摘要
城市多模式交通系统是一个高度复杂而多元的交通网络,旨在有效地满足城市内人员、货物和服务的流动需求。多模式交通系统复杂性源于许多因素,包括不同交通模式间的耦合性,交通需求和供应之间复杂的相互作用,以及开放、异质和自适应交通系统的固有随机特性。因此,理解和管理这样一个复杂系统是一个非常复杂且困难的任务。随着交通以及其他领域多源大数据可获取性的增加,计算机硬件算力的增强,以及机器学习模型的飞速发展,大模型的概念被许多领域应用与实践,包括计算机视觉、自然语言处理等。将大模型的概念应用于交通领域,提出了一种根据交通拓扑结构分层“点线面”的多模式交通大模型框架(Multimodal Transportation Generative Pre-trained Transformer,MT-GPT),旨在为复杂多模式交通系统中的多方位决策任务提供数据驱动的大模型。考虑到不同交通模式的特征,探讨了实现这一概念框架的核心技术及其整合方式,构思了适配交通的大模型数据范式与改进的分层多任务学习、分层联邦学习、分层迁移学习与分层Transformer框架。最后,通过搭建“任务岛”与“耦合桥”的框架讨论了这样一个多模式交通大模型框架在“点线面”3层大模型框架下的应用案例,从而为多尺度的多模式交通规划、网络设计、基础设施建设和交通管理提供智能化的支持。
The urban multimodal transportation system is a highly complex and diverse transportation network designed to efficiently meet the mobility needs of people,goods,and services within a city.Its complexity originates from many factors including the coupling between different transportation modes,complex interactions between transportation demand and supply,and intrinsic stochasticity and self-organization of an open,heterogeneous,and adaptive system.Therefore,understanding and managing such a complex system is a nontrivial task.However,with the increasing availability of multisource big data in multimodal transportation and other sectors,enhanced computational hardware capabilities,and rapid development of machine learning models,the concept of large models has been applied in various fields,including computer vision and natural language processing.In this study,a conceptual framework,multimodal transportation generative pretrained transformer(MT-GPT),of a data-driven foundation model for multifaceted decision-making in complex multimodal transportation systems was conceived.Considering the characteristics of different transportation modes,the core technologies and their integration methods were investigated to realize this conceptual framework.An expansive data paradigm is envisioned for a foundation model tailored to transportation,along with improvements in hierarchical multitask learning,hierarchical federated learning,hierarchical transfer learning,and hierarchical transformer framework.Application cases of MT-GPT within the“spots-corridors-networks”three-layer large model framework are discussed by constructing“task islands”and“coupling bridges”.MT-GPT aims to provide an intelligent support for tasks such as multiscale multimodal transportation planning,network design,infrastructure construction,and traffic management.
作者
周臻
顾子渊
曲小波
刘攀
刘志远
ZHOU Zhen;GU Zi-yuan;QU Xiao-bo;LIU Pan;LIU Zhi-yuan(School of Transportation,Southeast University,Nanjing 211189,Jiangsu,China;School of Vehicle and Mobility,Tsinghua University,Beijing 100084,China)
出处
《中国公路学报》
EI
CAS
CSCD
北大核心
2024年第2期253-274,共22页
China Journal of Highway and Transport
基金
国家自然科学基金重点项目(52131203)
国家自然科学基金青年项目(52102375)
江苏省自然科学基金青年项目(BK20210247)
江苏省“双创博士”(JSSCBS20220099)。
关键词
交通工程
多模式交通
大模型
交通管理与决策
TRANSFORMER
多任务学习
联邦学习
迁移学习
traffic engineering
multimodal transportation
foundation model
transportation decision-making task
transformer
multi-task learning
federated learning
transfer learning
