基于强化学习的自动驾驶伦理决策机制研究  

Research on autonomous driving ethical decision mechanism based on reinforcement learning

作　　者：刘国满[1] 盛敬[1] 周志龙 司永章 LIU Guoman;SHENG Jing;ZHOU Zhilong;SI Yongzhang(School of Mechanical Engineering,Nanchang Institute of Technology,Nanchang 330099,China)

机构地区：[1]南昌工程学院机械工程学院,南昌330099

出　　处：《南昌工程学院学报》2023年第1期74-79,共6页Journal of Nanchang Institute of Technology

基　　金：江西省教育厅自然科学基金项目(GJJ201911);大学生创新创业训练项目(S202211319014)。

摘　　要：由于自动驾驶技术的道德伦理和法律法规研究相对滞后,自动驾驶车辆遇到伦理困境时,还不能做出决策,影响了人类对自动驾驶车辆的信任和认可,阻碍了自动驾驶技术的应用和发展。根据Moral Machine伦理困境下调查结果,选择6个特征作为伦理决策的特征因素,初步建立基于差异化特征的车道危险度预测模型,对车道危险度进行预测和比较,做出一个决策行为。再根据人类决策调查结果,判断自动驾驶车辆所做的决策行为是否符合人类决策,若是,则进行激励和相应特征权重系数的强化,优化车道危险度预测模型,建立自动驾驶车辆伦理决策机制。最后运用测试伦理困境分别对初始和建立自动驾驶车辆决策机制进行测试,测试结果表明两者与人类决策一致率分别为71.4%和85.7%,说明训练后自动驾驶车辆伦理决策机制的决策行为更能符合人类决策标准,为自动驾驶伦理决策提供了一种新的研究思路和方法。Because of the relatively backward research on ethics,laws and regulations of autonomous driving technology,autonomous vehicles cannot make decisions when encountering ethical dilemmas,which affects human trust and recognition of autonomous vehicles and hinders the application and development of autonomous driving technology.According to the investigation results in Moral Machine,6 features are selected as factors of ethical decision.Then a prediction model of lane risk based on differentiated features is preliminarily established to predict and compare lane risk and make a decision behavior.Then it is judged whether the decision behavior of autonomous vehicles conforms to human decision.If so,the corresponding feature weight coefficients are strengthened to optimize the prediction model of lane risk and establish an autonomous driving ethical decision mechanism.Lastly,the initial and established autonomous driving ethical decision mechanisms are tested and verified by using ethical dilemmas respectively.The test results show the rates of both consistent with human decision are 71.4%and 85.7%respectively,which indicates that the trained autonomous driving ethical decision mechanism is more suitable for human decision standard and provides a new research ideas and method for autonomous driving ethical decision in the future.

关 键 词：自动驾驶 控制决策 强化学习 伦理困境 

分 类 号：U495[交通运输工程—交通运输规划与管理]