演讲标题:AI and Mathematical Optimization: New progress in solving huge-scale mathematical programs and applications
摘要:In recent years, with the rapid increase in the scale of data from real-world problems and AI training tasks, the corresponding decision-making and computational problems have also become larger and larger. The scale of some linear programs (in terms of non-zero elements) has reached the tens of billions, while the matrix size of some semi-define programs has reached hundreds of millions. These problems often also require high-precision solutions. In this presentation, we discuss some recent frontier explorations in response to these challenges, particularly advancements in first-order methods, interior-point algorithms, and fast solutions for large-scale problems under the GPU/CUDA architecture. In addition, we describe how to integrate the AI Large Language Models and OR Optimization Solvers and tackle some real-world applications.
简介:曾任芬兰科学院与芬兰国家创新局“芬兰杰出教授”、德国联邦教育与研究部“洪堡人工智能教席教授”。分别于1988年、1991年、1996年获浙江大学工学学士、工学硕士、和工学博士学位,并于2001年获德国波鸿鲁尔大学工学博士学位。1991年至1997年在浙江大学电机系任助教、讲师和副教授,1998年至1999年在美国新泽西州立大学工业工程系从事博士后研究, 2001年至2010年在德国本田(欧洲)研究院任资深科学家、主任科学家, 2010年加入英国萨里大学计算机系任计算智能讲席教授,2019年升任“萨里杰出教授”, 2021至2023年担任德国比勒菲尔德大学工学院洪堡人工智能教席教授。曾任《IEEE认知与发育系统汇刊》主编,现任IEEE计算智能学会主席。2023年10月全职加入西湖大学工学院,受聘人工智能讲席教授,并创立“可信与通用人工智能实验室”。演讲标题:Graph Neural Networks for Combinatorial Optimization
摘要:Graph neural networks have been found successful in solving combinatorial optimization problems. This talk starts with a simple example of solving the travelling salesman problem using graph neural networks. Then, we present an approach to multi-objective facility location using two graph neural networks with supervised training. Finally, we showcase how a graph neural network with negative message passing can be trained using unsupervised training for solving graph coloring problems. We conclude the talk with a summary and discussion of future work.
叶杰平
阿里云智能集团副总裁
简介:美国明尼苏达大学博士,IEEE Fellow,曾任美国密西根大学教授。曾担任多个国际顶级期刊编委及国际顶级会议程序委员会主席和领域主席,发表高质量学术论文近400篇 (H-index: 103)。曾先后荣获美国国家自然科学基金会生涯奖、CCF科学技术奖科技进步一等奖、国际运筹学领域顶级实践奖-瓦格纳运筹学杰出实践奖、KDD China技术转化奖、以及多个国际顶级会议最佳论文奖等。演讲标题:Large Language Models: An In-Depth Mechanism Analysis and Applications
摘要:Large Language Models (LLMs) have emerged as pivotal tools in natural language processing, transforming the way machines understand, generate, and interact with human language. This presentation aims to explore the inner workings of LLMs, offering a comprehensive analysis of their underlying mechanisms, and showcasing their applications in multiple tasks.
研究成果发表于ANOR、COR、EJOR、MSOM、OR Spectrum、TRB、TRC、TRE、TS等运筹学和交通科学领域学术期刊。曾任美国运筹学与管理学研究协会(INFORMS)交通科学与物流分会(TSL)秘书长和INFORMS Journal on Computing(IJOC)期刊副编(AE)。现任Transportation Science(TS)期刊副编和Transportation Research Part E(TRE)期刊编委。演讲标题:序贯决策问题与仿真优化:在交通与物流中的应用摘要:基于仿真优化方法和序贯决策问题的基本特征,探讨仿真优化方法与马尔可夫决策过程模型在解决序贯决策问题时的结合。通过若干交通与物流应用领域的研究课题,分享其数学建模及策略设计和训练过程的研究与思考。
演讲标题:Thoughts on the Design and Implementation of Column Generation Method for Crew Pairing Optimization
摘要:In a forum intended for OR professionals, the topics are mostly on math models, algorithms and their applications. Through the example of crew pairing optimization,the purpose of this talk is to shed light on the design and implementation of optimization systems, or specifically the column generation method with a rule engine to encapsulate the business logics, so that the system can be easily set up to solve problems of great variation in business rules, and therefore serve for business purpose beyond just finding an optimization solution.袁晓明
香港大学数学系教授
简介:研究方向为优化算法与理论、云计算、最优控制、人工智能。Clarivate Analytics 高被引学者。带领香港大学与华为云的研究队伍进入2023年INFORMS Franz Edelman Award决赛。与华为云合作项目被评为2023年华为公司“公司级优秀合作项目奖”。
演讲标题:Revisiting First-order Algorithms for Optimization Problems in Industry
摘要:We will revisit some classic first-order algorithms for optimization problems arising in the industries of AI and Cloud Computing, including post-training pruning for large language models, bandwidth allocation for live streaming, and digital human simulation. In particular, we will show their adaptability to GPUs and other parallel architectures for real industrial applications. We will also showcase how to save computation substantially for classic first-order algorithms by mathematically-driven heuristics. 高磊
简介:工学博士,副高,中电海康创新研究院研究员。曾获武汉市科技创新人才(3551)、粤港澳大湾区创新青年。主要从事多传感器数据融合、无人系统融合感知与定位、多智能体仿真的研究,主要理论方法:基于多传感器数据的融合和特征分析,包括SLAM建图、隐马尔可夫链模型、卷积神经网络等。围绕以上研究,主持武汉市重点基金项目1项,参与浙江省基金重点项目1项(排名8/10) 。近年来,在 IJGI、IEEE Transactions on Transport、SCIENCE CHINA Information Sciences、《图形图像学》等重要学术期刊上发表学术论文9篇,出版学术专著1部,授权发明专利5项,获得软件著作权2项,广州市技进步奖三等奖1项。