical results of some classical examples published in a lecture notes of \textit{W. Hock} and \textit{K. Schittkowski} [Text examples for nonlinear programming codes. (1981; Zbl 0452.90038)] are added. Unfortunately, these examples are not repeated here.