On the Complexity of Inverse Mixed Integer Linear Optimization

TL;DR

This paper explores the computational complexity of inverse mixed integer linear optimization, revealing its equivalence to the separation problem and establishing its placement within complexity classes like coNP-complete and D^P.

Contribution

It establishes the complexity classifications of inverse MILP problems and their close relationship with forward and separation problems, providing new insights into their computational nature.

Findings

Inverse MILP decision problem is coNP-complete.

Primal bound verification for inverse and forward problems is D^P-complete.

A cutting-plane algorithm links inverse problems to the separation problem.

Abstract

Inverse optimization is the problem of determining the values of missing input parameters for an associated forward problem that are closest to given estimates and that will make a given target vector optimal. This study is concerned with the relationship of a particular inverse mixed integer linear optimization problem (MILP) to both the forward problem and the separation problem associated with its feasible region. We show that a decision version of the inverse MILP in which a primal bound is verified is coNP-complete, whereas primal bound verification for the associated forward problem is NP-complete, and that the optimal value verification problems for both the inverse problem and the associated forward problem are complete for the complexity class D^P. We also describe a cutting-plane algorithm for solving inverse MILPs that illustrates the close relationship between the separation problem for the convex hull of solutions to a given MILP and the associated inverse problem. The inverse problem is shown to be equivalent to the separation problem for the radial cone defined by all inequalities that are both valid for the convex hull of solutions to the forward problem and binding at the target vector. Thus, the inverse, forward, and separation problems can be said to be equivalent.