Asynchronous Parallel Computation of Lagrangian Bounds for Mixed-Integer Stochastic Programs by Frank-Wolfe Methods within Projective Hedging

Abstract

We describe an asynchronous parallel method to calculate Lagrangian bounds for mixed-integer stochastic programming problems. As in some prior work, we use a fully corrective Frank-Wolfe method on subproblems within a decomposition method, eventually optimizing subproblems for individual scenarios over the convex hulls of their integer-feasible vertices. However, instead of using progressive hedging as our decomposition algorithm, we employ an asynchronous projective hedging method. Like our earlier work on projective hedging, our algorithm need only solve a subset of the scenario subproblems at each iteration. Here, we further exploit the capabilities of projective hedging to create a parallel implementation capable of fully asynchronous operation. As we demonstrate empirically, this capability can lead to significant efficiency improvements when the subproblems are solved by Frank-Wolfe and have highly variable solve times.