Now more than ever the world calls for our leadership, our collaboration and our innovation. That is why we’ve made the decision to host the HHL 2020 Conference online.
As we approach the conference date, we encourage you to Register and stay up to date with new developments through our website and email communications.
]]>Now more than ever the world calls for our leadership, our collaboration and our innovation. That is why we’ve made the decision to host the HHL2020 Conference online.
]]>
Pol Boada-Collado is a 5th year Ph.D. candidate in the Department of Industrial Engineering and Management Sciences at Northwestern University, working with Karen Smilowitz and Sunil Chopra. Before his Ph.D., Pol worked as a research assistant at IESE Business School and as a data scientist in the Neuroscience Department of IDIBAPS (Barcelona). During his first summer at Northwestern, Pol did an internship as a supply chain consultant, an industry collaboration that has motivated his doctoral dissertation. His research studies how traditional supply chain models can become more nimble to adapt to market changes such as demand surges or customer behavior trends, with applications to retail and home delivery operations. His research has been published in Transportation Science and M&SOM and has been awarded the Nemhauser Prize for Best Paper. Pol received a Transportation Center Dissertation Year Fellowship and a McCormick Terminal Year Fellowship. Pol served as President and Treasurer of the Northwestern University INFORMS student chapter and received the IEMS Department Award for Exceptional Leadership and Service in recognition of these efforts.
]]>Abstract:
Getting students to the right school at the right time can pose a challenge for school districts in the United States, which must balance educational objectives with operational ones, often on a shoestring budget. Examples of such operational challenges include deciding which students should attend, how they should travel to school, and what time classes should start. From an optimizer’s perspective, these decision problems are difficult to solve in isolation, and present a formidable challenge to solve together. In this paper, we develop an optimization-based approach to three key problems in school operations: school assignment, school bus routing, and school start time selection. Our methodology is comprehensive, flexible enough to accommodate a variety of problem specifics, and relies on a tractable decomposition approach. In particular, it comprises a new algorithm for jointly scheduling school buses and selecting school start times, that leverages the simplifying assumption of fixed route arrival times to allow for multiple objectives and enhance tractability. We show that our methodology can significantly streamline the operations of Boston Public Schools Extended School Year (ESY) summer program for special education students. Using summer 2019 data, we find that replacing the actual student-to-school assignment with our method could lead to total cost savings of up to 8%. Our models can also be used to quantify the costs and benefits of particular operational policies, providing administrators with an analytics framework to evaluate potential decisions.
Bio:
Arthur Delarue is a fifth-year PhD student in the MIT Operations Research Center, advised by Dimitris Bertsimas. His primary aim as a researcher is to leverage data, optimization, and machine learning, in order to solve practical problems that matter to society. In particular, he is interested in applications of mixed-integer optimization in transportation, machine learning, educational operations and public policy. He is a recipient of the MIT ORC Best Student Paper Award and the William Pierskalla Best Paper Award, as well as a Franz Edelman Laureate.
]]>
It will be a fun time!
We cordially invite you to participate in the workshop. Registration is free for everyone and can be found here: https://sites.gatech.edu/eso-2020/. Workshop talks and discussions will be recorded. We ask everyone who would like to have their image/sound recorded to sign a document, which will come in an email after you submit registration. Online meeting links will be made available on the workshop website shortly.
Please feel free to email us if you have any questions concerning the workshop.
Hope to see you all!
Organizing Committee
Andy Sun (andy.sun@isye.gatech.edu)
Valerie Thomas (Valerie.thomas@isye.gatech.edu)
Pascal Van Hentenryck (pascal.vanhentryck@isye.gatech.edu)
H. Milton Stewart School of Industrial and Systems Engineering
Georgia Institute of Technology
Atlanta, GA 30332
Abstract: We consider a parallel processing network with removable servers. Beginning with the single server model with power and service rate control, we study the importance of a delayed restart when the server is off. In particular, we show that an optimal policy exists (under the average cost criterion) that delays restarting until a “safety stock” of work is in the system. It then behaves similarly to that of the classic service rate control models. With that as the backdrop, we consider scheduling with the ability to remove servers. We introduce “delay-JSQ” (join the shortest queue) policies, show their stability and asymptotic optimality in the two-server case, and conclude with a detailed numerical study that shows they outperform JSQ by up to 80%. This is joint work with Professor Douglas Down from McMaster University and Dr. Pamela Badian-Pessot (now at Proctor and Gamble).
Bio: Professor Lewis received his Ph.D. from Georgia Tech’s ISyE department in 1998 (under the expert tutelage of Hayriye Ayhan and Bob Foley). He spent a year as a postdoctoral researcher in the Centre for Operations Excellence at the University of British Columbia. Mark’s first (tenure track) academic job was at the University of Michigan’s Department of Industrial and Operation Engineering. Professor Lewis joined Cornell’s School of Operations Research and Information Engineering in 2005. His research interest are at the nexus of dynamic decision-making and stochastic processes, where he splits his time studying Markov decision process theory and dynamic queueing control. Before becoming director of the school in 2019, Mark was appointed Senior Associate Dean for Diversity and Faculty Development in the College of Engineering from 2015 to 2020. Some of his accolades include NSF’s Presidential Early Career Award for Science and Engineering (PECASE) and the Sloan Foundation’s Mentor of the Year Award.
]]>IPIC 2020 aims to provide an open forum for researchers, industry representatives, government officials and citizens to together explore, discuss, introduce leading edge concepts, methodologies, recent projects, technological advancements,start-up initiatives, for current and future Physical Internet implementation.
Conference topics include inter-connected logistics, PI fundamentals, business models, governance and implementation, cross-chain control, synchromodal transportation, IT systems, stakeholders and their roles. New business models, enabling technologies and experimentations already underway will be presented, making this meeting a unique opportunity to learn, network and discuss the latest results and challenges about interconnected logistics. And, because logistics is global, participants will be from all over the world including researchers, industrial and international institution members, local authorities and standardization committees.
Georgia Tech Supply Chain Students
Please plan on joining us for Virtual Supply Chain Day! Visit https://www.scl.gatech.edu/supplychainday/students to learn how your can participate online.
Interested Organizations/Recruiters
If you are a organization who would like to participate, please email event@scl.gatech.edu.
The event will host supply chain representatives from supply chain and logistics companies who will be on online to educate supply chain students about their organizations, available employment, internships, and capstone project opportunities.
We strongly encourage students to act now to seek full-time employment, internships, and projects (rather than waiting until the end of the semester).
For more information relating to the session, please visit https://www.scl.gatech.edu/supplychainday.
]]>
This course focuses on defining, executing, and improving the S&OP process. Participants will be introduced to the appropriate stakeholders of S&OP, the importance of S&OP to corporate performance, S&OP cadence, and the use of visionary technology to bring S&OP to the next level. Business cases will be used to show concrete examples of companies where S&OP is effectively applied.
Upon completion of this course, you will be able to:
This course is the second in the four-course Supply Chain Analytics Professional certificate program. It emphasizes operational performance metrics to align supply chain management with strategic business goals. You’ll learn several statistics concepts (e.g. variance analysis, hypothesis testing, forecasting methods) along with inventory management models. You’ll use diagnostic analytics with PowerBI and Python to conduct demand and service profiling, undertake root cause analysis, and use time series forecasting in inventory management.
Experienced business professionals who perform or want to perform analytics to improve their supply chain management processes. They want to tackle strategic goals and to perform leading edge analytics projects that address the full complexity of supply chains.
Introduced by ISyE School Chair Edwin Romeijn, McKeehan will talk about being a woman in a male-dominated field and how to leverage your Georgia Tech education to expand your influence.
Register here for the event on Tuesday, October 20, 2020, at 11:00 a.m.
]]>
Bio: Laurent Massoulié is research director at Inria, head of the Microsoft Research – Inria Joint Centre, and professor at the Applied Maths Centre of Ecole Polytechnique. His research interests are in machine learning, probabilistic modelling and algorithms for networks. He has held research scientist positions at: France Telecom, Microsoft Research, Thomson-Technicolor, where he headed the Paris Research Lab. He obtained best paper awards at IEEE INFOCOM 1999, ACM SIGMETRICS 2005, ACM CoNEXT 2007, NeurIPS 2018, was elected "Technicolor Fellow" in 2011, received the "Grand Prix Scientifique" of the Del Duca Foundation delivered by the French Academy of Science in 2017, and is a Fellow of the “Prairie” Institute.
