Proximity control in bundle methods for convex nondifferentiable minimization (Q911994)

The author discusses a mathematical programming problem involving minimization of a convex, possibly nondifferentiable, function. The considered approach is the proximal bundle method. The method is based on the approximation of the objective function at a current iteration \(x_ k\) by a sum of a polyhedral function and the quadratic term \(u_ k\| x-x_ k\|^ 2\). The author argues that the usual choice of \(u_ k=1\) may yield a very slow convergence of the method, and proposes a technique for choosing \(\{u_ k\}\) adaptively that eliminates sensitivity to objective scaling. Some preliminary numerical experience reported by the author suggests that the method can compete successfully with other descent algorithms.