Improved Price of Anarchy via Predictions

TL;DR

This paper introduces decentralized mechanisms enhanced with predictions to improve the efficiency of multiagent systems, analyzing their performance in scheduling and network formation games based on prediction accuracy.

Contribution

It proposes a novel approach of integrating predictions into decentralized mechanisms to reduce the price of anarchy in complex game-theoretic settings.

Findings

Mechanisms with predictions outperform traditional methods under certain prediction accuracies.

The price of anarchy decreases as prediction error diminishes.

Applicable to scheduling and multicast network formation games.

Abstract

A central goal in algorithmic game theory is to analyze the performance of decentralized multiagent systems, like communication and information networks. In the absence of a central planner who can enforce how these systems are utilized, the users can strategically interact with the system, aiming to maximize their own utility, possibly leading to very inefficient outcomes, and thus a high price of anarchy. To alleviate this issue, the system designer can use decentralized mechanisms that regulate the use of each resource (e.g., using local queuing protocols or scheduling mechanisms), but with only limited information regarding the state of the system. These information limitations have a severe impact on what such decentralized mechanisms can achieve, so most of the success stories in this literature have had to make restrictive assumptions (e.g., by either restricting the structure of the networks or the types of cost functions). In this paper, we overcome some of the obstacles that the literature has imposed on decentralized mechanisms, by designing mechanisms that are enhanced with predictions regarding the missing information. Specifically, inspired by the big success of the literature on "algorithms with predictions", we design decentralized mechanisms with predictions and evaluate their price of anarchy as a function of the prediction error, focusing on two very well-studied classes of games: scheduling games and multicast network formation games.