]]>Joint work with Luca Ganassali
Abstract:
In this work we consider alignment of sparse graphs, for which we introduce the Neighborhood Tree Matching Algorithm (NTMA). For correlated Erdős-Rényi random graphs, we prove that the algorithm returns -- in polynomial time -- a positive fraction of correctly matched vertices, and a vanishing fraction of mismatches. This result holds with average degree of the graphs in O(1) and correlation parameter s bounded away from 1, conditions under which random graph alignment is particularly challenging. As a byproduct of the analysis we introduce a matching metric between trees and characterize it for several models of correlated random trees. These results may be of independent interest, yielding for instance efficient tests for determining whether two random trees are correlated or independent.
]]>
Abstract: In this talk, we introduce a new approach for the vehicle routing problem with stochastic demands for the case in which customer demands are revealed before vehicles are dispatched. Our approach combines ideas from vehicle routing and manufacturing process flexibility to propose overlapped routing strategies with customer sharing. We characterize the asymptotic performance of the overlapped routing strategies under probabilistic analysis. Using the characterization, we demonstrate that our overlapped routing strategies perform close to the theoretical lower-bound derived from the reoptimization strategy, and significantly outperforms the routing strategy without overlapped routes. Furthermore, we verify effectiveness of the proposed overlapped routing strategies in non-asymptotic regimes through numerical analysis. This is joint work with Kirby Ledvina, Hanzhang Qin, and David Simchi-Levi.
Bio: Yehua Wei is an Associate Professor in Decision Sciences at Duke University. His research interest includes optimizing decisions under uncertainty, designing sparse resource pooling systems, and strategic customers in queueing networks. His research has won a number of awards, including second place in the 2011 George Nicholson Paper Competition, winner of the 2014 Daniel H. Wagner Prize for Excellence in Operations Research Practice, winner of the 2019 Service Management SIG Prize, and finalist of the 2019 M&SOM Practice-Based Paper Competition.
]]>
Bio: Professor Paul Harper is Professor of Operational Research in the School of Mathematics at Cardiff University, Deputy Head of School, and Director of the University’s Data Innovation Research Institute. His research interests are in stochastic OR, including queueing theory, simulation methods, optimisation and game theory, and applications to healthcare. Author of more than 90 peer-reviewed papers and book chapters, Paul has been a named investigator on £11million of funded research grants, recipient of a Times Higher Education award for ‘Outstanding Contribution to Innovation and Technology, founding editor-in-chief of the international journal Health Systems (Taylor & Francis) and Director of Health Modelling Centre Cymru. He is also an elected Fellow of the Learned Society of Wales (FLSW) and recipient of the 2018 'Companion of OR' Award (The UK OR Society). profpaulharper.com
]]>
Georgia Tech Supply Chain Students
Please plan on joining us for our first Virtual Supply Chain Day! Visit https://www.scl.gatech.edu/supplychainday/students to learn how your can participate online.
Interested Organizations/Recruiters
If you are a organization who would like to participate, please email event@scl.gatech.edu.
The event will host supply chain representatives from Castellini Company, Chainalytics, Cummins, Dematic, Deposco, Diageo North America, Kinaxis, Micron Technology, Motion Industries, Norfolk Southern, Novelis, OM Partners, ORTEC International USA, Parker Hannifin Corporation, SF Technology, Siemens Healthineers, Smith & Nephew, STORD, The Home Depot, Vector Global Logistics, and Verusen (Autit, Inc) who will be on online to educate supply chain students about their organizations, available employment, internships, and capstone project opportunities.
We strongly encourage students to act now to seek full-time employment, internships, and projects (rather than waiting until the end of the semester).
]]>
We focus on a two-layer distribution network that has been implemented in practice by major e-retailers. We analyze the performance of fulfillment policies that do not rely on demand forecasts. We show that, under an assumption that aligns with the practice of most e-retailers, a myopic policy achieves a constant performance ratio compared to an optimal clairvoyant algorithm. We further show that this ratio is tight. More generally, we characterize the performance guarantees of the myopic policy in terms of the cost param- eters of the model. Finally, we complement our theoretical results by conducting a numerical study and demonstrate the good performance of the myopic policy.
This is joint work with Xinshang Wang (Alibaba) and Yanyang Zhao (Chicago Booth).
Bio:
Linwei Xin is an assistant professor of Operations Management at the University of Chicago Booth School of Business. He graduated from ISyE in 2015, advised by David A. Goldberg and Alexander Shapiro. His research interests include supply chain, inventory and revenue management, optimization under uncertainty, and data-driven decision-making. His work has been recognized with several INFORMS paper competition awards, including the 2019 Applied Probability Society Best Publication Award, First Place in the 2015 George E. Nicholson Student Paper Competition, Second Place in the 2015 Junior Faculty Interest Group Paper Competition, and a finalist in the 2014 Manufacturing and Service Operations Management Student Paper Competition. His research has been published in journals such as Operations Research and Management Science. He won a NSF grant as PI. He also has worked with companies/organizations through research collaboration including Alibaba Group and Walmart Global eCommerce.
]]>
We focus on a two-layer distribution network that has been implemented in practice by major e-retailers. We analyze the performance of fulfillment policies that do not rely on demand forecasts. We show that, under an assumption that aligns with the practice of most e-retailers, a myopic policy achieves a constant performance ratio compared to an optimal clairvoyant algorithm. We further show that this ratio is tight. More generally, we characterize the performance guarantees of the myopic policy in terms of the cost param- eters of the model. Finally, we complement our theoretical results by conducting a numerical study and demonstrate the good performance of the myopic policy.
This is joint work with Xinshang Wang (Alibaba) and Yanyang Zhao (Chicago Booth).
]]>
This course is the first in the four-course Supply Chain Analytics Professional certificate program. It prepares you to apply leading-edge analytical methods and technology enablers across the supply chain. You’ll learn the dynamics of supply chains, the most relevant planning challenges, and the roles of different types of analytics. Next, you’ll learn about data cleansing, exploratory data analysis, and visualization. You’ll use Python and PowerBI to analyze the causes of underperformance and to build dashboards to visualize supply chain data. You will leave knowing how to gather, analyze, and prepare your data through descriptive analytics before you dig into deeper applications.
Experienced business professionals who perform or want to perform analytics to improve their supply chain management processes. They want to tackle strategic goals and to perform leading edge analytics projects that address the full complexity of supply chains.
The course is designed to help participants understand how decisions impact financial performance, identify initiatives to improve company performance, and build better business cases with the overall goal of improving financial acumen and decision-making. The course utilizes hands-on applications, group discussion, and exercises.
The course is comprised of (3) 90-minute instructor-led LIVE group webinars (Sept 9, 16 , 23 | 1:30-3pm EDT) with each webinar requiring (1) 90-minute session of online course work to be completed before each webinar (total of 9 hours). Participants will be able to access the online "pre-work" material Wednesday, August 26th. Please note that the course begins September 9, 2020 and participants must participate in all LIVE webinars.
Early or middle stage career professionals who are or will be responsible for executing organizational strategy tied to an integrated view of the organization (including professionals from distribution and logistics, production, and operations).
Often the lack of cooperation and coordination between organizations or stakeholders lead to inefficiencies, despite having common goals. A systems view is needed to ensure appropriate use of scarce resources to meet the multiple, and often conflicting, short- and long-term goals from multiple constituents. This course will focus on conceptual and modeling skills to understand and effectively manage supply chains and operations from a systems perspective. Models will address system characteristics (e.g., demand dependencies) that drive system dynamics and policies to regulate performance. Course topics include methods for improving coordination and collaboration, addressing demand dependencies, and reliably measuring and evaluating system performance.
This course is designed for representatives from governmental or non-governmental organizations, private corporations, military, and foundations, including but not limited to senior executives overseeing administrative and operational functions of an organization, logistics and supply chain managers, program managers, directors of field operations, directors of emergency/disaster preparedness and response, and public health professionals.
NOTE: Pre-course activities will be conducted online using the Canvas platform online learning management system. Access instructions will be provided to registrants when details become available.
