Gerard Cornuejols
IBM University Professor of Operations Research
Tepper School of Business
Carnegie Mellon University
Abstract
This talk will be based on joint work with Borozan, Basu, Conforti and Zambelli. We extend a theorem of Lovasz characterizing maximal lattice-free convex sets. This result has implications in integer programming. In particular we show a ono-to-one correspondance between these sets and minimal inequalities.
Bio
Gerard Cornuejols was Editor-in-Chief of MOR during the period of 1999-2003 and has been serving in the Advisory Board of MOR since then. He is very famous for his works on Integer Programming, for which he has been awarded several important prizes, namely:
Paul Glasserman
Jack R. Anderson Professor
Graduate School of Business
Columbia University
Abstract
Among the solutions proposed to the problem of banks "too big to fail" is contingent capital in the form of debt that converts to equity when a bank's regulatory capital ratio falls below a threshold. We analyze the dynamics of such a security with continuous conversion and derive closed form expressions for its value when the firm's assets are modeled as geometric Brownian motion and the conversion trigger is an asset-based capital ratio. A key step in the analysis is an explicit formula for the fraction of equity held by the original holders of the contingent capital debt as a function of the maximum drop in asset value. We contrast this analysis with the case of a market-based (rather than accounting-based) conversion trigger.
This is joint work with Bezhad Nouri.
Bio
Paul Glasserman is the Jack R. Anderson Professor of Business at Columbia Business School, where he served as senior vice dean in 2004-2008. Since 2008, he has also been an academic visitor at the Federal Reserve Bank of New York. His research focuses on derivative securities, risk management, stochastic models, and simulation. His publications include the book Monte Carlo Methods in Financial Engineering, which received the 2005 I-Sim Outstanding Simulation Publication Award and the 2006 Lanchester Prize. He also received Risk magazine's 2007 Quant of the Year Award.
]]>Professor
Director of the Computational Science Institute,
Northwestern University
Abstract
We present a "semi-stochastic" Newton method motivated by machine learning problems with very large training sets as well as by the availability of powerful distributed computing environments. The method employs sampled Hessian information to accelerate convergence and enjoys convergence guarantees. We illustrate its performance on multiclass logistic models for the speech recognition system developed at Google. An extension of the method to the sparse L1 setting as well as a complexity analysis will also be presented. This is joint work with Will Neveitt (Google), Richard Byrd (Colorado) and Gillian Chin (Northwestern).
Bio
Jorge Nocedal is a professor in the IEMS and EECS departments at Northwestern University. He obtained a BS from the National University of Mexico and a PhD from Rice University. His research interests are in optimization and scientific computing, and in their application to machine learning, computer-aided design and financial engineering. He is the author (with Steve Wright) of the book Numerical Optimization.
He is a SIAM Fellow, an ISI Highly Cited Researcher (Mathematics Category), and was an invited speaker at the 1998 International Congress of Mathematicians. He serves in the editorial board of Mathematical Programming, and in 2011 he will become editor-in-chief of SIAM Journal on Optimization. In 1998 he was appointed Bette and Neison Harris Professor of Teaching Excellence at Northwestern.
]]>Edward H. Kaplan
William N and Marie A Professor of Management Sciences,
Yale School of Management;
Professor of Public Health,
Yale School of Public Health;
Professor of Engineering,
Yale School of Engineering and Applied Science
Abstract
This article presents the first model developed specifically for understanding the infiltration and interdiction of ongoing terror plots by undercover intelligence agents, and does so via novel application of ideas from queueing theory and Markov population processes. The resulting "terror queue" models predict the number of undetected terror threats in an area from agent activity/utilization data, and also estimate the rate with which such threats can be interdicted. The models treat terror plots as customers and intelligence agents as servers. Agents spend all of their time either detecting and infiltrating new terror plots (in which case they are "available"), or interdicting already detected terror plots (in which case they are "busy"). Initially we examine a Markov model assuming that intelligence agents, while unable to detect all plots, never err by falsely detecting fake plots. While this model can be solved numerically, a simpler Ornstein-Uhlenbeck diffusion approximation yields some results in closed form while providing nearly identical numerical performance. The transient behavior of the terror queue model is discussed briefly along with a sample sensitivity analysis to study how model predictions compare to simulated results when using estimated versus known terror plot arrival rates. The diffusion model is then extended to allow for the false detection of fake plots. Such false detection is a real feature of counterterror intelligence given that intelligence agents or informants can make mistakes, as well as the proclivity of terrorists to deliberately broadcast false information. The false detection model is illustrated using suicide bombing data from Israel.
Bio
Professor Kaplan's research has been reported on the front pages of the New York Times and the Jerusalem Post, editorialized in the Wall Street Journal (article), recognized by the New York Times Magazine's Year in Ideas, and discussed between the covers of Time (article), Newsweek (see article), US News and World Report, Consumer Reports and the New Yorker, and in person on NBC's Today Show, the Cronkite Report, and National Public Radio
(transcript).
