Resource-Aware Cost-Sharing Mechanisms with Priors

TL;DR

This paper introduces resource-aware cost-sharing mechanisms that leverage prior information about users to achieve low price of anarchy in selfish scheduling games, extending applicability beyond convex or concave cost functions.

Contribution

It demonstrates that incorporating minimal prior information about users enables resource-aware mechanisms to maintain low PoA across a broader range of cost functions.

Findings

A mechanism with knowledge of two users guarantees constant PoA.

Using probabilistic user presence, the mechanism achieves logarithmic expected PoA.

Enhanced mechanisms outperform previous models for non-convex, non-concave cost functions.

Abstract

In a decentralized system with $m$ machines, we study the selfish scheduling problem where each user strategically chooses which machine to use. Each machine incurs a cost, which is a function of the total load assigned to it, and some cost-sharing mechanism distributes this cost among the machine's users. The users choose a machine aiming to minimize their own share of the cost, so the cost-sharing mechanism induces a game among them. We approach this problem from the perspective of a designer who can select which cost-sharing mechanism to use, aiming to minimize the price of anarchy (PoA) of the induced games. Recent work introduced the class of \emph{resource-aware} cost-sharing mechanisms, whose decisions can depend on the set of machines in the system, but are oblivious to the total number of users. These mechanisms can guarantee low PoA bounds for instances where the cost functions of the machines are all convex or concave, but can suffer from very high PoA for cost functions that deviate from these families. In this paper we show that if we enhance the class of resource-aware mechanisms with some prior information regarding the users, then they can achieve low PoA for a much more general family of cost functions. We first show that, as long as the mechanism knows just two of the participating users, then it can assign special roles to them and ensure a constant PoA. We then extend this idea to settings where the mechanism has access to the probability with which each user is present in the system. For all these instances, we provide a mechanism that achieves an expected PoA that is logarithmic in the expected number of users.