Operational efficiency in last-mile logistics (LML) is often hindered by fluctuating e-commerce demand, unforeseen disruptions, and diverse stakeholders with evolving objectives. This paper aims to evaluate the effectiveness of Physical Internet hubs (PI-hubs) in addressing LML challenges by developing an intelligent multi-agent system (iMAS) that focuses on stakeholders’ interactions. In the iMAS, carriers, shippers, and Physical Internet managers (PI-Managers) are considered learning agents. In this complex scenario, the distribution network (DN) structure is dynamic, transitioning from a single-tier system to a two-tier network when carriers and shippers utilize PI-hubs. Bayesian Q-learning optimizes action selection by balancing exploration and exploitation, while fair reward distribution aligns agent incentives, improving cooperation, stability, and performance in dynamic, multi-agent environments. Simulations involving varying combinations of learning agents are performed. Two delivery vehicle types are also included in the collaborative vehicle routing problem, forming the iMAS environment. The simulation results are compared with the base case where agents do not engage in learning. Findings suggest that when PI-managers engage in learning, there is an increase in the percentage of PI-hub usage and a decrease in total vehicle kilometers traveled (VKT), highlighting the effectiveness of PI-hubs in alleviating the adverse impacts of freight vehicle mobility within metropolitan areas. The impact of the initial PI-hub fee policy on DN efficiency, including PI-hub usage, VKT, carriers’ and shippers’ costs, and PI-Manager profit, is assessed through extensive sensitivity analysis. The iMAS acts as a decision support system enabling policymakers to evaluate various policies and actions, aiding the identification of optimal decisions within the LML framework.

中文翻译：

---

用于最后一公里物流的智能多代理系统

最后一英里物流 （LML） 的运营效率经常受到电子商务需求波动、不可预见的中断以及目标不断变化的不同利益相关者的阻碍。本文旨在通过开发专注于利益相关者互动的智能多代理系统 （iMAS） 来评估物理互联网中心 （PI-hubs） 在应对 LML 挑战方面的有效性。在 iMAS 中，运营商、托运人和物理互联网经理 （PI-Manager） 被视为学习代理。在这种复杂的场景中，分销网络 （DN） 结构是动态的，当承运商和托运人使用 PI 中心时，它会从单层系统过渡到两层网络。贝叶斯 Q-learning 通过平衡探索和开发来优化作选择，而公平的奖励分配则使代理激励保持一致，从而在动态的多代理环境中提高合作、稳定性和性能。执行涉及学习代理的不同组合的模拟。协作车辆配送 （VRP） 问题中还包括两种配送车辆类型，形成 iMAS 环境。将模拟结果与智能体不参与学习的基本情况进行比较。研究结果表明，当 PI 经理参与学习时，PI 中心使用的百分比会增加，车辆总行驶里程 （VKT） 会减少，这凸显了 PI 中心在减轻大都市地区货运车辆流动性的不利影响方面的有效性。通过广泛的敏感性分析，评估初始 PI-hub 费用政策对 DN 效率的影响，包括 PI-hub 使用率、VKT、承运商和托运人的成本以及 PI-Manager 利润。 iMAS 充当决策支持系统，使政策制定者能够评估各种政策和行动，帮助在 LML 框架内确定最佳决策。