A branch-and-cut algorithm for submodular interdiction games

TL;DR

This paper introduces an exact branch-and-cut algorithm for submodular interdiction games, effectively solving complex interdiction problems with improved performance over existing MIBLP methods.

Contribution

The authors develop a novel branch-and-cut framework utilizing interdiction cuts that exploit submodularity, advancing solution methods for interdiction games with submodular objectives.

Findings

The proposed algorithm outperforms MIBLP solvers on benchmark instances.

Interdiction cuts significantly improve solution efficiency.

The method successfully solves most instances to near-optimality.

Abstract

Many relevant applications from diverse areas such as marketing, wildlife conservation, or defending critical infrastructure can be modeled as interdiction games. In this work, we introduce interdiction games whose objective is a monotone and submodular set function. Given a ground set of items, the leader interdicts the usage of some of the items of the follower in order to minimize the objective value achievable by the follower, who seeks to maximize a submodular set function over the uninterdicted items subject to knapsack constraints. We propose an exact branch-and-cut algorithm for these kind of interdiction games. The algorithm is based on interdiction cuts which allow to capture the follower's objective function value for a given interdiction decision of the leader and exploit the submodularity of the objective function. We also present extensions and liftings of these cuts and discuss additional preprocessing procedures. We test our solution framework on the weighted maximal covering interdiction game and the bipartite inference interdiction game. For both applications, the improved variants of our interdiction cut perform significantly better than its basic version. For the weighted maximal covering interdiction game for which a mixed-integer bilevel linear programming (MIBLP) formulation is available, we compare the results with those of a state-of-the-art MIBLP solver. While the MIBLP solver yields a minimum of 54% optimality gap within one hour, our best branch-and-cut setting solves all but 4 of 108 instances to optimality with a maximum of 3% gap among unsolved ones.