Prediction-Aware Learning in Multi-Agent Systems

TL;DR

This paper introduces a prediction-aware learning framework for multi-agent systems in time-varying games, improving theoretical guarantees and empirical performance by leveraging payoff predictions.

Contribution

It presents the POWMU algorithm, a novel extension of optimistic multiplicative weight updates that incorporates payoff predictions for better adaptation in dynamic environments.

Findings

The framework achieves performance comparable to static settings under bounded prediction errors.

POWMU provides theoretical guarantees on social welfare and convergence.

Empirical results show improved traffic routing performance.

Abstract

The framework of uncoupled online learning in multiplayer games has made significant progress in recent years. In particular, the development of time-varying games has considerably expanded its modeling capabilities. However, current regret bounds quickly become vacuous when the game undergoes significant variations over time, even when these variations are easy to predict. Intuitively, the ability of players to forecast future payoffs should lead to tighter guarantees, yet existing approaches fail to incorporate this aspect. This work aims to fill this gap by introducing a novel prediction-aware framework for time-varying games, where agents can forecast future payoffs and adapt their strategies accordingly. In this framework, payoffs depend on an underlying state of nature that agents predict in an online manner. To leverage these predictions, we propose the POWMU algorithm, a contextual extension of the optimistic Multiplicative Weight Update algorithm, for which we establish theoretical guarantees on social welfare and convergence to equilibrium. Our results demonstrate that, under bounded prediction errors, the proposed framework achieves performance comparable to the static setting. Finally, we empirically demonstrate the effectiveness of POWMU in a traffic routing experiment.