A splitting method for stochastic programs

This paper derives a new splitting-based decomposition algorithm for convex stochastic programs. It combines certain attractive features of the progressive hedging algorithm of Rockafellar and Wets, the dynamic splitting algorithm of Salinger and Rockafellar and an algorithm of Korf. We give two derivations of our algorithm. The first one is very simple, and the second one yields a preconditioner that resulted in a considerable speed-up in our numerical tests. The first author was supported by the Finnish Foundation for Economic Education under grants 20728 and 21599, and by Jenny ja Antti Wihuri Foundation.