PG-DyRA: Potential-Game Guided Multi-Agent Collaboration for Dynamic Urban Resource Allocation

Abstract

Efficient allocation of urban public resources such as transportation, logistics, and emergency response is crucial for modern cities to maintain operational efficiency and service equity. This task requires coordinating multiple mobile units to serve spatiotemporally heterogeneous demands under coupled physical constraints such as finite service capacity, energy budgets, and mobility ranges. Beyond physical limitations, practical deployments also impose information constraints: sensing costs, communication bandwidth, and privacy regulations restrict each agent to observing only nearby peers, yet effective coordination demands reasoning about unobserved actions across the entire system. We propose PG-DyRA, a decentralized coordination framework that operates under partial observability without inter-agent communication. Building on a potential game perspective, we design a marginal contribution utility under which the resulting game admits pure strategy Nash equilibria that coincide with the global optimum, providing a theoretically grounded coordination signal for independent decision-making. To operationalize this signal under this observability constraint, we develop a dual-branch intention inference network that estimates other agents' service capacity allocation from local observations, with one branch modeling visible agents via trajectory-based state bank alignment and the other inferring invisible agents from demand residuals. Experiments on four urban datasets demonstrate that PG-DyRA achieves 2.23-8.04 percentage points higher task coverage rates while reducing energy consumption by 2.99-8.35 percentage points compared to the state-of-the-art methods.