Pre-Course Activities - Conducted online via GTPE platform
Coordination and Collaboration – 2 hours
Game Theory/Incentives – 1 hour
System Dynamics - 2 hour
Provided
Recommended
For those interested in earning the Health and Humanitarian Supply Chain Management Certificate, this course is the third and final of the three-course certificate program. To earn the certificate, participants must register and complete the following courses within three years:
The Georgia Tech Health and Humanitarian Logistics Professional Certificate Program is an executive education program designed for practitioners in non-governmental organizations (NGOs), government, industry, and the military who are active participants in health and humanitarian sectors. The courses are developed for practitioners seeking to build skills to improve decision making in preparedness, response operations planning, and system design. Courses include many interactive components, such as case studies and games, which help professionals in the humanitarian world to link the challenges and decision-making tradeoffs they face in practice with the systematic approaches, tools, and techniques presented.
This course provides for 1.40 continuing education units (CEUs).
Standard: $2,400.00, Alumni/Org Discount: $2,160.00, Certificate: $1,992.00 (cost of each course when signing up for and paying for a multi-course certificate program).
First time attendees pay the listed course fee. If you are a returning student of the Supply Chain & Logistics Institute (SCL) courses, you will receive a 10% discount which you will enter at the "Check Out" page. Use Coupon Code SCL-Alum.
There are also discounts available for multiple-team member registrations, to those who prepay for all the courses in a specific certificate, to active/retired military, or to members of certain organizations (click this link for a listing).
Discounts cannot be combined or used for online formatted courses. To receive the coupon code for these discounts, call 404-385-8663 or send us an email prior to registration
The program fee for In Person courses (non-online) includes continental breakfasts, lunches, breaks, parking, internet access, and all classroom materials.
If CHHS must cancel a program, registrants will receive a full refund. Georgia Tech, however, cannot assume the responsibility for other costs incurred. Due to program enrollment limits, early registration is encouraged. Registrations will be acknowledged by a letter of confirmation from Professional Education.
**The 2020 Virtual Program, will also be offered at a reduced cost of $4,200 (originally $6,000). Courses taken individually (not the full 3 course program) will be $1800 each rather than the original fee of $2400.
On the first day, please check in at least 30 minutes before the class start time.
Each course will run for 1 week Monday through Thursday from 9:30am to 1:00pm EDT each day.
]]>Meeting demand in a timely and cost-effective manner is important both in public and private supply chains, and heavily depend on the design and operation of these supply chains. Demand is affected by ongoing factors such as local economy, infrastructure, and geographic location, as well as unexpected events such as natural or manmade disasters or other large-scale disruptions. Designing and operating responsive supply chains requires the consideration of uncertainty in timing, scope, scale, and understanding of various topics such as forecasting, distribution network design, and inventory management. This course will examine methods and models for making supply chain design and operational decisions and explore the significant value that is obtained through informed decision-making in advance of an unpredictable event or long-term strategy for meeting the need of customers and beneficiaries.
This course is designed for representatives from governmental or non-governmental organizations, private corporations, military, and foundations, including but not limited to senior executives overseeing administrative and operational functions of an organization, logistics and supply chain managers, program managers, directors of field operations, directors of emergency/disaster preparedness and response, and public health professionals.
NOTE: Pre-course activities will conducted online using the CANVAS online learning management system. Access instructions will be provided to registrants when details become available.
Provided
Recommended
For those interested in earning the Health and Humanitarian Supply Chain Management Certificate, this course is the first of the three-course certificate program. To earn the certificate, participants must register and complete the following courses within three years:
The Georgia Tech Health and Humanitarian Logistics Professional Certificate Program is an executive education program designed for practitioners in non-governmental organizations (NGOs), government, industry, and the military who are active participants in health and humanitarian sectors. The courses are developed for practitioners seeking to build skills to improve decision making in preparedness, response operations planning, and system design. Courses include many interactive components, such as case studies and games, which help professionals in the humanitarian world to link the challenges and decision-making tradeoffs they face in practice with the systematic approaches, tools, and techniques presented.
This course provides for 1.40 continuing education units (CEUs).
Julie Swann, Pinar Keskinocak, Ozlem Ergun
Standard: $2,400.00, Alumni/Org Discount: $2,160.00, Certificate: $1,992.00 (cost of each course when signing up for and paying for a multi-course certificate program).
First time attendees pay the listed course fee. If you are a returning student of the Supply Chain & Logistics Institute (SCL) courses, you will receive a 10% discount which you will enter at the "Check Out" page. Use Coupon Code SCL-Alum.
There are also discounts available for multiple-team member registrations, to those who prepay for all the courses in a specific certificate, to active/retired military, or to members of certain organizations (click this link for a listing).
Discounts cannot be combined or used for online formatted courses. To receive the coupon code for these discounts, call 404-385-8663 or send us an email prior to registration
The program fee for In Person courses (non-online) includes continental breakfasts, lunches, breaks, parking, internet access, and all classroom materials.
If CHHS must cancel a program, registrants will receive a full refund. Georgia Tech, however, cannot assume the responsibility for other costs incurred. Due to program enrollment limits, early registration is encouraged. Registrations will be acknowledged by a letter of confirmation from Professional Education.
**The 2020 Virtual Program, will also be offered at a reduced cost of $4,200 (originally $6,000). Courses taken individually (not the full 3 course program) will be $1800 each rather than the original fee of $2400.
On the first day, please check in at least 30 minutes before the class start time.
Each course will run for 1-week Monday through Thursday from 9:30am to 1:00pm EDT each day.
]]>Dr. Benoit Montreuil will emphasize the impactful interplay between supply chains and pandemics, across the world and in each locality. The impact of collectively improving pandemic supply chain readiness and response in each of the five world states - healthy, outbreak, epidemics, pandemics, and recovery - will be highlighted.
The focus will be on critical and essential supply chains to demonstrate the capital importance of supply chains in fighting and surviving a pandemics, as revealed by the COVID-19 crisis.
Key avenues for collectively and comprehensively enabling effective and efficient pandemic supply chain readiness and response in each of the world states will be presented. These avenues combine common-sense pragmatic practices to leading-edge concepts and technologies, notably based on Artificial Intelligence and the Physical Internet. They notably enable:
The presentation results from the interdisciplinary collaboration of 28 professors from Georgia Tech Supply Chain and Logistics Institute.
Please register online via Eventbrite if you plan to attend.
The webinar series aims to challenge us beyond the immediate crisis and onto a newer state where we have another chance to build strong community-research partnerships for good. By focusing on complex, societal problems that communities all over GA and the rest of the world face, we aim to provide innovative research and create partnerships to empower all. We adhere to GT’s Strategic Plan and Mission on “developing leaders who advance technology and improve the human condition.”
Let's all be #SMARTer Together.
]]>Abstract:
Modeling and tractable computation form two fundamental but competing pillars of data science; indeed, fitting good models to data is often computationally challenging in modern applications. Focusing on the canonical tasks of ranking and regression, I introduce problems where this tension is immediately apparent, and present methodological solutions that are both statistically sound and computationally tractable.
I begin by describing a class of “permutation-based" models as a flexible alternative to parametric modeling in a host of inference problems including ranking from ordinal data. I introduce procedures that narrow a conjectured statistical-computational gap, demonstrating that carefully chosen non-parametric structure can significantly improve robustness to mis-specification while maintaining interpretability. Next, I address a complementary question in the context of convex regression, where I show that the curse of dimensionality inherent to non-parametric modeling can be mitigated via parametric approximation. Our provably optimal methodology demonstrates that it is often possible to enhance the interpretability of non-parametric models while maintaining important aspects of their flexibility.
Bio:
Ashwin Pananjady is a PhD student in the Department of Electrical Engineering and Computer Sciences at the University of California Berkeley, advised by Martin Wainwright and Thomas Courtade. His interests lie broadly in statistics, machine learning, information theory, and optimization, and include ranking and permutation estimation, high-dimensional and non-parametric statistics, high-dimensional probability, and reinforcement learning. He is a recipient of the inaugural Lawrence Brown PhD student award from the Institute of Mathematical Statistics, an Outstanding Graduate Student Instructor award from UC Berkeley, and the Governor's Gold Medal from IIT Madras.
]]>Abstract:
Modeling and tractable computation form two fundamental but competing pillars of data science; indeed, fitting good models to data is often computationally challenging in modern applications. Focusing on the canonical tasks of ranking and regression, I introduce problems where this tension is immediately apparent, and present methodological solutions that are both statistically sound and computationally tractable.