The author of more than 100 research articles, Professor Kaplan received both the Lanchester Prize and theEdelman Award, the two top honors in the operations research field. An elected member of both the National Academy of Engineering and the Institute of Medicine of the US National Academies, he has also twice received the prestigious Lady Davis Visiting Professorship at the Hebrew University of Jerusalem, where he has investigated AIDS policy issues facing the State of Israel.
Kaplan's current research focuses on the application of operations research to problems in counterterrorism and
homeland security.
Boeing Leaders for Global Operations Professor
Operations Research/Statistics
Sloan School of Management
Abstract
In this presentation, we show a significant role that symmetry, a fundamental concept in convex geometry, plays in determining the power of robust and finitely adaptable solutions in multi-stage stochastic and adaptive optimization problems. We consider a fairly general class of multi-stage mixed integer stochastic and adaptive optimization problems and propose a good approximate solution policy with performance guarantees that depend on the geometric properties such as symmetry of the uncertainty sets. In particular, we show that a class of finitely adaptable solutions is a good approximation for both the multi-stage stochastic as well as the adaptive optimization problem. A finitely adaptable solution specifies a small set of solutions for each stage and the solution policy implements the best solution from the given set depending on the realization of the uncertain parameters in the past stages. To the best of our knowledge, these are the first approximation results for the multi-stage problem in such generality.
(joint work with Vineet Goyal, Columbia University and Andy Sun, MIT)
Bio
Dimitris Bertsimas is currently the Boeing Professor of Operations Research and the codirector of the Operations Research Center at the Massachusetts Institute of Technology. He has received a BS in Electrical Engineering and Computer Science at the National Technical University of Athens, Greece in 1985, a MS in Operations Research at MIT in 1987, and a Ph.D in Applied Mathematics and Operations research at MIT in 1988. Since 1988, he has been in the MIT faculty.
His research interests include optimization, stochastic systems, data mining, and their application. He has co-authored more than 120 scientific papers and he has co-authored the following books: ``Introduction to Linear Optimization'' (with J. Tsitsiklis, Athena Scientific and Dynamic Ideas, 2008), ``Data, models and decisions'' (with R. Freund, Dynamic Ideas, 2004) and ``Optimization over Integers'' (with R. Weismantel, Dynamic Ideas, 2005). He is currently department editor in Optimization for Management Science and former area editor in Operations Research in Financial Engineering. He has supervised 42 doctoral students and he is currently supervising 10 others.
He is a member of the National Academy of Engineering, and he has received several research awards including: the Farkas prize (2008), the Erlang prize (1996), the SIAM prize in optimization (1996), the Bodossaki prize (1998) and the Presidential Young Investigator award (1991-1996).
]]>Sheldon Ross
Epstein Chair Professor
Industrial and Systems Engineering
University of Southern California
Abstract
Suppose there are r gamblers, with gambler i initially having a fortune of ni. In our first model we suppose that at each stage two of the gamblers are chosen to play a game, equally likely to be won by either player, with the winner of the game receiving 1 from the loser. Any gambler whose fortune becomes 0 leaves, and this continues until there is only a single gambler left. We are interested in the probability that player i is the one left, and in the the mean number of games played between specified players i and j. In our second model we suppose that all remaining players contribute 1 to a pot, which is equally likely to be won by each of them. The problem here is to determine the expected number of games played until one player has all the funds.
]]>Egon Balas
University Professor of Industrial Administration and Applied Mathematics
The Thomas Lord Professor of Operations Research
Carnegie Mellon University
Abstract
Intersection cuts are generated from a polyhedral cone and a convex set S whose interior contains no feasible integer point. We generalize these cuts by replacing the cone with a more general polyhedron C. The resulting generalized intersection cuts dominate the original ones. This leads to a new cutting plane paradigm under which one generates and stores the intersection points of the extreme rays of C with the boundary of S rather than the cuts themselves. These intersection points can then be used to generate deeper cuts in a non-recursive fashion.
(This talk is based on joint work with Francois Margot)
Bio
Egon Balas is University Professor of Industrial Administration and Applied Mathematics, as well as the Thomas Lord Professor of Operations Research, at Carnegie Mellon University. He has a doctorate in Economic Science from the University of Brussels and a doctorate in Mathematics from the University of Paris.
Professor Balas's research interests are in mathematical programming, primarily integer and combinatorial optimization. He has played a leading role in the developmant of enumerative and cutting plane techniques for 0-1 programming, and is mainly known as the developer of the approach called disjunctive programming or lift-and-project. He has also developed scheduling algorithms and software. Dr. Balas has served or is serving on the editorial boards of Operations Research, Discrete Applied Mathematics, the Journal of Combinatorial Optimization, Computational Optimization and Applications, the European Journal of Operational Research, Annals of Operations Research etc. In 1980 Dr. Balas received the US Senior Scientist Award of the Alexander von Humboldt Foundation; in 1995 he received the John von Neumann Theory Prize of INFORMS; and in 2001 he was the first American to be awarded the EURO Gold Medal of the European Association of Operational Research Societies.
