Monte Carlo Optimization for Solving Multilevel Stackelberg Games

TL;DR

This paper introduces a stochastic algorithm to approximate solutions for complex multilevel Stackelberg games, which are difficult to solve due to their nested leader-follower structures and NP-hardness.

Contribution

The paper presents a novel stochastic algorithm for approximating local equilibria in multilevel Stackelberg games, including convergence proof and empirical performance analysis.

Findings

Algorithm effectively approximates solutions within reasonable error margins.

Converges asymptotically with proven theoretical guarantees.

Outperforms existing solution methods in experiments.

Abstract

Stackelberg games originate where there are market leaders and followers, and the actions of leaders influence the behavior of the followers. Mathematical modelling of such games results in what's called a Bilevel Optimization problem. There is an entire area of research dedicated to analyzing and solving Bilevel Optimization problems which are often complex, and finding solutions for such problems is known to be NP-Hard. A generalization of Stackelberg games is a Multilevel Stackelberg game where we may have nested leaders and followers, such that a follower is, in turn, a leader for all lower-level players. These problems are much more difficult to solve, and existing solution approaches typically require extensive cooperation between the players (which generally can't be assumed) or make restrictive assumptions about the structure of the problem. In this paper, we present a stochastic algorithm to approximate the local equilibrium solutions for these Multilevel games. We then construct a few examples of such Multilevel problems, including: a) a nested toll-setting problem; and b) an adversarial initial condition determination problem for Robust Trajectory Optimization. We test our algorithm on our constructed problems as well as some trilevel problems from the literature, and show that it is able to approximate the optimum solutions for these problems within a reasonable error margin. We also provide an asymptotic proof for the convergence of the algorithm and empirically analyze its accuracy and convergence speed for different parameters. Lastly, we compare it with existing solution strategies from the literature and demonstrate that it outperforms them.