CAREER: Urban Transport Network Design with Privacy-Aware Agent Learning

交通系统的技术进步正在迅速发展，从70年前的高速公路材料和交通控制技术发展到支持“智能城市”的技术：自动驾驶汽车、按需移动服务和机器学习交通控制等。2016年，美国交通部提议在自动驾驶汽车上投入40亿美元，并承诺投入4000万美元应对智能城市这一重大挑战。然而，这些技术在大规模、高度拥挤的城市中的成功运营仍然容易出现运营陷阱和障碍。例如，服务运营商应该如何最好地部署他们的车辆或真实的通知他们的旅行者，以同时优化服务和学习潜力，同时承认他们的隐私？一些引人注目的失败包括赫尔辛基的按需运输服务Kutsuplus和圣地亚哥的Car2Go汽车共享服务。这些系统是高度动态的，但机器学习和动态优化的方法并不是为城市交通网络的独特复杂性而设计的。这个教师早期职业发展（CAREER）计划项目是综合研究和高级教育，为这些动态系统创造新技术，培训学生和专业人士在这些技术，并与私人运营商和孵化器在纽约市参与运作。该项目将使用来自共享出行和自动驾驶汽车系统运营行业合作伙伴的真实的数据，并通过定义混合交通、计算机科学和经济学的新功能角色来推动创新和创业。这将在纽约市的一个试验台上达到高潮，预计该试验台将在未来十年内形成一个关于“智能交通”的下一代国家跨学科研究中心。PI每年夏天将招募当地高中生与他的博士生一起工作;高中生将通过大学的ARISE（科学与工程应用研究创新）计划确定，该计划为妇女，少数民族和低收入背景的学生创造STEM教育经验。 该研究将三种理论结合在一起，以解决智能城市中的这些新的移动性问题：不确定性下的动态资源分配，基于代理的机器学习和网络环境中的隐私优化。所有这三个都是必要的，因为交通系统需要整体优化，但数据通常是从多个旅行者或车辆中获得的。因此，在系统优化中需要代理学习和最小化隐私问题。该方法通过将其与动态网络优化相结合，将隐私控制作为对估计参数的约束，扩展了逆优化的科学。该研究将使所有类型的动态移动系统受益：它将使共享自动驾驶车队运营可行，并使按需服务车队更具可持续性和弹性。交通系统设计背景下的数据隐私和安全性也将得到提高，使旅行者的数据更容易获得。这项研究将有利于智慧城市，传感器部署，人工智能，集体行为，差异隐私，服务系统，公共政策和网络经济。

Technological advances for transportation systems are quickly evolving from their roots in highway materials and traffic control 70 years ago to technologies that support "smart cities": autonomous vehicles, on-demand mobility services, and traffic control with machine learning, among others. In 2016, the U.S. Department of Transportation proposed spending $4 billion on autonomous vehicles, and pledged $40 million in tackling smart cities as a grand challenge. However, successful operation of these technologies in a large scale, highly congested city remains prone to operational pitfalls and obstacles. For example, how should a service operator best deploy their vehicles or inform their travelers in real time to optimize service and learning potential simultaneously, while acknowledging their privacy? Some high profile failures include the on-demand transit service Kutsuplus in Helsinki and the Car2Go car share service in San Diego. These systems are highly dynamic, but methods in machine learning and dynamic optimization are not designed for the unique intricacies of urban transport networks. This Faculty Early Career Development (CAREER) Program project is for integrated research and advanced education to create new technologies for these dynamic systems, train students and professionals in these technologies, and engage with private operators and incubators in New York City to operationalize them. The project will use real data from industry partners in ridesharing and autonomous vehicle systems operations, and drive innovation and entrepreneurship by defining new functional roles that mix transportation, computer science, and economics. This will culminate in a test bed in New York City that is expected to shape a next-generation national interdisciplinary research center on "smart transit" over the next decade. The PI will recruit local high school students to work with his PhD students each summer; the high school students will be identified through the university's ARISE (Applied Research Innovations in Science and Engineering) program which creates STEM education experiences for women, minorities, and students from low-income backgrounds. The research marries three theories together in order to address these new mobility problems in smart cities: dynamic resource allocation under uncertainty, agent-based machine learning, and privacy optimization in a network context. All three are necessary because transportation systems need to be optimized holistically, but data is typically obtained from multiple travelers or vehicles in operation. As such, agent learning and minimization of privacy concerns in the system optimization is needed. The methodology expands the science of inverse optimization by integrating it with dynamic network optimization with privacy control as constraints on the estimated parameters. The research benefits all types of dynamic mobility systems: it will make shared autonomous vehicle fleet operations viable and on-demand service fleets more sustainable and resilient. Data privacy and security in a transport system design context will also be advanced, allowing data from travelers to be more accessible. This research will benefit smart cities, sensor deployment, artificial intelligence, collective behavior, differential privacy, service systems, public policy, and network economics.