]]>Yurii Nesterov
CORE/INMA,
Catholic University og Louvain (UCL) Belgium
Abstract
In this talk we describe the new methods for solving huge-scale optimization problems. For problems of this size, even the simplest full-dimensional vector operations are very expensive. Hence, we suggest to apply an optimization technique based on random partial update of decision variables. For these methods, we prove the global estimates for the rate of convergence. Surprisingly enough, for certain classes of objective functions, our results are better than the standard worst-case bounds for deterministic algorithms. We present constrained and unconstrained versions of the method, and its accelerated variant. Our numerical test confirms a high efficiency of this technique on problems of very big size.
The paper can be downloaded as CORE Discussion Paper 2010/2:
http://www.uclouvain.be/en-310431.html
]]>Costis Maglaras
David and Lyn Silfen Professor of Business
Decision, Risk and Operations Division,
Graduate School of Business,
Columbia University
Abstract
The first part of the talk will offer a brief overview of algorithmic trading in the US equities market, touching, in passing, upon on the topic of high-frequency trading and the nature of current market structure. The emphasis will be on highlighting different facets of this area and discussing corresponding quantitative research problems. The second half of the talk describes a queueing model of limit order book dynamics, and explores questions of optimal limit order placement, market impact, and optimal trade execution.
Bio
Costis Maglaras is the David and Lyn Silfen Professor of Business at Columbia University. His research focuses on quantitative pricing and revenue management, the economics, design and operations of service systems, and financial engineering. He is the author of many research articles spanning the theory and application of stochastic modeling in a variety of fields, more recently in pricing, risk management and valuation of multi-unit real estate portfolios, and in the design of portfolio trading systems and algorithms. He holds editorial positions in many of the flagship journals of his fields of study, he is the recipient of several research and teaching awards, and he teaches and serves as faculty director for the executive education course on Risk Management offered by Columbia Business School.
]]>Adrian Lewis
Cornell University
Abstract
Concrete optimization problems, while often nonsmooth, are not pathologically so. The class of "semi-algebraic" sets and functions - those arising from polynomial inequalities - nicely exemplifies nonsmoothness in practice. Semi-algebraic sets (and their generalizations) are common, easy to recognize, and richly structured, supporting powerful variational properties. In particular I will discuss a generic property of such sets - partial smoothness - and its relationship with a proximal algorithm for nonsmooth composite minimization, a versatile model for practical optimization.
Bio
Adrian S. Lewis was born in England in 1962. He is a Professor at Cornell University in the School of Operations Research and Industrial Engineering. Following his B.A., M.A., and Ph.D. degrees from Cambridge, and Research Fellowships at Queens' College, Cambridge and Dalhousie University, Canada, he worked in Canada at the University of Waterloo (1989-2001) and Simon Fraser University (2001-2004). He is an Associate Editor of the SIAM Journal on Optimization, Mathematics of Operations Research, and the SIAM/MPS Book Series on Optimization, and is a Co-Editor for Mathematical Programming. He received the 1995 Aisenstadt Prize, from the Canadian Centre de Recherches Mathematiques, the 2003 Lagrange Prize for Continuous Optimization from SIAM and the Mathematical Programming Society, and an Outstanding Paper Award from SIAM in 2005. He co-authored "Convex Analysis and Nonlinear Optimization" with J.M. Borwein.
Lewis' research concerns variational analysis and nonsmooth optimization, with a particular interest in optimization problems involving eigenvalues.
]]>Yinyu Ye
Professor of Management Science and Engineering
and, by courtesy, Electrical Engineering
Affiliation: Department of Management Science and Engineering
Stanford University
Abstract
A natural optimization model that formulates many online resource allocation and revenue management problems is the online linear program (LP) where the constraint matrix is revealed column by column along with the objective function. We provide a near-optimal algorithm for this surprisingly general class of online problems under the assumption of random order of arrival and some mild conditions on the size of the LP right-hand-side input. Our learning-based algorithm works by dynamically updating a threshold price vector at geometric time intervals, where the dual prices learned from revealed columns in the previous period are used to determine the sequential decisions in the current period. Our algorithm has a feature of learning by doing", and the prices are updated at a carefully chosen pace that is neither too fast nor too slow. In particular, our algorithm doesn't assume any distribution information on the input itself, thus is robust to data uncertainty and variations due to its dynamic learning capability. Applications of our algorithm include many online multi-resource allocation and multi-product revenue management problems such as online routing and packing, online combinatorial auctions, adwords matching, inventory control and yield management.
This is a joint work with Shipra Agrawal and Zizhuo Wang.
Bio
Yinyu Ye received the B.S. degree in System Engineering from the Huazhong University of Science and Technology, Wuhan, China, and the M.S. and Ph.D. degrees in Management Science & Engineering from Stanford University, Stanford. Currently, he is a full Professor of Management Science and Engineering and Institute of Computational and Mathematical Engineering and the Director of the MS&E Industrial Affiliates Program, Stanford University. His current research interests include Continuous and Discrete Optimization, Mathematical Programming, Algorithm Design and Analysis, Computational Game/Market Equilibrium, Metric Distance Geometry, Graph Realization, Dynamic Resource Allocation, and Stochastic and Robust Decision Making, etc.