I begin by describing a class of “permutation-based" models as a flexible alternative to parametric modeling in a host of inference problems including ranking from ordinal data. I introduce procedures that narrow a conjectured statistical-computational gap, demonstrating that carefully chosen non-parametric structure can significantly improve robustness to mis-specification while maintaining interpretability. Next, I address a complementary question in the context of convex regression, where I show that the curse of dimensionality inherent to non-parametric modeling can be mitigated via parametric approximation. Our provably optimal methodology demonstrates that it is often possible to enhance the interpretability of non-parametric models while maintaining important aspects of their flexibility.
Abstract:
Single-sourcing lost-sales inventory systems with lead times are notoriously difficult to optimize. Recent numerical experiments have suggested that a so-called capped base-stock policy demonstrates superior performance compared with existing heuristics. However, the superior performance lacks of a theoretical foundation (in the stochastic setting) and why such policies generally perform so well remains a major open question. In this paper, we provide a theoretical foundation for this phenomenon. In a continuous review lost-sales inventory model with lead times and Poisson demand, we prove that this policy has a worst-case performance guarantee of 1.79 by conducting an asymptotic analysis under large penalty cost and lead time following Reiman (2004). This result provides a deeper understanding of the superior numerical performance of capped base-stock policies, and presents a new approach to proving worst-case performance guarantees of simple policies in notoriously hard inventory problems.
Bio:
Linwei Xin is an assistant professor of Operations Management at the University of Chicago Booth School of Business. He graduated from ISyE in 2015, advised by David A. Goldberg and Alexander Shapiro. His research interests include supply chain, inventory and revenue management, optimization under uncertainty, and data-driven decision-making. His work has been recognized with several INFORMS paper competition awards, including the 2019 Applied Probability Society Best Publication Award, First Place in the 2015 George E. Nicholson Student Paper Competition, Second Place in the 2015 Junior Faculty Interest Group Paper Competition, and a finalist in the 2014 Manufacturing and Service Operations Management Student Paper Competition. His research has been published in journals such as Operations Research and Management Science. He won a NSF grant as PI. He also has worked with companies/organizations through research collaboration including Alibaba Group and Walmart Global eCommerce.
]]>Abstract:
Single-sourcing lost-sales inventory systems with lead times are notoriously difficult to optimize. Recent numerical experiments have suggested that a so-called capped base-stock policy demonstrates superior performance compared with existing heuristics. However, the superior performance lacks of a theoretical foundation (in the stochastic setting) and why such policies generally perform so well remains a major open question. In this paper, we provide a theoretical foundation for this phenomenon. In a continuous review lost-sales inventory model with lead times and Poisson demand, we prove that this policy has a worst-case performance guarantee of 1.79 by conducting an asymptotic analysis under large penalty cost and lead time following Reiman (2004). This result provides a deeper understanding of the superior numerical performance of capped base-stock policies, and presents a new approach to proving worst-case performance guarantees of simple policies in notoriously hard inventory problems.
]]>These problems have been addressed in the academic literature using a variety of modeling and algorithmic frameworks, including dynamic programming, stochastic programming, stochastic control, simulation optimization, approximate dynamic programming/reinforcement learning, and even multiarmed bandit problems.
I will describe a universal modeling framework that can be used for any sequential decision problem in the presence of different sources of uncertainty. The framework is centered on an optimization problem that optimizes over policies (rules for making decisions), where we show that there are two fundamental strategies for designing policies (policy search and policies based on lookahead approximations), each of which further divide into two classes, creating four (meta)classes of policies that are the foundation of any solution approach that has ever been proposed for a sequential problem. I will demonstrate these policies in two broad contexts: pure learning problems (“bandit problems”) and dynamic resource allocation problems, where I will use a simple energy storage problem to show that each of the four classes (and a hybrid) can be made to work best.
]]>These problems have been addressed in the academic literature using a variety of modeling and algorithmic frameworks, including dynamic programming, stochastic programming, stochastic control, simulation optimization, approximate dynamic programming/reinforcement learning, and even multiarmed bandit problems.
I will describe a universal modeling framework that can be used for any sequential decision problem in the presence of different sources of uncertainty. The framework is centered on an optimization problem that optimizes over policies (rules for making decisions), where we show that there are two fundamental strategies for designing policies (policy search and policies based on lookahead approximations), each of which further divide into two classes, creating four (meta)classes of policies that are the foundation of any solution approach that has ever been proposed for a sequential problem. I will demonstrate these policies in two broad contexts: pure learning problems (“bandit problems”) and dynamic resource allocation problems, where I will use a simple energy storage problem to show that each of the four classes (and a hybrid) can be made to work best.
]]>A Unified Framework on Sequential Decisions under Uncertainty
Georgia Tech, ISyE
Warren B Powell
Princeton University
April 3, 2020
Sequential decision problems arise in application areas that include engineering, the sciences, transportation, logistics, health services, medical decision making, energy, e-commerce and finance, along multiagent problems that arise with drones and robotics. These problems have been addressed in the academic literature using a variety of modeling and algorithmic frameworks, including dynamic programming, stochastic programming, optimal control, simulation optimization, approximate dynamic programming/reinforcement learning, and even multiarmed bandit problems.
In sharp contrast with the field of deterministic math programming which enjoys a canonical framework that is used around the world, sequential decision problems are described in the literature using at least eight fundamental modeling languages plus at least six more derivative dialects. Further, application communities have developed a wide range of solution methods that reflect the characteristics of specific problem classes.
In this workshop, I will introduce a single, canonical modeling framework that covers every sequential decision problem. The framework consists of five dimensions (drawing heavily on the approach used by stochastic control) that naturally represent real-world problems and map directly into software (and vice versa). The framework clearly lays out the elements of any sequential decision problem that have to be modeled, which helps to clarify the understanding of complex systems.
Special attention will be given to the modeling of state variables, which are a surprising source of confusion in the research literature. I will illustrate the different classes of state variables, including belief states, using a series of energy storage problems. Through proper handling of state variables, I show how we can model pure learning problems, pure resource allocation problems, hybrid learning/resource allocation problems, and contextual problems.
Our modeling strategy is centered on the challenge of optimizing over policies (functions for making decisions). I will describe four (meta) classes of policies, that can be divided into two broad groups: the policy search class (finding the best function that works well over time), and lookahead policies that approximate the downstream impact of making a decision now. I will claim that these four classes are universal: any method proposed in the literature (or used in practice) falls into one of these four classes, or a hybrid of two or more. I will illustrate parametric cost function approximations that are widely used in practice, but almost completely ignored by the academic community.
I will illustrate the four classes of policies in the context of pure learning problems, and the much richer class of state-dependent problems. In the process, I will highlight the strengths of policies in the policy search class (simplicity, ability to incorporate structure) and the weaknesses (tunable parameters). I will use an energy storage problem to show that we can make each of the four classes of policies (and possibly a hybrid) work best depending on the characteristics of the data.
]]>In sharp contrast with the field of deterministic math programming which enjoys a canonical framework that is used around the world, sequential decision problems are described in the literature using at least eight fundamental modeling languages plus at least six more derivative dialects. Further, application communities have developed a wide range of solution methods that reflect the characteristics of specific problem classes.
In this workshop, I will introduce a single, canonical modeling framework that covers every sequential decision problem. The framework consists of five dimensions (drawing heavily on the approach used by stochastic control) that naturally represent real-world problems and map directly into software (and vice versa). The framework clearly lays out the elements of any sequential decision problem that have to be modeled, which helps to clarify the understanding of complex systems.
Special attention will be given to the modeling of state variables, which are a surprising source of confusion in the research literature. I will illustrate the different classes of state variables, including belief states, using a series of energy storage problems. Through proper handling of state variables, I show how we can model pure learning problems, pure resource allocation problems, hybrid learning/resource allocation problems, and contextual problems.
Our modeling strategy is centered on the challenge of optimizing over policies (functions for making decisions). I will describe four (meta) classes of policies, that can be divided into two broad groups: the policy search class (finding the best function that works well over time), and lookahead policies that approximate the downstream impact of making a decision now. I will claim that these four classes are universal: any method proposed in the literature (or used in practice) falls into one of these four classes, or a hybrid of two or more. I will illustrate parametric cost function approximations that are widely used in practice, but almost completely ignored by the academic community.
