• Citer

Formulations and Algorithms for General and Security Stackelberg Games

par Casorran Amilburu, Carlos
Président du jury Massart, Thierry
Promoteur Labbé, Martine ;Ordóñez, Fernando
Publication Non publié, 2017-10-16

Thèse de doctorat

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

General Stackelberg games (GSG) confront two contenders, each wanting to optimize their rewards. One of the players, referred to as the leader, can commit to a given action or strategy first, and the other player, referred to as the follower, then responds by selecting an action or strategy of his own. The objective of the game is for the leader to commit to a reward-maximizing strategy, anticipating that the follower will best respond.Finding an optimal mixed strategy for the leader in a GSG is NP-hard when the leader faces one out of a group of several followers and polynomial when there exists a single follower. Additionally, GSGs in which the strategies of the leader consist in covering a subset of at most $m$ targets and the strategies of the followers consist in attacking some target, are called Stackelberg security games (SSG) and involve an exponential number of pure strategies for the leader.The goal of this thesis is to provide efficient algorithms to solve GSGs and SSGs. These algorithms must not only be able to produce optimal solutions quickly, but also be able to solve real life, and thus large scale, problems efficiently. To that end, the main contributions of this thesis are divided into three parts:First, a comparative study of existing mixed integer linear programming (MILP) formulations is carried out for GSGs, where the formulations are ranked according to the tightness of their linear programming (LP) relaxations. A formal theoretical link is established between GSG and SSG formulations through projections of variables and this link is exploited to extend the comparative study to SSG formulations. A new strong SSG MILP formulation is developed whose LP relaxation is shown to be the tightest among SSG formulations. When restricted to a single attacker type, the new SSG formulation is ideal, i.e., the constraints of its LP relaxation coincide with its convex hull of feasible solutions. Computational experiments show that the tightest formulations in each setting are the fastest. Notably, the new SSG formulation proposed is competitive with respect to solution time, and due to the tightness of its LP relaxation, it is better suited to tackle large instances than competing formulations.Second, the bottleneck encountered when solving the formulations studied in the first part of the thesis is addressed: The tightest formulations in each setting have heavy LP relaxations which can be time-consuming to solve and thus limit the effectiveness of the formulations to tackle instances. To address this issue, in both the general and the security case, Benders cuts from the LP relaxation of the tightest MILP formulations are embedded into a Cut and Branch scheme on a sparse equivalent formulation in each setting. By combining the tightness of the bound provided by the strong formulations with the resolution speed of the formulations, the proposed algorithm efficiently solves large GSG and SSG instances which were out of the scope of previous methods.Third, a special type of SSG, defined on a network, is studied, where the leader has to commit to two coverage distributions, one over the edges of the network and one over the targets, which are contained inside the nodes. A particular case of this SSG is used to tackle a real life border patrol problem proposed by the Carabineros de Chile in which the use of their limited security resources is optimized while taking into account both global and local planning considerations. A methodology is provided to adequately generate the game's parameters. Computational experiments show the good performance of the approach and a software application developed for Carabineros to schedule their border resources is described.