To Appear in SIAM J . on Optimization INCREMENTAL SUBGRADIENT METHODS 1 FOR NONDIFFERENTIABLE OPTIMIZATION

We consider a class of subgradient methods for minimizing a convex function that consists of the sum of a large number of component functions. This type of minimization arises in a dual context from Lagrangian relaxation of the coupling constraints of large scale separable problems. The idea is to perform the subgradient iteration incrementally, by sequentially taking steps along the subgradients of the component functions, with intermediate adjustment of the variables after processing each component function. This incremental approach has been very successful in solving large differentiable least squares problems, such as those arising in the training of neural networks, and it has resulted in a much better practical rate of convergence than the steepest descent method. In this paper, we establish the convergence properties of a number of variants of incremental subgradient methods, including some that are stochastic. Based on the analysis and computational experiments, the methods appear very promising and effective for important classes of large problems. A particularly interesting discovery is that by randomizing the order of selection of component functions for iteration, the convergence rate is substantially improved. 1 Research supported by NSF under Grant ACI-9873339. 2 Dept. of Electrical Engineering and Computer Science, M.I.T., Cambridge, Mass., 02139.