自动驾驶车辆混行集聚MAS控制模型

随着车路协同技术和自动驾驶技术的不断发展，越来越多的网联自动驾驶车辆（Connected and Autonomous Vehicle， CAV）涌入道路交通，与传统人工驾驶车辆（Human Pilot Vehicle， HPV）形成混合交通流（Mixed Traffic Stream， MTS）。为在提高MTS交通流量的同时保证交通安全，面向未来的混行交通环境，结合交通工程中人、车、路等要素，设计基于多智能体系统的CAV集聚控制模型（Agglomeration Control Model of Connected and Autonomous Vehicle Based on Multi-Agent System，ACMCAV-MAS）。该模型针对CAV的可控性和HPV的随机性，意在通过集聚控制，促使道路中分散行驶的CAV集聚成行驶条件更优的队列。具体以Agent的形式设计与集聚控制相关的底层车辆Agent（CAV-Agent和HPV-Agent两类）和上层管理Agent。同时，针对同质要素间的匹配和异质要素间的风险规避，区别于常规的无集聚（No Agglomeration，NOA）策略，提出车队级集聚（Platoon Level Agglomeration， PLA）和车道级集聚（Lane Level Agglomeration，LLA）2种策略及相关的CAV-Agent集聚控制算法。基于ACMCAV-MAS及元胞自动机模型，在不同交通流密度和不同CAV-Agent渗透率下进行仿真试验。结果表明：集聚策略能在60%的CAV-Agent渗透率下取得最佳效益，同时，在60 veh·km^-1密度条件下，车队级集聚策略平均能提升38.14%的交通流量，比车道级集聚的提升效果高9.73%，并能在40~50 veh·km^-1的密度范围和50%~70%的CAV-Agent渗透率条件下有效缓解交通拥堵；通过对中高密度交通流下的纵向风险分析，发现2种集聚策略在低CAV-Agent渗透率下的风险发生率无显著差异，且最大风险降低比例都能达到20%以上，然而，在实际交通情况下，集聚策略可能会在一定程度上导致横向碰撞风险的增加。在未来的工作中，将继续探究降低横向碰撞风险的方法，同时着力解决目前仿真框架中对于人工驾驶行为异质性建模不够完善的缺陷，不断优化ACMCAV-MAS，为未来MTS中自动驾驶策略的制定提供理论依据。

Agglomeration Control Model Based on Multi-agents for Autonomous Vehicles in Mixed Traffic Environment

With the continuous development of cooperative vehicle infrastructure systems and automatic driving technology, an increasing number of connected and autonomous vehicles (CAVs) flow into road traffic, with traditional human-pilot vehicles(HPVs) forming mixed traffic streams (MTSs). To improve the traffic flow of the MTS and ensure traffic safety, considering that CAVs require less headway and have fewer speed fluctuations when driving in a platoon, an agglomeration control model of connected and autonomous vehicle based on multi-agent system(ACMOCAV-MAS) was designed. Based on the controllability of CAVs and the randomness of HPV, the model aimed to promote the scattered driving CAVs to agglomerate into a platoon with better driving conditions. The underlying vehicle agents (CAV agent and HPV agent) and the upper management agent were designed as an agent. This paper proposes platoon-level agglomeration (PLA) and lane-level agglomeration (LLA), which differ from no aggregation (NOA) as strategies, to match homogeneous elements and risk aversion among heterogeneous elements. In addition, algorithms related to the agglomeration of CAV agents are also proposed. Simulation experiments, based on the ACMCAV-MAS and cellular automata models, were conducted at different traffic flow densities and different CAV agent penetration rates, with the results showing that the agglomeration strategy achieves the best benefit at a CAV agent penetration rate of 60%. Concomitantly, at a density of 60 veh·km^-1, PLA can increase the traffic flow by 38.14% on average, which is 9.73% higher than that of LLA. Platoon-level agglomeration can also effectively alleviate traffic congestion in the density range of 40-50 veh·km^-1 at a 50%-70% CAV-agent penetration rate. Through a longitudinal risk analysis of medium-and high-density traffic flows, no significant difference was found between the two agglomeration strategies at low CAV agent penetration rates, and the maximum risk reduction ratio reached more than 20%. However, in actual traffic situations, the agglomeration strategy, to some extent, may increase the risk of lateral collisions. In future work, methods to reduce the risk of lateral collisions will continue to be explored. Meanwhile, efforts were expended to solve the deficiency of heterogeneous modeling of artificial driving behavior in the current simulation framework, and the ACMCAV-MAS will be improved to provide a theoretical basis for the formulation of automatic driving strategies in future MTSs.