I will illustrate the four classes of policies in the context of pure learning problems, and the much richer class of state-dependent problems. In the process, I will highlight the strengths of policies in the policy search class (simplicity, ability to incorporate structure) and the weaknesses (tunable parameters). I will use an energy storage problem to show that we can make each of the four classes of policies (and possibly a hybrid) work best depending on the characteristics of the data.
]]>The finite dimensional McCormick second-order sufficiency theory for non-linear programming problems with a finite number of constraints is now a classical part of the optimization literature. It was introduced by McCormick in 1967 and an improved version was given by Fiacco and McCormick in their 1968 award winning book. Later it was learned that Pennisi in 1953 had presented exactly the same theory. Many authors, most notably Maurer and Zowe in a highly visible paper in 1978 argue that the Pennisi- McCormick theory cannot be extended to infinite dimensions without adding further assumptions, by producing a counter example. They then extend the theory to infinite dimensions, allowing for an infinite number of constraints, by adding additional assumptions. In the current work we use a fundamental principle for second-order sufficiency to extend, the Pennisi-McCormick second-order theory as stated in Rn to infinite dimensional normed vector spaces, without adding additional assumptions. The Maurer and Zowe infinite dimensional counter example carried an infinite number of constraints. Hence they seemed to be unaware that the extension of the Pennisi-McCormick theory to infinite dimensions was possible provided the original feature of a finite number of constraints was maintained.
Bio:
Richard A. Tapia, University Professor and Maxfield-Oshman Professor of Engineering, Rice University was born in Los Angeles to parents who emigrated from Mexico when they were children, seeking educational opportunities. He was the first in his family to attend college, earning his B.A., M.A., and Ph.D. degrees in mathematics from the University of California, Los Angeles. Due to his efforts, Rice University has received national recognition for its educational outreach programs, and the Rice Computational and Applied Mathematics Department has become a national leader in producing women and underrepresented minority Ph.D.s in the mathematical sciences.
Dr. Tapia’s major research contributions have been in the area of computational optimization, both linear and nonlinear programming, where he pioneered the exploration and settlement of the important computational methods in numerical optimization known as primal-dual interior point methods. Tapia has authored or co-authored two books and more than 100 mathematical research papers, and is currently authoring a graduate level textbook on the foundations of optimization.
Dr. Tapia’s honors include: election to the National Academy of Engineering (1992) for his seminal work in interior point methods; being the first recipient of the A. Nico Habermann Award from the Computing Research Association (1994) for outstanding contributions in aiding members of underrepresented groups within the computing community; the Presidential Award for Excellence in Science, Mathematics, and Engineering Mentoring from President Bill Clinton (1996); appointment by President Clinton to the National Science Board, the governing body of the National Science Foundation (1996); the Lifetime Mentor Award from the American Association for the Advancement of Science (1997); and the establishment of a lecture series to honor Dr. Tapia and African American mathematician David Blackwell at Cornell University (2000). The Richard Tapia Celebration of Diversity in Computing honors his many contributions to diversity (2001). He received the Hispanic Engineer of the Year Award from Hispanic Engineer Magazine in 1996, and was inducted into the Hispanic Engineer National Achievement Awards Conference Hall of Fame in 1997. Hispanic Engineer & Informational Technology Magazine also selected him as one of the 50 Most Important Hispanics in Technology and Business for 2004. That same year Dr. Tapia was inducted into the Texas Science Hall of Fame.
Dr. Tapia has been named one of 20 most influential leaders in minority math education by the National Research Council; listed as one of the 100 most influential Hispanics in the U.S. by Hispanic Business magazine (2008); and given the “Professor of the Year” award by the Association of Hispanic School Administrators, Houston Independent School District, Houston, Texas. In 2005, Tapia was elected to the Board of Directors for TAMEST, comprising the Texas members of the National Academy of Engineering, National Academy of Sciences and the Institute of Medicine. In 2009, Tapia received the Hispanic Heritage Award for Math and Science. In 2011, President Obama named Dr. Tapia one of the recipients of the National Medal of Science, the highest honor bestowed by the United States government on scientists and engineers.
]]>The finite dimensional McCormick second-order sufficiency theory for non-linear programming problems with a finite number of constraints is now a classical part of the optimization literature. It was introduced by McCormick in 1967 and an improved version was given by Fiacco and McCormick in their 1968 award winning book. Later it was learned that Pennisi in 1953 had presented exactly the same theory. Many authors, most notably Maurer and Zowe in a highly visible paper in 1978 argue that the Pennisi- McCormick theory cannot be extended to infinite dimensions without adding further assumptions, by producing a counter example. They then extend the theory to infinite dimensions, allowing for an infinite number of constraints, by adding additional assumptions. In the current work we use a fundamental principle for second-order sufficiency to extend, the Pennisi-McCormick second-order theory as stated in Rn to infinite dimensional normed vector spaces, without adding additional assumptions. The Maurer and Zowe infinite dimensional counter example carried an infinite number of constraints. Hence they seemed to be unaware that the extension of the Pennisi-McCormick theory to infinite dimensions was possible provided the original feature of a finite number of constraints was maintained.
]]>Learning & Exploiting Low-Dimensional Structure in High-Dimentional Data
This talk will focus on the problem of learning low-dimensional geometric structure in high-dimensional data. We allow the lower-dimensional subspace to be non-linear. There are a variety of algorithms available for “manifold learning” and non-linear dimensionality reduction, mostly relying on locally linear approximations and not providing a likelihood-based approach for inferences. We propose a new class of simple geometric dictionaries for characterizing the subspace, along with a simple optimization algorithm and a model-based approach to inference. We provide strong theory support, in terms of tight bounds on covering numbers, showing advantages of our approach relative to local linear dictionaries. These advantages are shown to carry over to practical performance in a variety of settings including manifold learning, manifold de-noising, data visualization, classification (providing a competitor to deep neural networks that requires fewer training examples), and geodesic distance estimation. We additionally provide a Bayesian nonparametric methodology for inference, using a new class of kernels, which is shown to outperform current methods, such as mixtures of multivariate Gaussians.
David Dunson is Arts & Sciences Distinguished Professor of Statistical Science and Mathematics at Duke University. His research focuses on developing methodology for analysis and interpretation of complex and high-dimensional data, with a particular emphasis on Bayesian and probability modeling approaches. He is particularly interested in work at the intersection of statistics, differential geometry, and computer science. Methods development and theory are directly motivated by applications in neuroscience, genomics, environmental health, and ecology among others. In these settings, it is common for data to have a structured form, consisting of replicated networks/graphs, trees, functions, tensors, etc. A focus is on developing fundamentally new frameworks for statistical inferences in challenging settings, including improving robustness to modeling assumptions and scalability to large datasets. He has won numerous awards, including the 2010 COPSS Presidents’ Award, which is widely viewed as the most prestigious award in statistics and represents statistics version of the Field’s Medal, being given to one outstanding researcher under the age of 41 per year internationally. His work has had substantial impact, with ~48,000 citations on google scholar and an H-index of 75.
]]>This talk will focus on the problem of learning low-dimensional geometric structure in high-dimensional data. We allow the lower-dimensional subspace to be non-linear. There are a variety of algorithms available for “manifold learning” and non-linear dimensionality reduction, mostly relying on locally linear approximations and not providing a likelihood-based approach for inferences. We propose a new class of simple geometric dictionaries for characterizing the subspace, along with a simple optimization algorithm and a model-based approach to inference. We provide strong theory support, in terms of tight bounds on covering numbers, showing advantages of our approach relative to local linear dictionaries. These advantages are shown to carry over to practical performance in a variety of settings including manifold learning, manifold de-noising, data visualization, classification (providing a competitor to deep neural networks that requires fewer training examples), and geodesic distance estimation. We additionally provide a Bayesian nonparametric methodology for inference, using a new class of kernels, which is shown to outperform current methods, such as mixtures of multivariate Gaussians.