The following is a list of some of his main achievements:
Mark S. Squillante,
IBM Research
Abstract
In this talk we present an integrated suite of operations research models and methods that supports the effective and efficient management and planning of human capital supply chains by addressing distinct features and characteristics of human talent and skills. This consists of solutions for: (1) the statistical forecasting of future demand and resource requirements; (2) a new form of risk-based stochastic resource capacity planning; (3) the stochastic modeling and optimization/control of supply evolutionary dynamics over time; (4) a new form of optimal multi-skill supply-demand matching; and (5) the stochastic optimization of business decisions to manage resource shortages and overages. These solutions include contributions in the areas of stochastic models and stochastic optimization/control. The suite of models and methods constitutes an end-to-end solution that is deployed as an important part of the human capital management and planning process within IBM.
Bio
Mark S. Squillante is a Research Staff Member in the Mathematical Sciences Department at the IBM Thomas J. Watson Research Center, where he leads the Applied Probability and Stochastic Optimization team. His research interests concern mathematical foundations of the analysis, modeling and optimization of the design and control of stochastic systems, including stochastic processes, applied probability, stochastic optimization and control, and their applications. He is the author of many research articles across these areas, and has received several internal (IBM) and external research awards. He is a Fellow of ACM and IEEE, and currently serves on the editorial boards of Operations Research, Stochastic Models and Performance Evaluation.
]]>SPEAKER: Peter J. Haas, IBM Research
ABSTRACT:
The generalized semi-Markov process (GSMP) is the usual model forthe underlying stochastic process of a complex discrete-eventsystem. It is important to understand fundamental behavioralproperties of the GSMP model, such as the conditions under whichthe states of a GSMP are recurrent. For example, recurrence isnecessary for the validity of steady-state simulation outputanalysis methods such as the regenerative method, spectralmethod, and the method of batch means. We review some sufficientconditions for recurrence in irreducible finite-stateGSMPs. These conditions include requirements on the "clocks" thatgovern the occurrence times of state transitions. For example,each clock-setting distribution must have finite mean. We thenshow that, in contrast to ordinary semi-Markov processes, an
irreducible finite-state GSMP can have transient states in thepresence of multiple clock-setting distributions with heavytails. (Joint work with Peter Glynn.)
SPEAKER: Margaret L. Brandeau
ABSTRACT:
What is the most cost-effective way to uselimitedHIVprevention and treatment resources? Howshould the Centers for Disease Control and Prevention revise nationalimmunization recommendations so that gaps in vaccination coverage willbefilled in a cost-effective manner? Towhat extent should local communities stockpile antibiotics for responseto apotential bioterror attack? How canhumanitarian relief organizations manage their inventories mosteffectively? This talk will describeexamples from past and ongoing model-based analyses of public healthpolicyquestions. We also provide perspectiveson key elements of a successful policy analysis and discuss ways inwhich suchanalysis can influence policy.
About the Speaker
MargaretBrandeau isProfessor of Management Science and Engineering and Professor ofMedicine (byCourtesy) at Stanford University. Her research focuses on the development ofapplied mathematical and economic models to support health policydecisions. Her recent work has focusedon HIV prevention and treatment programs, programs to control thespread of HepatitisB virus, and evaluating preparedness plans for bioterror response. She is a Fellow of the Institute forOperations Research and Management Science (INFORMS), and has receivedthePresident’s Award from INFORMS (recognizing important contributions tothewelfare of society), the Pierskalla Prize from INFORMS (for researchexcellencein health care management science), and a Presidential YoungInvestigator Awardfrom the National Science Foundation, among other awards. Professor Brandeau earned a BS in Mathematicsand an MS in Operations Research from MIT, and a PhD inEngineering-EconomicSystems from Stanford University.
]]>SPEAKER: Professor Joseph Wu
ABSTRACT:
Early estimates of transmissibility and severity of an emerginginfluenza pandemic is an urgent public health priority. This ischallenging because many influenza infections are subclinical.
Population-based serologic surveillance allows accurate estimates ofinfection attack rates (IAR) and severity estimates. During 2009, wetested more than 17,800 serological specimens in Hong Kong throughoutthe first wave of the H1N1 pandemic. Using these data we estimated
that the basic reproductive number was 1.38 and that 0.6% ofinfections led to hospitalization. We developed a novel statisticalmethod for real-time serologic surveillance and estimated that about1,000 specimens per week would allow accurate estimates of IAR andseverity as soon as the true IAR has reached 2%. Serologic monitoringshould be considered in updated pandemic plans.
