A successive quadratic programming algorithm with global and superlinear convergence properties

From MaRDI portal





The author suggests an algorithm for solving the nonlinear programming problem (NPL): Minimize f(x) subject to \(c_ i(x)\leq 0\) \((i=1,...,m')\), \(c_ i(x)=0\), \((i=m'+1,...,m)\) where the functions f, \(c_ i\) are twice continuously differentiable. In order to ensure a fast ultimate convergence, direction-finding subproblems are derived from the second- order approximation to (NLP) which in itself is not tractable and is modified to a quadratic programming problem. The author claims that this modification has the added advantage of avoiding the Maratos effect, a phenomenon where the use of a nonsmooth exact penalty function results in the deterioration of the desirable local convergence property. Such devices have been incorporated in algorithms for ensuring global and superlinear convergence properties by various authors including \textit{T. F. Coleman} and \textit{A. R. Conn} [Math. Program. 24, 123-136, 137-161 (1982; Zbl 0501.90077, Zbl 0501.90078)] and \textit{R. M. Chamberlin, M. J. D. Powell, C. Lemarechal} and \textit{H. C. Pedersen} [Math. Program. Study 16, 1-17 (1982; Zbl 0501.90078)]. The algorithm suggested for solving (NLP) either terminates at a Kuhn-Tucker point of (NLP) or generates an infinite sequence \(\{\) \(x\}\) whose limit points are Kuhn-Tucker points of (NLP). Under suitable assumptions, it is shown that the sequence \(\{\) \(x\}\) generated by the algorithm converges superlinearly.




Cited in
(52)








This page was built for publication: A successive quadratic programming algorithm with global and superlinear convergence properties

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3731379)