]]>We study a discrete-time interacting particle system, which can be called a totally asymmetric simple exclusion process with holdback (TASEP-H). There are rho*n particles, rho < 1, moving clockwise, in discrete time, on n sites arranged in a circle. The “holdback" refers to the property that the probability of a particle moving forward to a vacant site depends on the presence of a particle immediately “behind” it. The model is motivated by a communication network with packets moving along a sequence of nodes under a “standard” random access algorithm. Another motivation is a slow-to-start model of car traffic. We focus on the dependence of the steady-state flux (throughput) on the density rho, when n is large. We show that when rho exceeds a certain threshold, a phase transition occurs in that large particle clusters are formed and persist, making the "typical" flux different from the formal one. (Joint work with Seva Shneer, Heriot-Watt Univ.)
Bio:
Since 2017 Aleksandr Stolyar is a Founder Professor in the ISE Department and Coordinated Science Lab at UIUC. His research interests are in stochastic processes, queueing theory, and stochastic modeling of information, communication and service systems. He received Ph.D. in Mathematics from the Institute of Control Science, Moscow, in 1989, and was a research scientist at the Institute of Control Science in 1989-1991. In 1992-1998 he was working on stochastic models in telecommunications at Motorola and AT&T Research. From 1998 to 2014 he was with the Bell Labs Mathematical Sciences Research, Murray Hill, New Jersey, working on stochastic networks and resource allocation problems in a variety of applications, including wireless and wireline communications, service systems, network clouds. In 2014-2016 he was a Timothy J. Wilmott Endowed Chair Professor in the ISE Department at Lehigh University. He received INFORMS Applied Probability Society 2004 Best Publication award, SIGMETRICS'96 Best Paper award.
We study a discrete-time interacting particle system, which can be called a totally asymmetric simple exclusion process with holdback (TASEP-H). There are rho*n particles, rho < 1, moving clockwise, in discrete time, on n sites arranged in a circle. The “holdback" refers to the property that the probability of a particle moving forward to a vacant site depends on the presence of a particle immediately “behind” it. The model is motivated by a communication network with packets moving along a sequence of nodes under a “standard” random access algorithm. Another motivation is a slow-to-start model of car traffic. We focus on the dependence of the steady-state flux (throughput) on the density rho, when n is large. We show that when rho exceeds a certain threshold, a phase transition occurs in that large particle clusters are formed and persist, making the "typical" flux different from the formal one. (Joint work with Seva Shneer, Heriot-Watt Univ.)
Testing and estimation for clustered signals
We propose a change-point detection method for large scale multiple testing problems with data having clustered signals. Unlike the classic change-point setup, the signals can vary in size within a cluster. The clustering structure on the signals enables us to effectively delineate the boundaries between signal and non-signal segments. New test statistics are proposed for observations from one and/or multiple realizations. Their asymptotic distributions are derived. We also study the associated variance estimation problem. We allow the variances to be heteroscedastic in the multiple realization case, which substantially expands the applicability of the proposed method. Simulation studies demonstrate that the proposed approach has a favorable performance. Our procedure is applied to an array CGH dataset.
Hongyuan Cao is currently an associate professor of statistics at Florida State University. She got her Ph.D in statistics from UNC-Chapel Hill. Her research interests include high dimensional data, machine learning, longitudinal data analysis, survival analysis and causal inference. She currently serves as associate editor of Biometrics.
]]>We propose a change-point detection method for large scale multiple testing problems with data having clustered signals. Unlike the classic change-point setup, the signals can vary in size within a cluster. The clustering structure on the signals enables us to effectively delineate the boundaries between signal and non-signal segments. New test statistics are proposed for observations from one and/or multiple realizations. Their asymptotic distributions are derived. We also study the associated variance estimation problem. We allow the variances to be heteroscedastic in the multiple realization case, which substantially expands the applicability of the proposed method. Simulation studies demonstrate that the proposed approach has a favorable performance. Our procedure is applied to an array CGH dataset.
]]>Abstract:
Metal Additive Manufacturing (AM) offers tremendous freedom to create complex parts without the design constraints of traditional manufacturing routes. Despite its promise and potential, however, AM is still largely a solution that is used for rapid prototyping and small-batch production. While the process-structure-property-performance relationship is largely unknown for AM, it is further masked by the information-disconnected AM process chain. The silos of information hamper data exchange among the different steps in the AM process chain, causing manufacturing to be less efficient, prone to error, and lacking traceability. Therefore, how to harness the data from every step of the AM process chain, powered with data science, is urgent but promising to make AM transparent and optimized, and to produce quality parts. Digital thread is the seamless flow of data throughout the product development chain, including design concept, modeling, build plan, monitoring, quality assurance, the build process itself, and post-processing and inspection. The ability to dissect, understand, and apply the potentially massive amounts of data and intense computing demands within the digital thread allows users to enhance and scale their AM capabilities and manage the complexities of AM production. By collecting and analyzing detailed logs of real-time data from process monitoring, it is possible to recognize patterns which reveal where potential defects might occur and where process adjustments may be beneficial.
This talk will highlight our recent work toward the digital thread of AM by focusing on melt pool modeling and porosity prediction in laser metal deposition (LMD). Porosity produced in LMD hampers its application due to the absence of an effective prediction method. Measured thermal images of the melt pool provide a unique opportunity for porosity analytics. The first part of this talk will present a hierarchical spatial-temporal model for melt pool thermal images, which enables melt pool modeling and monitoring. On the other hand, a physical model may provide complementary rich data that cannot be measured otherwise. How to leverage both types of data to predict porosity is very challenging. The second part of the talk will present a physics-driven deep learning model predict porosity by integrating both measured and predicted data of the melt pool.
Bio:
Weihong “Grace” Guo is an Assistant Professor in the Department of Industrial and Systems Engineering. She earned her B.S. degree in Industrial Engineering from Tsinghua University, China, in 2010 and her Ph.D. in Industrial & Operations Engineering from the University of Michigan, Ann Arbor, in 2015. Her research focuses on developing novel methodologies for extracting and analyzing massive and complex data to facilitate effective monitoring of operational quality, early detection of system anomalies, quick diagnosis of fault root causes, and intelligent system design and control. She has collaborated with a domestic logistics/supply chain company, a university-affiliated health system and worldwide manufacturers of automobiles and personal care products. Her research has been funded by NSF, DOT, and Ford Motor Company. She received the Barbara M. Fossum Outstanding Young Manufacturing Engineer Award from the Society of Manufacturing Engineers in 2019. She also received several best paper awards from ASME. She is a member of INFORMS, IISE, ASME, SME, and Tau Beta Pi.
]]>Abstract:
Metal Additive Manufacturing (AM) offers tremendous freedom to create complex parts without the design constraints of traditional manufacturing routes. Despite its promise and potential, however, AM is still largely a solution that is used for rapid prototyping and small-batch production. While the process-structure-property-performance relationship is largely unknown for AM, it is further masked by the information-disconnected AM process chain. The silos of information hamper data exchange among the different steps in the AM process chain, causing manufacturing to be less efficient, prone to error, and lacking traceability. Therefore, how to harness the data from every step of the AM process chain, powered with data science, is urgent but promising to make AM transparent and optimized, and to produce quality parts. Digital thread is the seamless flow of data throughout the product development chain, including design concept, modeling, build plan, monitoring, quality assurance, the build process itself, and post-processing and inspection. The ability to dissect, understand, and apply the potentially massive amounts of data and intense computing demands within the digital thread allows users to enhance and scale their AM capabilities and manage the complexities of AM production. By collecting and analyzing detailed logs of real-time data from process monitoring, it is possible to recognize patterns which reveal where potential defects might occur and where process adjustments may be beneficial.
This talk will highlight our recent work toward the digital thread of AM by focusing on melt pool modeling and porosity prediction in laser metal deposition (LMD). Porosity produced in LMD hampers its application due to the absence of an effective prediction method. Measured thermal images of the melt pool provide a unique opportunity for porosity analytics. The first part of this talk will present a hierarchical spatial-temporal model for melt pool thermal images, which enables melt pool modeling and monitoring. On the other hand, a physical model may provide complementary rich data that cannot be measured otherwise. How to leverage both types of data to predict porosity is very challenging. The second part of the talk will present a physics-driven deep learning model predict porosity by integrating both measured and predicted data of the melt pool.