SPEAKER: Paul Glasserman
ABSTRACT:
Among the solutions proposed to the problem of banks "too big to fail" is contingent capital in the form of debt that converts to equity when a bank's regulatory capital ratio falls below a threshold. We analyze the dynamics of such a security with continuous conversion and derive closed form expressions for its value when the firm's assets are modeled as geometric Brownian motion and the conversion trigger is an asset-based capital ratio. A key step in the analysis is an explicit formula for the fraction of equity held by the original holders of the contingent capital debt as a function of the maximum drop in asset value. We contrast this analysis with the case of a market-based (rather than accounting-based) conversion trigger. This is joint work with Bezhad Nouri.
Paul Glasserman is the Jack R. Anderson Professor of Business at Columbia Business School, where he served as senior vice dean in 2004-2008. Since 2008, he has also been an academic visitor at the Federal Reserve Bank of New York. His research focuses on derivative securities, risk management, stochastic models, and simulation. His publications include the book Monte Carlo Methods in Financial Engineering, which received the 2005 I-Sim Outstanding Simulation Publication Award and the 2006 Lanchester Prize. He also received Risk magazine's 2007 Quant of the Year Award.
At the invitation of the Georgia Tech Center for Health and Humanitarian Logistics, Dr. John Hardman, president and CEO of the Carter Center, and François Grünewald, director of URD (Urgency, Recovery and Development), will discuss the situation in Haiti and what still needs to be done to ensure recovery.
This lecture will be moderated by Reginald DesRoches, professor and associate chair of Civil & Environmental Engineering and key technical leader in the response to the Haiti Earthquake.
It will be followed by the screening of a series of documentaries on Haiti.
To register: http://www.france-atlanta.org/spip.php?article50
For a complete listing of events taking place on the Georgia Tech campus, visit http://www.global.gatech.edu/france-atlanta/. For other lectures,workshops, conferences, and symposiums associated with the full France-Atlanta collaboration, visit http://www.france-atlanta.org/
]]>SPEAKER: Andrew Johnson
ABSTRACT:
The Finnish electricity markethas a competitive energy generation market and a monopolistic transmissionsystem. To regulate the local monopoly power of network operators, thegovernment regulator uses frontier estimation methods (e.g., StochasticFrontier Analysis (SFA) and nonparametric Data Envelopment Analysis (DEA)) toidentify excessive transmission costs, taking into account outputs and theoperating environment. We describe the new regulatory system developed for theFinnish regulator, which is based on the method Stochastic Non-smoothEnvelopment of Data (StoNED) and utilizes panel data to detect the excessivecosts from random noise.
The literature ofproductive efficiency analysis is divided into two main branches: the parametricSFA and nonparametric DEA. StoNED is a new frontier estimation framework thatcombines the virtues of both DEA and SFA in a unified approach to frontieranalysis. StoNED follows the SFA approach by including a stochastic component. Incontrast to SFA, however, the proposed method does not make any priorassumptions about the functional form of the production function. In thatrespect, StoNED is similar to DEA, and only imposes free disposability,convexity, and some returns to scale specification.
The main advantage ofthe StoNED approach to the parametric SFA approach is the independence of the adhoc parametric assumptions about the functional form of the production function(or cost/distance functions). In contrast to the flexible functional forms, onecan impose monotonicity, concavity and homogeneity constraints withoutsacrificing the flexibility of the regression function. Additionally, the mainadvantage of StoNED to the nonparametric DEA approach is robustness tooutliers, data errors, and other stochastic noise in the data. In DEA thefrontier is spanned by a relatively small number of efficient firms, however,in our method all observations influence the shape of the frontier. Also manystandard tools from parametric regression such as goodness of fit statisticsand statistical tests are directly applicable in our approach. This is collaboratework with Timo Kuosmanen of Aalto University in Finland.
Andrew L Johnson is an AssistantProfessor in the Department of Industrial and Systems Engineering at TexasA&M University. He obtained his B.S. in Industrial and Systems Engineeringfrom Virginia Tech and his M.S. and Ph.D. from the H. Milton Stewart School ofIndustrial and Systems Engineering from Georgia Tech. His research interestsinclude productivity and efficiency measurement, warehouse design andoperations, material handling and mechanism design. He is a member of theINFORMS, National Eagle Scout Association, and German Club of Virginia Tech.
]]>SPEAKER: Rustam Ibragimov
ABSTRACT:
Many risk, inequality, poverty and concentration measures are extremely sensitive to outliers, dependence, heterogeneity and heavy tails. In this paper we focus on robust measurement of risk, inequality, poverty and concentration under heterogeneity, dependence and heavy-tailednessof largely unknown form using the recent results on t-statistic based heterogeneity and correlation robust inference in Ibragimov and Muller (2007). The robust large sample inference on risk, inequality, poverty and concentration measures is conducted as follows: partition the observations into q>=2 groups, calculate the empirical measures for each group and conduct a standard test with the resulting q estimators of the population measures.
