Mixed method for solving the general convex programming problem (Q1288666): Difference between revisions

A method for solving the following convex optimization problem is suggested: \[ \text{minimize }f(x) \text{ subject to }x\in\{x\in \mathbb{R}^n\mid f_j(x)\leq 0,\;j=1,\dots, m,\;x\in M\}, \] where \(f,f_j\) are continuous convex functions and \(M\) is a convex polyhedron. The proposed method combines ideas of the following three methods known from the literature: The linearization method, the cutting plane method and the exact penalty method. The algorithm has two main features: (1) It converges under very general assumptions; (2) It provides an estimate for the Kuhn-Tucker multiplier. Since the theoretical version of the algorithm, the convergence of which the authors proved, may involve accumulating a large number of constraints, the authors suggested in the paper also a modified version of the algorithm, which overcomes this disadvantage. Although the convergence of this modified version of the algorithm was not proved, the authors report about its good efficiency and reliability during tests on numerical examples.