Towards Global Solutions for Nonconvex Two-Stage Stochastic Programs: A Polynomial Lower Approximation Approach

TL;DR

This paper presents a polynomial lower approximation method for solving nonconvex two-stage stochastic programs, enabling global optimality certification and efficient handling of continuous distributions or large scenario sets.

Contribution

It introduces a novel two-phase approach using polynomial lower bounds and semidefinite relaxations to find global solutions for complex stochastic programs.

Findings

Provides a polynomial lower bound for the recourse function.

Enables global optimality certification for solutions.

Effective for problems with continuous distributions or many scenarios.

Abstract

This paper tackles the challenging problem of finding global optimal solutions for two-stage stochastic programs with continuous decision variables and nonconvex recourse functions. We introduce a two-phase approach. The first phase involves the construction of a polynomial lower bound for the recourse function through a linear optimization problem over a nonnegative polynomial cone. Given the complex structure of this cone, we employ semidefinite relaxations with quadratic modules to facilitate our computations. In the second phase, we solve a surrogate first-stage problem by substituting the original recourse function with the polynomial lower approximation obtained in the first phase. Our method is particularly advantageous for two reasons: it not only generates global lower bounds for the nonconvex stochastic program, aiding in the certificate of global optimality for prospective solutions like stationary solutions computed from other methods, but it also yields an explicit polynomial approximation for the recourse function through the solution of a linear conic optimization problem, where the number of variables is independent of the support of the underlying random vector. Therefore, our approach is particularly suitable for the case where the random vector follows a continuous distribution or when dealing with a large number of scenarios. Numerical experiments are conducted to demonstrate the effectiveness of our proposed approach.