S&AS:FND:COLLAB:Unsupervised Rare Event Learning - With Applications on Autonomous Vehicles

S

• 批准号: 1849280
• 负责人: Henry Lam
• 金额: $ 25.6万
• 依托单位: Columbia University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2019
• 资助国家: 美国
• 起止时间: 2019-02-15 至 2023-01-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1849280&HistoricalAwards=false
• 关键词: S& FND; COLLAB; Unsupervised; Rare

Intelligent physical systems (IPSs) architecture must be cognizant, taskable, adaptive, and ethical, particularly when users need to reliably address uncertain, dynamically changing situations. Despite the collection and availability of massive data on IPSs, a significant challenge, faced by the autonomous vehicle (AV) industry is the capability to efficiently anticipate and avoid rare but catastrophic failure events. Developing an information structure that enables users to make the best use of massive data and to model the complexity and rarity of critical situations would contribute to addressing this challenge. This project investigates a novel, unsupervised rare-event learning framework for AV research that utilizes the collected data to assess and alleviate the risks, limitations, and failure modes of IPSs. The framework combines the flexibility in modeling high-dimensional time-series driving data via unsupervised learning algorithms, with the statistical robustness and computational efficiency from rare-event analysis. The specific goals of the research include: 1) formulate an effective framework to extract safety-critical information from driving database; 2) develop a solid probabilistic measurement scheme for rare but adverse events in AV driving contexts; 3) disseminate the framework and research outcomes to industry and government units in urgent need of reliable evaluation and development methods; and 4) cross-fertilize research thrusts between the academic communities of rare-event estimation and the study of IPSs. The research will merge high-fidelity driving representations learned from data and statistically rigorous rare event analysis so that AV design evaluation in simulated complex driving situations will become highly efficient. It will synthesize machine learning with simulation methodologies, including rare-event analysis, stochastic optimization, and particle methods to boost the performances of state-of-the-art algorithms. The project will foster transformative academic cross-fertilization between researchers in rare-event estimation (operations research) and IPSs (machine learning/robotics), by disseminating the outcomes in top-tier journals, conferences, tutorials and invited seminars. The research will train under-represented minority students at the PIs' institutions, provide opportunities to meet key AV academic and industry players, and show them how to apply mathematical and engineering skills to improve transportation mobility, safety, and roadway efficiency.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.

智能物理系统（IPS）架构必须是认知的，可分配的，自适应的和道德的，特别是当用户需要可靠地解决不确定的，动态变化的情况。尽管收集了大量IPS数据，但自动驾驶汽车（AV）行业面临的一个重大挑战是有效预测和避免罕见但灾难性故障事件的能力。发展一种信息结构，使用户能够最佳地利用大量数据，并对危急情况的复杂性和罕见性进行建模，将有助于应对这一挑战。该项目研究了一种用于AV研究的新型无监督罕见事件学习框架，该框架利用收集的数据来评估和减轻IPS的风险，局限性和故障模式。该框架结合了通过无监督学习算法对高维时间序列驾驶数据建模的灵活性，以及稀有事件分析的统计鲁棒性和计算效率。研究的具体目标包括：1）建立一个有效的框架，从驾驶数据库中提取安全关键信息; 2）为自动驾驶环境中罕见但不利的事件开发可靠的概率测量方案; 3）将框架和研究成果传播给迫切需要可靠评估和开发方法的行业和政府单位;（4）稀有事件估计学术界与IPS研究之间的交叉研究。该研究将融合从数据中学习到的高保真驾驶表现和统计上严格的罕见事件分析，以便在模拟复杂驾驶情况下进行高效的AV设计评估。它将机器学习与仿真方法相结合，包括稀有事件分析，随机优化和粒子方法，以提高最先进算法的性能。该项目将通过在顶级期刊、会议、教程和特邀研讨会上传播成果，促进罕见事件估计（运筹学）和IPS（机器学习/机器人技术）研究人员之间的变革性学术交叉。该研究将在PI的机构中培训代表性不足的少数民族学生，提供与关键AV学术和行业参与者会面的机会，并向他们展示如何应用数学和工程技能来提高交通流动性，安全性和道路效率。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

Certifiable Evaluation for Autonomous Vehicle Perception Systems using Deep Importance Sampling (Deep IS)

• DOI: 10.1109/itsc55140.2022.9922202
• 发表时间: 2022-10
• 期刊: 2022 IEEE 25th International Conference on Intelligent Transportation Systems (ITSC)
• 作者: Mansur Arief;Zhepeng Cen;Zhen-Yan Liu;Zhiyuan Huang;Bo Li;H. Lam;Ding Zhao

On the impacts of tail model uncertainty in rare-event estimation

尾部模型不确定性对罕见事件估计的影响

• 发表时间: 2019
• 期刊: Proceedings of the Winter Simulation Conference
• 作者: Huang, Z.;Lam, H.
• 通讯作者: Lam, H.

Scalable Safety-Critical Policy Evaluation with Accelerated Rare Event Sampling

通过加速罕见事件采样进行可扩展的安全关键政策评估

• DOI: 10.1109/iros47612.2022.9981867
• 发表时间: 2022
• 期刊: 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS
• 作者: Xu, Mengdi;Huang, Peide;Li, Fengpei;Zhu, Jiacheng;Qi, Xuewei;Oguchi, Kentaro;Huang, Zhiyuan;Lam, Henry;Zhao, Ding
• 通讯作者: Zhao, Ding

Evaluation Uncertainty in Data-Driven Self-Driving Testing

数据驱动自动驾驶测试中的评估不确定性

• DOI: 10.1109/itsc.2019.8917406
• 发表时间: 2019
• 期刊: 2019 IEEE Intelligent Transportation Systems Conference (ITSC
• 作者: Huang, Zhiyuan;Arief, Mansur;Lam, Henry;Zhao, Ding
• 通讯作者: Zhao, Ding

Rare-event Simulation for Neural Network and Random Forest Predictors

神经网络和随机森林预测器的罕见事件模拟

• DOI: 10.1145/3519385
• 发表时间: 2022
• 期刊: ACM Transactions on Modeling and Computer Simulation
• 影响因子: 0.9
• 作者: Bai, Yuanlu;Huang, Zhiyuan;Lam, Henry;Zhao, Ding
• 通讯作者: Zhao, Ding