Duality and sensitivity analysis of multistage linear stochastic programs

Abstract

In this paper we investigate the dual of a Multistage Stochastic Linear Program (MSLP) to study two related questions for this class of problems. The first of these questions is the study of the optimal value of the problem as a function of the involved parameters. For this sensitivity analysis problem, we provide formulas for the derivatives of the value function with respect to the parameters and illustrate their application on an inventory problem. Since these formulas involve optimal dual solutions, we need an algorithm that computes such solutions to use them, i.e., we need to solve the dual problem.In this context, the second question we address is the study of solution methods for the dual problem. Writing Dynamic Programming equations for the dual, we can use an SDDP type method, called Dual SDDP, which solves these Dynamic Programming equations computing a sequence of nonincreasing deterministic upper bounds on the optimal value of the problem. However, applying this method will only be possible if the Relatively Complete Recourse (RCR) holds for the dual. Since the RCR assumption may fail to hold (even for simple problems), we design two variants of Dual SDDP, namely Dual SDDP with penalizations and Dual SDDP with feasibility cuts, that converge to the optimal value of the dual (and therefore primal when there is no duality gap) problem under mild assumptions. We also show that optimal dual solutions can be obtained computing dual solutions of the subproblems solved when applying Primal SDDP to the original primal MSLP.The study of this second question allows us to take a fresh look at the notoriously difficult to solve class of MSLP with interstage dependent cost coefficients. Indeed, for this class of problems, cost-to-go functions are non-convex and solution methods were so far using SDDP for a Markov chain approximation of the cost coefficients process. For these problems, we propose to apply Dual SDDP with penalizations to the cost-to-go functions of the dual which are concave. This algorithm converges to the optimal value of the problem.Finally, as a proof of concept of the tools developed, we present the results of numerical experiments computing the sensitivity of the optimal value of an inventory problem as a function of parameters of the demand process and compare Primal and Dual SDDP on the inventory and a hydro-thermal planning problems.

Multi-Stage Stochastic Programming

“…where µ t , a t are m dimensional vectors, Φ t is an m × m coefficient matrix and a t is a white noise vector process with zero mean in (2. A recent development in [3] is to consider the dual of the multistage stochastic program. The uncertainties in objective for the primal problem are then transformed into the uncertainties on right hand sides for the dual problem.…”

“…1],[2,3],[4,5]], transition_matrix=[[[1]],[[0.2,0.8]],[[0.3,0.7],[0.4,0.6]]] ) for t in range(3): m = MC[t] now, past = m.addStateVar(uncertainty_dependent=0)…”