《2016 Level bundle-like algorithms for convex optimization》.pdf

LEVEL BUNDLE-LIKE ALGORITHMS FOR CONVEX OPTIMIZATION J. Y. BELLO CRUZ1 AND WELINGTON OLIVEIRA2 Abstract. We propose two restricted memory level bundle-like algorithms for minimizing a convex function over a convex set. If the memory is restricted to one linearization of the objective function, then both algorithms are variations of the projected subgradient method. The first algorithm, proposed in Hilbert space, is a conceptual one. It is shown to be strongly convergent to the solution that lies closest to the initial iterate. Furthermore, the entire sequence of iterates generated by the algorithm is contained in a ball with diameter equal to the distance between the initial point and the solution set. The second algorithm is an implementable version. It mimics as much as possible the conceptual one in order to resemble convergence properties. The implementable algorithm is validated by numerical results on several two-stage stochastic linear programs found in the literature. Keywords: Convex minimization, Nonsmooth optimization, Level bundle method, Strong convergence. 1. Introduction We are interested in solving problems of the form (1) f∗ := min f (x) , x∈C where C is a nonempty convex and closed subset of a real Hilbert space H with inner product ·, ·, and f : H → is a lower semicontinuous and convex function, possibly nondifferentiable. It is well known that lower semicontinuous functions are locally bounded in Hilbert spaces and therefore, the subdifferential set