EAGER: Real-Time: Learning-Mediated Control for Traffic Shaping

EAGER：实时：以学习为中介的流量整形控制

• 批准号: 1839816
• 负责人: Nicholas Duffield
• 金额: $ 30万
• 依托单位: Texas A&M Engineering Experiment Station
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2018
• 资助国家: 美国
• 起止时间: 2018-10-01 至 2022-09-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1839816&HistoricalAwards=false
• 关键词: EAGER; Real; Time; Learning; Mediated

Efficient Management of Vehicular Traffic via Real-time Machine-Learning-Mediated Control and Traffic ShapingWhile connectivity and automation promise orders of magnitude gains in the safety and efficiency of vehicular transportation networks, these gains cannot be realized without monitoring, learning the behavior, and control of vehicles at different aggregation levels. Indeed, current congestion mitigation methods, such as speed harmonization that uses a sequence of variable speed limits along a highway do not reliably control congestion, and may exacerbate it (e.g., via shocks propagated through speed limit changes) due to the inconsistency between congestion prediction and real-time control. The objective of this project is to develop a holistic approach using machine-learning methods to identify and predict macroscopic congestion behavior of traffic based on both vehicle-borne and transportation infrastructure measurements, while designing fine-grain control systems for individual vehicles that can help to mitigate congestion effects. In doing so, the project recognizes that these designs must account for the possibility of low take-up rates of connected, automated vehicles (CAVs) over the next decade, and the consequent dominance of human-mediated vehicle operation for some time to come. The project also includes the development of educational materials on data analytics and vehicular control systems. Intrinsic to the program are efforts at outreach to involve high-school students via demonstrations and lectures based on the technology developed.The goal of this project is to develop the theory of and evaluate a novel approach to traffic management entitled "real-time learning-mediated control". The key idea is to meld large-scale real-time learning about macroscopic phenomena in a physically interpretable manner, with distributed dynamic control of individual vehicles in a provably safe and efficient manner. The work comprises two thrusts, namely (i) Traffic State Prediction, which offers a Graph Signal Processing (GSP)-based congestion prediction approach for planned and unplanned congestion-causing events, and (ii) Traffic Shaping and Control, which offers novel vehicular control methods that shape traffic in a stable manner over the multiple dimensions of target time headway and velocities over space and time, and candidate time-gap and velocity profiles in a mixed environment of both connected, automated vehicles and human driven ones. Thus, the overall aim is to combine the ability of learning methods to provide predictions about complex interconnected systems, with control laws that are safe and consistent with the laws of physics. The value of this research to broader society is in combining traffic prediction, control and learning, which can result in accurate congestion mitigation and increased throughput. Incorporating analytical concepts into senior design projects and courses enhances the project via educational impact. The project also contributes to development of systems-design expertise for students, as well as to diversity enhancement through minority student engagement.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

通过实时机器学习介导的控制和交通整形有效管理车辆交通虽然连接和自动化承诺在车辆运输网络的安全性和效率方面获得数量级的收益，但如果不监控，学习行为和控制不同聚合级别的车辆，这些收益就无法实现。实际上，当前的拥塞缓解方法（诸如使用沿高速公路沿着的一系列可变速度限制的速度协调）不能可靠地控制拥塞，并且可能加剧拥塞（例如，经由通过速度限制变化传播的冲击），这是由于拥塞预测和实时控制之间的不一致。该项目的目标是开发一种整体方法，使用机器学习方法来识别和预测基于车载和交通基础设施测量的宏观交通拥堵行为，同时为单个车辆设计细粒度控制系统，以帮助减轻拥堵影响。 在这样做的过程中，该项目认识到，这些设计必须考虑到未来十年互联自动驾驶汽车（CAV）的低使用率的可能性，以及未来一段时间以人为媒介的车辆操作的主导地位。 该项目还包括编写关于数据分析和车辆控制系统的教材。该项目的核心是努力通过基于开发的技术的演示和讲座来吸引高中生参与。该项目的目标是开发一种名为“实时学习介导控制”的交通管理新方法的理论并对其进行评估。 其关键思想是以物理可解释的方式融合对宏观现象的大规模实时学习，以可证明安全和有效的方式对单个车辆进行分布式动态控制。这项工作包括两个方面，即（i）交通状态预测，它为计划内和计划外的交通拥堵事件提供了一种基于图形信号处理（GSP）的拥堵预测方法，以及（ii）交通整形和控制，它提供了新的车辆控制方法，在目标时间间隔和空间和时间速度的多个维度上以稳定的方式整形交通，以及在连接的自动车辆和人类驾驶车辆的混合环境中的候选时间间隔和速度曲线。 因此，总体目标是将学习方法提供关于复杂互连系统的预测的能力与安全且符合物理定律的控制律联合收割机相结合。 这项研究对更广泛的社会的价值在于将交通预测、控制和学习结合起来，这可以准确地缓解拥堵并增加吞吐量。 将分析概念融入高级设计项目和课程，通过教育影响增强了项目。该项目还有助于学生系统设计专业知识的发展，以及通过少数民族学生的参与来增强多样性。该奖项反映了NSF的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

项目成果

Variational Graph Recurrent Neural Networks

• 发表时间: 2019-08
• 作者: Ehsan Hajiramezanali;Arman Hasanzadeh;N. Duffield;K. Narayanan;Mingyuan Zhou;Xiaoning Qian

Piecewise Stationary Modeling of Random Processes Over Graphs With an Application to Traffic Prediction

• DOI: 10.1109/bigdata47090.2019.9005965
• 发表时间: 2017-11
• 期刊: 2019 IEEE International Conference on Big Data (Big Data)
• 作者: Arman Hasanzadeh;Xi Liu;N. Duffield;K. Narayanan

Learning with Safety Constraints: Sample Complexity of Reinforcement Learning for Constrained MDPs

• DOI: 10.1609/aaai.v35i9.16937
• 发表时间: 2020-08
• 作者: Aria HasanzadeZonuzy;D. Kalathil;S. Shakkottai

Adaptive Shrinkage Estimation for Streaming Graphs

流图的自适应收缩估计

• 发表时间: 2020
• 期刊: Advances in neural information processing systems
• 作者: Ahmed, Nesreen;Duffield, Nick
• 通讯作者: Duffield, Nick

Semi-Implicit Stochastic Recurrent Neural Networks

• DOI: 10.1109/icassp40776.2020.9053491
• 发表时间: 2019-10
• 期刊: ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)
• 作者: Ehsan Hajiramezanali;Arman Hasanzadeh;N. Duffield;K. Narayanan;Mingyuan Zhou;Xiaoning Qian