Numerical results confirm the appealingproperties of tstatistic based robust inference method in this context, and its applicability to many widely used risk, inequality, poverty and concentration measures, including Sharpe ratio; value at risk and expected shortfall; Gini coecient; Theil index, mean logarithmic deviation and generalized entropy measures; Atkinson measures; coecient of variation and Herfindahl-Hirschman index; head count, poverty gap and squared poverty gap indices and other Foster-Greer-Thorbecke measures of poverty, among others. The results discussed in the paper further indicate a strong linkbetween the tstatistic based robust inference methods and stochasticanalogues of the majorization conditions that are usually imposed on risk, inequality, poverty and concentration measures related to
self-normalized sums or their transforms, as in the case of Sharpe ratio,coefficient of variation and Herndahl-Hirschman index.
SPEAKER: Ben Haaland
ABSTRACT:
A multistep procedure is introduced to statisticians for modeling large-scale computer experiments. In practice, the procedure shows substantial improvements in overall accuracy. We introduce the terms nominal and numeric error and decompose the overall error of an emulator into nominal and numeric portions. For the multistep procedure, we develop bounds on the numeric and nominal error. These bounds show that substantial gains in overall accuracy can be attained with the multistep approach.
]]>SPEAKER: Jose Blanchet
ABSTRACT:
Reflected stochastic networks arise in the analysis of a large class of queueing systems. The most popular model of this type is perhaps reflected Brownian motion, which arises in the heavy-traffic analysis of generalized Jackson networks. In this talk we discuss Monte Carlo
simulation strategies for the steady-state analysis of reflected stochastic networks. In particular, we show how to exactly (i.e. without bias) simulate the equilibrium distribution of a reflected stochastic network with compound Poisson input and how to provide samples that are
close (with explicit and controlled error bounds) to both the transient and the steady-state distribution of reflected Brownian motion in the positive orthant. (Joint work with Xinyun Chen.)
SPEAKER: Mark Squillante
ABSTRACT:
In this talk we present an integrated suite of operations research models and methods that supports the effective and efficient management and planning of human capital supply chains by addressing distinct features and characteristics of human talent and skills. This consists of solutions for: (1) the statistical forecasting of future demand and resource requirements; (2) a new form of risk-based stochastic resource capacity planning; (3) the stochastic modeling and optimization/control of supply evolutionary dynamics over time; (4) a new form of optimal multi-skill supply-demand matching; and (5) the stochastic optimization of business decisions to manage resource shortages and overages. These solutions include contributions in the areas of stochastic models and stochastic optimization/control. The suite of models and methods constitutes an end-to-end solution that is deployed as an important part of the human capital management and planning process within IBM.
Mark S. Squillante is a Research Staff Member in the Mathematical Sciences Department at the IBM Thomas J. Watson Research Center, where he leads the Applied Probability and Stochastic Optimization team. His research interests concern mathematical foundations of the analysis, modeling and optimization of the design and control of stochastic systems, including stochastic processes, applied probability, stochastic optimization and control, and their applications. He is the author of many research articles across these areas, and has received several internal (IBM) and external research awards. He is a Fellow of ACM and IEEE, and currently serves on the editorial boards of Operations Research, Stochastic Models and Performance Evaluation.
SPEAKER: Hans-Georg Mueller
ABSTRACT:
Functional regression has emerged as a useful approach for the analysis ofcomplex data that combine functional or longitudinal predictors with scalaror functional responses. A major emphasis has been the functional linearregression model, which allows to implement dimension reduction in a simpleand straightforward way but may be too restrictive. We will discuss flexibleextensions of this model. These include functional quadratic, polynomial and
additive models. Of special interest is differentiation with respect to afunctional argument, for which additive models are particularly well suited.Another extension are local models, where the focus is on the dependency ofa Gaussian process or its derivatives at a given time on the value of apredictor process at the same or a different time. The methods will beillustrated with densely as well as sparsely sampled functional data. Thistalk is based on joint work with Wenjing Yang and Fang Yao.
SPEAKER: Costis Maglaras
ABSTRACT:
The first part of the talk will offer a brief overview of algorithmic trading in the US equities market, touching, in passing, upon on the topic of high-frequency trading and the nature of current market structure. The emphasis will be on highlighting different facets of this area and discussing corresponding quantitative research problems. The second half of the talk describes a queueing model of limit order book dynamics, and explores questions of optimal limit order placement, market impact, and optimal trade execution.
Bio:
Costis Maglaras is the David and Lyn Silfen Professor of Business at Columbia University. His research focuses on quantitative pricing and revenue management, the economics, design and operations of service systems, and financial engineering. He is the author of many research articles spanning the theory and application of stochastic modeling in a variety of fields, more recently in pricing, risk management and valuation of multi-unit real estate portfolios, and in the design of portfolio trading systems and algorithms. He holds editorial positions in many of the flagship journals of his fields of study, he is the recipient of several research and teaching awards, and he teaches and serves as faculty director for the executive education course on Risk Management offered by Columbia Business School.
