CPS: Medium: Emulating Emerging Autonomous Vehicle Technologies to Understand Their Impact on Urban Congestion

自动驾驶汽车将继续存在，这种新兴的自动驾驶汽车（AV）技术将改变我们的交通系统。自动驾驶技术的潜在好处包括提高安全性，增加更多车辆在道路上行驶的能力（通过排成一排的车辆彼此之间的距离非常近）。但是自动驾驶技术在未来将如何发展是高度不确定的，我们对它们对交通系统影响的理解也是如此。例如，文献中的AV模型都没有得到经验数据的验证，这使得对其影响的现有预测非常值得怀疑。最近对一组特斯拉汽车的研究表明，交通拥堵实际上可能会加剧。为了解决这个问题，该项目将使用商用自动驾驶汽车进行测量，并提出复制其行为的数学模型。这些模型将使我们更好地了解自动驾驶汽车在相互排成一排时的行为，并提出解决拥堵等不良后果的方法。该项目的教育部分将使本科生和研究生接触到一个蓬勃发展的生态系统，在这个生态系统中，汽车制造商、技术公司和应用程序开发人员通过开源软件、学习材料和数据集来培养自动驾驶技术所需的机器学习模型，从而促进创新。该项目的研究目标是开发一个分析和数值框架，以模拟当前自动驾驶技术在不久的将来对交通网络的影响。该研究方法将基于收集目前市场上2/3级自动驾驶汽车的大量经验数据，以训练行业正在实施的机器学习模型类型，包括深度神经网络和专家领域知识的结合。鉴于最近的经验证据表明，这些车辆可能比人类驾驶员表现出更多的弦不稳定性，该项目将确定稳定性约束，可以在训练过程中纳入，以避免不稳定。并建立相应的跟车模型，建立网络层面的宏观动力学。该项目将重点关注纵向加速/减速组件，因为它在管柱稳定性、网络容量和拥塞方面起着重要作用。它还可以用一般场景所需的一小部分数据来训练机器学习模型，理解这种简化的驾驶场景是成功分析更一般案例的第一步。该项目将建立机器学习模型和汽车跟随模型之间的联系，并将引导未来自动驾驶技术的研究和开发转向保证稳定的人工智能模型，预计其影响将是重大的。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Self-driving cars are here to stay, and this emerging automated vehicle (AV) technology will transform our transportation system. Potential benefits of AV technology include improved safety and greater capacity for more vehicles to travel on the road (by forming a platoon of vehicles with very close distance with each other). But how AV technologies will evolve in the future is highly uncertain, and so is our understanding of their impacts on our transportation system. For example, none of the AV models in the literature have been validated with empirical data, which makes existing predictions about their impacts highly questionable. Recent studies on a platoon of Tesla vehicles suggest that traffic congestion might actually increase. To address this problem, this project will conduct measurements using commercially available AV vehicles and come up with mathematical models that replicate their behavior. These models will allow us to better understand how AV vehicles behave when they form a platoon with each other and come up with methods to address undesirable consequences such as congestion. The educational component of this project will expose both undergrad and graduate students to a thriving ecosystem where car manufacturers, technology companies and application developers foster innovation via open source software, learning material and data sets to train the machine learning models needed for AV technologies.The research objective of this project is to develop an analytical and numerical framework to emulate the impacts that current AV technologies will have on the transportation networks of the near future. The research approach will be based on the collection of large amounts of empirical data from Level 2/3 AVs currently on the market to train the type of machine learning models that the industry is implementing, consisting of a combination of deep neural networks and expert domain knowledge. Given the recent empirical evidence revealing that these vehicles may exhibit more string instability than human drivers, the project will identify stability constraints that can be incorporated during training to avoid instability. Additionally, the corresponding car-following models that will establish macroscopic dynamics at the network level will be formulated. The project will focus on the longitudinal acceleration/deceleration component since it plays the major role in string stability, network capacity and congestion. It also makes it possible to train machine learning models with a fraction of the data needed for general scenarios, and understanding this simplified driving scenario is the first step towards a successful analysis of more general cases. The impact of this project is expected to be significant as it will establish the connection between machine learning models and car-following models, and will steer research and development of future AV technologies towards artificial intelligence models that are guaranteed to be stable.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.