Bicretieria Optimization in Routing Games

TL;DR

This paper investigates bicriteria routing games focusing on maximum edge congestion and path length, analyzing stability, convergence, and bounds on the price of anarchy for different game formulations.

Contribution

It introduces and analyzes max and sum bicriteria routing games, providing bounds on stability, convergence, and the price of anarchy, including a new sum-bucket game with improved bounds.

Findings

Max games are stable and convergent with a bounded price of anarchy.

Sum games may lack Nash equilibria, leading to the proposal of sum-bucket games.

Sum-bucket games have a bounded and often superior price of anarchy, close to optimal under certain conditions.

Abstract

Two important metrics for measuring the quality of routing paths are the maximum edge congestion $C$ and maximum path length $D$. Here, we study bicriteria in routing games where each player $i$ selfishly selects a path that simultaneously minimizes its maximum edge congestion $C_i$ and path length $D_i$. We study the stability and price of anarchy of two bicriteria games: - {\em Max games}, where the social cost is $\max(C,D)$ and the player cost is $\max(C_i, D_i)$. We prove that max games are stable and convergent under best-response dynamics, and that the price of anarchy is bounded above by the maximum path length in the players' strategy sets. We also show that this bound is tight in worst-case scenarios. - {\em Sum games}, where the social cost is $C+D$ and the player cost is $C_i+D_i$. For sum games, we first show the negative result that there are game instances that have no Nash-equilibria. Therefore, we examine an approximate game called the {\em sum-bucket game} that is always convergent (and therefore stable). We show that the price of anarchy in sum-bucket games is bounded above by $C^* \cdot D^* / (C^* + D^*)$ (with a poly-log factor), where $C^*$ and $D^*$ are the optimal coordinated congestion and path length. Thus, the sum-bucket game has typically superior price of anarchy bounds than the max game. In fact, when either $C^*$ or $D^*$ is small (e.g. constant) the social cost of the Nash-equilibria is very close to the coordinated optimal $C^* + D^*$ (within a poly-log factor). We also show that the price of anarchy bound is tight for cases where both $C^*$ and $D^*$ are large.