SPEAKER: Jorge Nocedal
ABSTRACT:
We present a "semi-stochastic" Newton method motivated by machine learning problems with very large training sets as well as by the availability of powerful distributed computing environments. The method employs sampled Hessian information to accelerate convergence and enjoys convergence guarantees. We illustrate its performance on multiclass logistic models for the speech recognition system developed at Google. An extension of the method to the sparse L1 setting as well as a complexity analysis will also be presented. This is joint work with Will Neveitt (Google), Richard Byrd (Colorado) and Gillian Chin (Northwestern).
Short Bio:
Jorge Nocedal is a professor in the IEMS and EECS departments at Northwestern University. He obtained a BS from the National University of Mexico and a PhD from Rice University. His research interests are in optimization and scientific computing, and in their application to machine learning, computer-aided design and financial engineering. He is the author (with Steve Wright) of the book "Numerical Optimization."
He is a SIAM Fellow, an ISI Highly Cited Researcher (Mathematics Category), and was an invited speaker at the 1998 International Congress of Mathematicians. He serves in the editorial board of Mathematical Programming, and in 2011 he will become editor-in-chief of SIAM Journal on Optimization. In 1998 he was appointed Bette and Neison Harris Professor of Teaching Excellence at Northwestern.
SPEAKER: Yajun Mei
ABSTRACT:
In the modern information age one often monitors a large number of data streams with the aim of offering the potential for early detection of a "trigger" event. In this talk, we are interested in detecting the event as soon as possible, but we do not know when the event will occur, nor do we know which subset of data streams will be affected by the event. Motivated by the applications in censoring sensor networks and by the case when one has a prior knowledge that at most r data streams will be affected, we propose scalable global monitoring schemes based on the sum of the local detection statistics that are "large" under either hard thresholding or top-r thresholding rules or both. The proposed schemes are shown to possess certain asymptotic optimality properties.
]]>SPEAKER: Sofia Olhede
ABSTRACT:
High angular resolution diffusion imaging data is the observedcharacteristic function for the local diffusion of water molecules intissue. This data is used to infer structural information in brainimaging. Non-parametric scalar measures are proposed to summarizesuch data, and to locally characterize spatial features of thediffusion probability density function (PDF), relying on the geometryof the characteristic function. Summary statistics are defined sothat their distributions are, to first order, both independent ofnuisance parameters and analytically tractable. The dominantdirection of the diffusion at a spatial location (voxel) isdetermined, and a new set of axes are introduced in Fourier space.Variation quantified in these axes determines the local spatialproperties of the diffusion density. Non-parametric hypothesis testsfor determining whether the diffusion is unimodal, isotropic ormulti-modal are proposed. More subtle characteristics of white-mattermicrostructure, such as the degree of anisotropy of the PDF andsymmetry compared with a variety of asymmetric PDF alternatives, may
be ascertained directly in the Fourier domain without parametricassumptions on the form of the diffusion~PDF. We simulate a set ofdiffusion processes and characterize their local properties using thenewly introduced summaries. We show how complex white-matter
structures across multiple voxels exhibit clear ellipsoidal andasymmetric structure in simulation, and assess the performance of thestatistics in clinically-acquired magnetic resonance imaging data. Joint work with Brandon Whitcher, GSK.
BIO: Sofia C. Olhede was awarded the M.Sci. and Ph.D. degrees inmathematics from Imperial College London, London, U.K., in 2000 and 2003,respectively. She was a Lecturer (2002�2006) and Senior Lecturer(2006�2007) with the Mathematics Department, Imperial College London. In2007, she joined the Department of Statistical Science, University College
London, where she is Pearson Professor of Statistics and HonoraryProfessor of Computer Science. Her research interests include the analysisof nonstationary time series, inhomogeneous random fields and applicationsin geoscience, medical imaging and oceanography. Prof. Olhede is anAssociate Editor of the Journal of the Royal Statistical Society, Series B(Statistical Methodology) and of the IEEE Transactions on Signal
Processing. She is a member of the Programme Committee of theInternationalCentre for Mathematical Sciences, and is an Isaac Newton InstituteCorrespondent.
SPEAKER: Ming Yuan
ABSTRACT:
More and more often in practice, one needs to estimate a highdimensional covariance matrix. In this talk, we discuss how this taskis often related to the sparsity of the inverse covariance matrix. Inparticular, we consider estimating a (inverse) covariance matrix thatcan be well approximated by ``sparse'' matrices. Taking advantage ofthe connection between multivariate linear regression and entries ofthe inverse covariance matrix, we introduce an estimating procedurethat can effectively exploit such ``sparsity''. The proposed methodcan be computed using linear programming and therefore has thepotential to be used in very high dimensional problems. Oracleinequalities are established for the estimation error in terms ofseveral operator norms, showing that the method is adaptive todifferent types of sparsity of the problem.
