The Impact of Selfish Behavior in Load Balancing Games

TL;DR

This paper investigates how the structure of strategy spaces in load balancing games affects the efficiency of decentralized solutions, analyzing various solution concepts and their bounds under different conditions.

Contribution

It introduces new bounds and insights on the performance of approximate equilibria and greedy algorithms in load balancing games with various player and resource configurations.

Findings

Better bounds are not achievable with weighted or asymmetric players.

Improvements are possible for symmetric, weighted players with identical resources.

Lower bounds are established for one-round walk performances in unweighted, identical resource games.

Abstract

To what extent does the structure of the players' strategy space influence the efficiency of decentralized solutions in congestion games? In this work, we investigate whether better performance are possible when restricting to load balancing games in which players can only choose among single resources. We consider three different solutions concepts, namely, approximate pure Nash equilibria, approximate one-round walks generated by selfish players aiming at minimizing their personal cost and approximate one-round walks generated by cooperative players aiming at minimizing the marginal increase in the sum of the players' personal costs. The last two concepts can be interpreted as solutions of greedy online algorithms for the related resource selection problem. We show that, under fairly general latency functions on the resources, better bounds cannot be achieved if players are either weighted or asymmetric. On the positive side, we prove that, under mild assumptions on the latency functions, improvements on the performance of approximate pure Nash equilibria are possible for load balancing games with weighted and symmetric players in the case of identical resources. We also design lower bounds on the performance of one-round walks in load balancing games with unweighted players and identical resources.