]]>
Abstract: In recent years, there has been an explosion of interest in designing fast nonconvex optimization algorithms to solve statistical estimation and learning problems. However, in contrast to convex optimization that has become a real pillar of modern engineering, the theoretical foundations of nonconvex optimization are far from satisfactory, especially in terms of its computational and inferential properties. This talk will present two recent stories that advance our understanding of nonconvex statistical estimation. The first story focuses on computational efficiency in solving random quadratic systems of equations. Despite the nonconvexity of the natural least-squares formulation, gradient descent with random initialization finds its global solution within a logarithmic number of iterations. The second story is concerned with uncertainty quantification for nonconvex low-rank matrix completion. We develop a de-biased estimator — on the basis of a nonconvex estimator — that enables optimal construction of confidence intervals for the missing entries of the unknown matrix. All of this is achieved via an integrated view of statistics and optimization.
Bio: Cong Ma is currently a Ph.D. student in the Department of Operations Research and Financial Engineering at Princeton University, advised by Yuxin Chen and Jianqing Fan. His research interests include nonconvex optimization, high-dimensional statistics, machine learning as well as their applications to computational neuroscience.
Abstract: In recent years, there has been an explosion of interest in designing fast nonconvex optimization algorithms to solve statistical estimation and learning problems. However, in contrast to convex optimization that has become a real pillar of modern engineering, the theoretical foundations of nonconvex optimization are far from satisfactory, especially in terms of its computational and inferential properties. This talk will present two recent stories that advance our understanding of nonconvex statistical estimation. The first story focuses on computational efficiency in solving random quadratic systems of equations. Despite the nonconvexity of the natural least-squares formulation, gradient descent with random initialization finds its global solution within a logarithmic number of iterations. The second story is concerned with uncertainty quantification for nonconvex low-rank matrix completion. We develop a de-biased estimator — on the basis of a nonconvex estimator — that enables optimal construction of confidence intervals for the missing entries of the unknown matrix. All of this is achieved via an integrated view of statistics and optimization.
We strongly encourage students to act now to seek full-time employment, internships, and projects (rather than waiting until the end of the semester). Plus, enjoy a free pizza lunch!
Where: ISyE Main Bldg
When: Tuesday, February 4, 2020 | 11am - 2pm
What: The session will include:
Please plan on staying for the duration of the event and bring copies of your resume and business cards. Dress is business casual.
EVENT SPONSOR
The event is sponsored through the generosity and support of JP Morgan Chase & Co. and APICS - Atlanta Chapter. APICS is a non-profit educational organization addressing operations management and supply chain management issues, and providing professional development opportunities to its members. APICS Membership is free for full time students. Join today at www.apics.org/join and start networking at local APIC Atlanta events. Also make sure to stop by the APICS table at the event.
]]>Please RSVP at https://www.surveymonkey.com/r/Z6X2DDG if you plan on attending.
]]>Statistical Learning for Big Dependent Data
Most works in the machine learning and high-dimensional statistical inference assume independence. Most data, on the other hand, are either dynamically or spatially dependent. In this talk, I discuss the impact of dependence on statistical inference of high-dimensional data analysis, including LASSO regression and generalized linear models. The presentation is based on some recent joint papers, focusing on modeling big dependent data for which either the dimension or the sample size or both are large. We demonstrate the analyses using PM2.5 data collected at multiple monitoring stations and at various frequencies.
Ruey S. Tsay is H.G.B. Alexander Professor of Econometrics & Statistics, Booth School
of Business, University of Chicago. He earned his PhD from the University of Wisconsin
- Madison and was with Carnegie Mellon University before joining Chicago in 1989. His
research interest includes nancial econometrics, analysis of high-dimensional dependent
data, forecasting, and time-series analysis. He served as co-editor of the Journal of Busi-
ness and Economic Statistics from 1995 to 1997, Journal of Forecasting from 2006-2013,
and Statistica Sinica from 2014-2017. Currently, he is a co-editor of the Probability and
Statistics Book Series of Wiley. Professor Tsay published widely in leading econometric and statistical journals with more than 115 referred articles. He is the author of Analysis of Financial Time Series (3rd ed., 2010, Wiley), An Introduction to Analysis of Financial Data with R (2013, Wiley), and Multivariate Time Series Analysis (2014, Wiley), and co-author of Nonlinear Time Series Analysis (with R. Chen, 2019, Wiley) and Statistical Learning for Big Dependent Data (with D. Pe~na, 2020, Wiley, forthcoming). He received many honors and awards, including an elected member of Academia Sinica, Taiwan, and a fellow of the American Statistical Association and the Institute of Mathematical Statistics. He also serves on advisory boards of several research institutes and has given invited lectures at the IMF (Head quarter, Washington, DC) and Central Banks of several countries.
Most works in the machine learning and high-dimensional statistical inference assume independence. Most data, on the other hand, are either dynamically or spatially dependent. In this talk, I discuss the impact of dependence on statistical inference of high-dimensional data analysis, including LASSO regression and generalized linear models. The presentation is based on some recent joint papers, focusing on modeling big dependent data for which either the dimension or the sample size or both are large. We demonstrate the analyses using PM2.5 data collected at multiple monitoring stations and at various frequencies.
]]>The Americans with Disabilities Act of 1990 mandates that door-to-door transit options be provided to those who cannot use the regular transit system due to disability. Paratransit agencies operate fleets of vehicles to fulfill daily requests for transportation, which are collected one to seven days ahead of time. Although paratransit is an essential safety net, it is also expensive to operate and requires large government subsidies. These financial difficulties, combined with significant improvements in integer optimization solvers in recent years, have let to interest in developing large-scale optimization algorithms for paratransit. We provide a cluster-then-optimize approach to servicing paratransit requests subject to labor constraints; this approach shows strong performance while also being tractable for daily use. Our case study is based on real data from Boston, MA ranging from 3,000 to 7,000 requests per day, and our algorithms improve upon Boston’s current state by over 30%.
The second part of the talk concerns inference of transit demand data, which is an essential input to any decision model. It is sometimes possible to track anonymized users through their commutes, accomplished through previous studies on smart cards, license plates, and mobile phones. However, widely-available data sources are frequently in aggregated forms such as entry and exit counts, and one must recover the original demand from these aggregated counts. Such problems are generally underspecified. To address this, we present an optimization framework to recover origin-destination matrices under minimal assumptions, incorporating reasonable physical constraints such as flow conservation, smoothness, and symmetry. The proposed method is evaluated and shows strong improvement over the maximum entropy method on a variety of real-world data sets from Boston, New York City, and San Francisco, comprising tens to hundreds of stations.
]]>Abstract:
The Americans with Disabilities Act of 1990 mandates that door-to-door transit options be provided to those who cannot use the regular transit system due to disability. Paratransit agencies operate fleets of vehicles to fulfill daily requests for transportation, which are collected one to seven days ahead of time. Although paratransit is an essential safety net, it is also expensive to operate and requires large government subsidies. These financial difficulties, combined with significant improvements in integer optimization solvers in recent years, have let to interest in developing large-scale optimization algorithms for paratransit. We provide a cluster-then-optimize approach to servicing paratransit requests subject to labor constraints; this approach shows strong performance while also being tractable for daily use. Our case study is based on real data from Boston, MA ranging from 3,000 to 7,000 requests per day, and our algorithms improve upon Boston’s current state by over 30%.
The second part of the talk concerns inference of transit demand data, which is an essential input to any decision model. It is sometimes possible to track anonymized users through their commutes, accomplished through previous studies on smart cards, license plates, and mobile phones. However, widely-available data sources are frequently in aggregated forms such as entry and exit counts, and one must recover the original demand from these aggregated counts. Such problems are generally underspecified. To address this, we present an optimization framework to recover origin-destination matrices under minimal assumptions, incorporating reasonable physical constraints such as flow conservation, smoothness, and symmetry. The proposed method is evaluated and shows strong improvement over the maximum entropy method on a variety of real-world data sets from Boston, New York City, and San Francisco, comprising tens to hundreds of stations.
Bio: Julia Yan is a fifth-year PhD student at the Operations Research Center at MIT, advised by Dimitris Bertsimas. She is interested in large-scale, data-driven optimization, and is especially motivated by applications to urban operations and the public good. Prior to coming to MIT, Julia spent two years in operations consulting. She completed her undergraduate degree at Princeton University in 2013.
]]>
Abstract:
In many areas of medicine, domain knowledge is available in the forms of bio-mechanistic models, human anatomy, and even descriptive statements. Although typically approximate and incomplete, this knowledge represents a wealth of cumulative human intelligence, which can be leveraged and integrated with data-driven learning algorithms for greater efficiency, interpretability, and robustness. My research develops modeling frameworks and associated estimation/inference algorithms to integrate human intelligence and machine intelligence, which is called “knowledge-infused statistical machine learning.” The methodological developments are within the context of using medical imaging and other data to improve the characterization, diagnosis, and treatment of cancer and other diseases.