]]>SPEAKER: Apostolos Fertis
ABSTRACT:
Coherent risks can be expressed as the worst-case expectation when the probability distribution varies in some uncertainty set, according to the representation theorem. Very often, randomness can be divided in two stages, and there is additional information about the possible first stage scenarios. Traditional coherent risks, such as the CVaR, fail to make use of this information. In this talk, we introduce a new class of risk measures, called robust risk measures, which combine the uncertainty set of a traditional risk measure with the additional information about the first stage scenarios. We state and prove a representation theorem for the robust risk measures, which facilitates their computation. We define and show how to compute the Robust CVaR, the robust risk constructed based on CVaR. We compare the optimal-Robust CVaR and optimal-CVaR portfolios under diverse scenarios constructed using real New York Stock Exchange (NYSE) and NASDAQ data from 2005 to 2010.
Bio:
Apostolos Fertis completed his PhD at the Electrical Engineering and Computer Science Department of the Massachusetts Institute of Technology in 2009. Currently he is a researcher at the Institute for Operations Research (IFOR) in Zurich.
In PhD thesis, under the supervision of Professor Dimitris Bertsimas, he investigated the application of the robust optimization concept in confronting the uncertainty in the samples used to produce statistical estimates. In January 2010, he initiated the "Robust Risk Management" research project at theIFOR. The project aspires to introduce a new idea in uncertainty management by combining traditional risk management techniques with robust optimization.
SPEAKER: Jose Blanchet
ABSTRACT:
A stochastic perpetuity is the net present value, with i.i.d. random discount factors, of an infinite stream of i.i.d. rewards in time. Under reasonable assumptions on the rewards and discounts we describe how to generate exact (unbiased) samples of stochastic perpetuities. The algorithm is based on a variation of dominated coupling from the past. The dominating process involves exact sampling of the delay sequence of a single server queue starting from the distant past. (This is joint work with K. Sigman.)
]]>SPEAKER: Egon Balas
ABSTRACT:
Intersection cuts are generated from a polyhedral cone and a convex setS whose interior contains no feasible integer point. We generalizethese cuts by replacing the cone with a more general polyhedron C. The resulting generalized intersection cuts dominate the original ones.This leads to a new cutting plane paradigm under which one generatesand stores the intersection points of the extreme rays of C with theboundary of S rather than the cuts themselves. These intersectionpoints can then be used to generate deeper cuts in a non-recursivefashion.
(This talk is based on joint work with Francois Margot)
Bio:
Egon Balas is University Professor of Industrial Administration andApplied Mathematics, aswell as the Thomas Lord Professor of Operations Research, at CarnegieMellonUniversity. He has a doctorate in Economic Science from the Universityof Brusselsand a doctorate in Mathematics from the University of Paris.
ProfessorBalas's researchinterests are in mathematical programming, primarily integer andcombinatorialoptimization. He has played a leading role in the developmant ofenumerativeand cutting plane techniques for 0-1 programming, and is mainly knownas thedeveloper of the approach called disjunctive programming orlift-and-project.He has also developed scheduling algorithms and software. Dr. Balas hasservedor is serving on the editorial boards of Operations Research, DiscreteAppliedMathematics, the Journal of Combinatorial Optimization, ComputationalOptimization and Applications, the European Journal of OperationalResearch,Annals of Operations Research etc. In 1980 Dr. Balas received the USSeniorScientist Award of the Alexander von Humboldt Foundation; in 1995 hereceivedthe John von Neumann Theory Prize of INFORMS; and in 2001 he was thefirstAmerican to be awarded the EURO Gold Medal of the European AssociationofOperational Research Societies.
]]>
SPEAKER: Dr. Michal Tzur
ABSTRACT:
The multi-locationreplenishment and transshipment problem is concerned with several retailersfacing random demand for the same item at distinct markets, that may usetransshipments to eliminate excess inventory/shortages after demandrealization. For a centralized system we summarize recent advances related tooperational and design aspects of the problem. Our main focus in this talk ison the decentralized system, in which the retailers operate to maximize theirown profit. This causes incentive problems that prevent coordination, even withtwo retailers who may pay each other for transshipped units. We propose a newmechanism based on a transshipment fund which is the first to coordinate thesystem, in a fully non-cooperative setting, for all instances of two retailersas well as all instances of any number of retailers. The computation andinformation requirements of this mechanism are realistic and relatively modest.We also present necessary and sufficient conditions for coordination and provethey are always satisfied with our mechanism.
Bio:
Michal Tzur is an Associate Professor in the Industrial Engineeringdepartment at the Faculty of Engineering at Tel Aviv University in Israel,where she served as the head of theundergraduate program and as the department chair. Michal joined TelAviv University after spending three years at the Operations and InformationManagement department at Wharton School at the University of Pennsylvania. Shereceived her B.A. from Tel Aviv University and her M. Phil and Ph.D. from Columbia University. During the years 2002-2004she visited the IEMS department at Northwestern University. Her areas ofexpertise are supply chain management, inventory routing and transportation.
]]>SPEAKER: Professor Peter Bickel
ABSTRACT:
A great deal of attention has recentlybeen paid to determining sub-communities on the basis of relations,corresponding to edges, between individuals, corresponding tovertices of an unlabelled graph (Newman, SIAM Review 2003;Airoldi et al JMLR 2008; Leskovec & