基于多目标协同演化算法的大规模自动驾驶策略  

Optimal Driving Policies for Large-scale Autonomous Vehicles Based on Multi-objective Co-evolutionary Algorithms

作　　者：刘章杰 李慧云[1,2,3] LIU Zhangjie;LI Huiyun(Shenzhen Institutes of Advanced Technology,Chinese Academy of Sciences,Shenzhen 518055,China;CAS Key Laboratory of Human-Machine Intelligence-Synergy Systems,Shenzhen Institutes of Advanced Technology,Shenzhen 518055,China;Guangdong-Hong Kong-Macao Joint Laboratory of Human-Machine Intelligence-Synergy Systems,Shenzhen 518055,China)

机构地区：[1]中国科学院深圳先进技术研究院,深圳518055 [2]中国科学院人机智能协同系统重点实验室,深圳518055 [3]粤港澳人机智能协同系统联合实验室,深圳518055

出　　处：《集成技术》2020年第5期93-102,共10页Journal of Integration Technology

基　　金：深圳市无人驾驶感知决策与执行技术工程实验室项目(Y7D004);深圳电动汽车动力平台与安全技术重点实验室项目。

摘　　要：目前无人驾驶技术领域的研究重点主要集中在单车层面的感知、决策与控制,而缺少对多车之间交互及博弈的研究,因此无法有效降低交通系统整体事故率并提升通行效率。该文提出一种基于合作博弈理论的大规模自动驾驶策略涌现方法。通过建立面向网联汽车、多目标优化决策的合作博弈演化平台,并构造了一种网格道路模型和车辆运动学模型,使得系统中各车辆之间以近邻博弈的方式进行交互;同时系统采用分布式算法并具有间接交互的特点,最终模型计算复杂度与模拟车辆规模呈线性关系。实验结果表明,最佳策略涌现后,事故率和平均速度均取得明显改善,其中事故率降低了90%,模型计算速度提升了30%。该方法可应用于包含数百万辆自动驾驶汽车的城市级智能交通规划系统中。Research in current autonomous driving domain mainly focused on the problems of perception, decision-making and control based on single autonomous vehicle, but the interactions and games among different vehicles are usually ignored. That makes exiting techniques inapplicable to reduce the accident rate and to improve the traffic efficiency of the transportation system. To solve this problem, a decision-making emergence method is proposed for the large-scale autonomous driving system based on the principle of coevolutionary games. We have established a grid road model and a vehicle kinematics model in which each vehicle interacts by indirect interaction. Benefited from the distributed algorithms and the communication method between vehicles, the computational complexity can be kept linear with the simulated vehicle volume. By designing a multi-objectives reward function, and making the co-evolution process in a simulated environment, the emergence of dominant driving strategies can be observed efficiently. Experimental results showed that the accidents rate and the average computation speed can be greatly improved compared with conventional approach. In details, the accident rate can be reduced by 90% and the average speed can be increased by 30%. The proposed method have great potentials to explore the optimal driving strategy for urban traffic up to millions of autonomous vehicles.

关 键 词：自动驾驶策略 多目标 协同演化 博弈 

分 类 号：V323.19[航空宇航科学与技术—人机与环境工程]