In this talk, I will introduce several new models and algorithms developed under this theme, driven by the need of improving cancer treatment precision to tackle not only inter- but also intra-tumor heterogeneity. I will present a modeling framework that integrates bio-mechanistic models with MRI and biopsy data to predict the spatial distribution of treatment-informed molecular markers within each tumor. Several extensions of this framework will also be presented, including an algorithm for simultaneous feature and instance selection and a Gaussian process model with knowledge regularization for uncertainty reduction. Furthermore, I will briefly talk about a few other medical domains where knowledge-infusion statistical machine learning has been investigated. I will end the talk by briefly going over my other research efforts and plans.
Bio:
Dr. Jing Li is an Associate Professor in Industrial Engineering & Computer Engineering at Arizona State University (https://www.public.asu.edu/~jli09/). She received her B.S. from Tsinghua University, and an M.A. in Statistics and a Ph.D. in Industrial and Operations Engineering from the University of Michigan. Her research interests are data fusion and statistical machine learning intersecting with health/medical domains having complex data structures. Dr. Li’s research is sponsored by NIH, NSF, DOD, Arizona State, Mayo Clinic, and biomedical industry. She co-founded the ASU-Mayo Clinic Center for Innovative Imaging, conducting various collaborative projects with the Departments of Radiology, Neurology, Neurosurgery, and Radiation Oncology at Mayo Clinic. She is an NSF CAREER awardee, a recipient of a Best Paper Award and a Best Application Paper Award from IISE Transactions, a recipient of the Harold Wolff-John Graham Award (Best Paper) from the American Academy of Neurology, and a recipient of the Harold G. Wolff Lecture Award (Best Paper) by the American Headache Society. She is a former Chair for the Data Mining Subdivision of INFORMS. She is currently the Editor-in-Chief for Quality Technology and Quantitative Management, an Associate Editor for IEEE Transactions on Automation Science and Engineering, an Associate Editor for IISE Transactions on Healthcare Systems Engineering, and on the editorial board of Journal of Quality Technology.
]]>Abstract:
In many areas of medicine, domain knowledge is available in the forms of bio-mechanistic models, human anatomy, and even descriptive statements. Although typically approximate and incomplete, this knowledge represents a wealth of cumulative human intelligence, which can be leveraged and integrated with data-driven learning algorithms for greater efficiency, interpretability, and robustness. My research develops modeling frameworks and associated estimation/inference algorithms to integrate human intelligence and machine intelligence, which is called “knowledge-infused statistical machine learning.” The methodological developments are within the context of using medical imaging and other data to improve the characterization, diagnosis, and treatment of cancer and other diseases.
In this talk, I will introduce several new models and algorithms developed under this theme, driven by the need of improving cancer treatment precision to tackle not only inter- but also intra-tumor heterogeneity. I will present a modeling framework that integrates bio-mechanistic models with MRI and biopsy data to predict the spatial distribution of treatment-informed molecular markers within each tumor. Several extensions of this framework will also be presented, including an algorithm for simultaneous feature and instance selection and a Gaussian process model with knowledge regularization for uncertainty reduction. Furthermore, I will briefly talk about a few other medical domains where knowledge-infusion statistical machine learning has been investigated. I will end the talk by briefly going over my other research efforts and plans.
]]>This accelerated learning boot camp covers 2.5 days of instructor-led facilitation and is designed to cover key learning points of the CSCP learning system. The learning system content is included in the course.
This boot camp is an exam preparatory course for the self-study learner planning to take the APICS CSCP certification exam. The course accelerates learning by focusing on key concepts and terminology covered in the exam. This boot camp, in tandem with the reading materials and online learning courseware, provides a roadmap for those who intend to sit for the exam.
The APICS Certified Supply Chain Professional (CSCP) program takes a broad view of the field, extending beyond internal operations to encompass all the steps throughout the supply chain--from the supplier, through the company, to the end consumer--and how to effectively manage the integration of these activities to maximize the company’s value chain.
The CSCP program provides candidates with the necessary tools to effectively manage global supply chain activities and enables them to implement best practice approaches to increase supply chain efficiencies.
The APICS CSCP is designed for supply chain professionals and people managing and planning extended supply chains especially people in:
Supply Chain Design
Supply Chain Planning and Execution
Supply Chain Planning and Execution
Supply Chain Improvement and Best Practices
Abstract: Simulation-based prediction, for instance in discrete-event analysis and machine learning, relies on models that often are contaminated with errors when calibrating from data. These errors, if overlooked, can result in incorrect inference and underestimation of risks that degrade decision-making. Existing approaches to quantify these errors face several challenges from high computational demand, undercoverage, to the opaqueness in parameter tuning. We present several methods to combat these issues, by injecting subsampling, distributionally robust optimization, and random perturbation respectively into simulation runs. We explain the statistical mechanisms of these approaches and why they help resolve each of the discussed challenges.
Bio: Huajie Qian is a Ph.D. candidate in the department of Industrial Engineering and Operations Research at Columbia University, advised by Henry Lam. His research borrows tools from statistics and machine learning to develop data-driven methodologies for stochastic simulation and optimization that can deal with uncertainties from data in an efficient and principled way. He received his M.S. degree in Applied and Interdisciplinary Mathematics from University of Michigan, and B.S. degree in Mathematics from Fudan University.
]]>Abstract: Simulation-based prediction, for instance in discrete-event analysis and machine learning, relies on models that often are contaminated with errors when calibrating from data. These errors, if overlooked, can result in incorrect inference and underestimation of risks that degrade decision-making. Existing approaches to quantify these errors face several challenges from high computational demand, undercoverage, to the opaqueness in parameter tuning. We present several methods to combat these issues, by injecting subsampling, distributionally robust optimization, and random perturbation respectively into simulation runs. We explain the statistical mechanisms of these approaches and why they help resolve each of the discussed challenges.
]]>SESSION OVERVIEW
The success of on-demand platforms to obtain a ride, e.g., Uber and Lyft, which rely on crowd-sourced transportation capacity, has radically changed the view on the potential and benefits of crowd-sourced transportation and delivery. Many retail stores, for example, are examining the pros and cons of introducing crowd-sourced delivery in their omni-channel strategies. We discuss recent trends in this rapidly evolving area, and highlight challenges and opportunities.
SESSION SPEAKER
Martin Savelsbergh, James C. Edenfield Chair and Professor and Co-Director Supply Chain and Logistics Institute
Attendance to the sessions is complimentary for SCL corporate partners, SCL Industry Advisory Board members, SCL affiliated faculty and students, and students enrolled in the Masters in Supply Chain Engineering program.
To take advantage of the included lunch, you must register by deadline noted on the registration page.
If you have any questions, please email event@scl.gatech.edu.
]]>*To take advantage of the included lunch, you must register by the noted deadlines. If you have any questions, please email event@scl.gatech.edu.
]]>SESSION OVERVIEW
Urban air mobility (UAM) with electric vertical takeoff and landing (eVTOL) aircraft is emerging as a promising aviation market for both cargo delivery and passenger travel. The rise of UAM is being driven by the convergence of two technologies: autonomy and electric aircraft propulsion. Although promising, these technologies place limits on the discovery of viable markets, the timeline of introduction, and the design of operational paradigms. In particular, electric propulsion—when achieved with battery energy storage—is highly constraining in terms of aircraft payload, range, and speed performance and in terms of operational tempo and ground infrastructure. This talk will discuss recent work in modeling aircraft performance and operations for UAM and will highlight the challenges and opportunities in particular urban markets.
SESSION SPEAKER
Brian German, Daniel Guggenheim School of Aerospace Engineering
Attendance to the sessions is complimentary for SCL corporate partners, SCL Industry Advisory Board members, SCL affiliated faculty and students, and students enrolled in the Masters in Supply Chain Engineering program.
To take advantage of the included lunch, you must register by deadline noted on the registration page.
If you have any questions, please email event@scl.gatech.edu.
]]>*To take advantage of the included lunch, you must register by the noted deadlines. If you have any questions, please email event@scl.gatech.edu.
]]>