CAREER: Spatial-Temporal Imitation Learning

职业：时空模仿学习

• 批准号: 1942680
• 负责人: Yanhua Li
• 金额: $ 52.95万
• 依托单位: Worcester Polytechnic Institute
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2020
• 资助国家: 美国
• 起止时间: 2020-07-01 至 2025-06-30
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1942680&HistoricalAwards=false
• 关键词: CAREER; Spatial; Temporal; Imitation; Learning

Humans make daily decisions based on their own "strategies" (such as taxi drivers' passenger-seeking processes and commuters' transit mode choices). Understanding and incorporating human decision-making strategies will bring significant benefits to the growing gig-worker population and transportation marketplace. For example, learning the decision-making strategies from taxi drivers, personal vehicle drivers, and urban commuters can facilitate the service providers (e.g., taxi/ride-hailing companies) to better serve the passengers, and enable the urban planners to design better road networks and transit routes to meet the needs of urban travelers. The goal of this project is to develop, implement, and evaluate a unified framework to learn the decision-making strategies of human agents from their generated mobility data, with applications to explain and incentivize their decisions to promote individual and societal well-being. Moreover, in this project, the investigator will integrate research, education, and outreach by developing new courses for both undergraduate and graduate students, reaching out to K-12 students, and engaging women and underrepresented minorities.There are several technical challenges to learn human decision-making strategies from their mobility data: Human decision-making strategies may vary over time and space, i.e., spatial-temporal dynamics challenge. The mobility data collected may cover only a part of the spatial regions and time periods, i.e., spatial-temporal sparsity challenge. Most human agents are not "experts", thus the generated mobility data are noisy and uncertain, i.e., non-expert challenge. A large number of human agents interact with each other when making decisions, i.e., interaction and scalability challenge. Human decisions are governed by many (sometimes hidden) factors, which make it hard to infer explainable information from their decision-making strategies, i.e., explainability challenge. Human agents have diverse reactions to offered incentives, making it hard to design targeted incentive mechanisms to consider both agents' inherent decision-making strategies and their online feedback, i.e., incentive design challenge. This project will address these research challenges, and include the development of a novel spatial-temporal imitation learning framework for learning decision-making strategies from individual and interactive human agents, and an interactive system that provides human agents with explainable learning results and online incentives to promote their decision-making strategies. The spatial-temporal imitation learning framework and associated algorithms have the potential to be transformative in both the data and urban sciences by enabling efficient and accurate discovery of human decision-making strategies from their mobility data.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.

人类根据自己的“策略”(例如出租车司机的寻客过程和通勤者的交通方式选择)做出日常决策。理解并纳入人类决策策略将为不断增长的零工人口和运输市场带来重大好处。例如，从出租车司机、私人车辆司机和城市通勤者那里学习决策策略可以帮助服务提供商(例如出租车/叫车公司)更好地服务于乘客，并使城市规划者能够设计更好的道路网络和交通路线来满足城市出行者的需求。该项目的目标是开发、实施和评估一个统一的框架，以从人类代理生成的流动性数据中学习他们的决策策略，并使用应用程序来解释和激励他们促进个人和社会福祉的决策。此外，在这个项目中，研究人员将通过为本科生和研究生开发新课程，接触到K-12学生，并让妇女和代表性不足的少数民族参与进来，将研究、教育和推广整合在一起。从他们的流动性数据中学习人类决策策略存在几个技术挑战：人类决策策略可能会随时间和空间而变化，即时空动态挑战。所收集的移动性数据可以仅覆盖空间区域和时间周期的一部分，即时空稀疏性挑战。大多数人类智能体不是专家，因此生成的移动性数据具有噪声和不确定性，即非专家挑战。大量的人类智能体在决策时相互作用，即交互和可伸缩性挑战。人类的决策受到许多(有时是隐藏的)因素的支配，这使得人们很难从他们的决策策略中推断出可解释的信息，即可解释性挑战。人类智能体对提供的激励有不同的反应，这使得很难设计有针对性的激励机制来同时考虑智能体固有的决策策略和他们的在线反馈，即激励设计挑战。该项目将解决这些研究挑战，包括开发一个新的时空模拟学习框架，用于从个人和互动的人类代理人那里学习决策策略，以及一个互动系统，向人类代理人提供可解释的学习结果和在线激励措施，以促进其决策战略。时空模仿学习框架和相关算法能够从数据和城市科学的流动性数据中高效而准确地发现人类决策策略，从而在数据和城市科学中都具有变革的潜力。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势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

f-GAIL: Learning f-Divergence for Generative Adversarial Imitation Learning

• 发表时间: 2020-10
• 期刊: ArXiv
• 作者: Xin Zhang;Yanhua Li;Ziming Zhang;Zhi-Li Zhang

HintNet: Hierarchical Knowledge Transfer Networks for Traffic Accident Forecasting on Heterogeneous Spatio-Temporal Data

HintNet：异构时空数据交通事故预测的分层知识转移网络

• 发表时间: 2022
• 期刊: 2022 SIAM International Conference on Data Mining (SDM
• 作者: An, Bang;Vahedian, Amin;Zhou, Xun;Street, Nick W.;Li, Yanhua
• 通讯作者: Li, Yanhua

EgoSpeed-net: forecasting speed-control in driver behavior from egocentric video data

• DOI: 10.1145/3557915.3560946
• 发表时间: 2022-09
• 期刊: Proceedings of the 30th International Conference on Advances in Geographic Information Systems
• 作者: Yichen Ding;Ziming Zhang;Yanhua Li;Xun Zhou

MTrajRec: Map-Constrained Trajectory Recovery via Seq2Seq Multi-task Learning

• DOI: 10.1145/3447548.3467238
• 发表时间: 2021-08
• 期刊: Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining
• 作者: Hu Ren;Sijie Ruan;Yanhua Li;Jie Bao;Chuishi Meng;Ruiyuan Li;Yu Zheng

xGAIL: Explainable Generative Adversarial Imitation Learning for Explainable Human Decision Analysis

• DOI: 10.1145/3394486.3403186
• 发表时间: 2020-07
• 期刊: Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining
• 作者: Menghai Pan;Weixiao Huang;Yanhua Li;Xun Zhou;Jun Luo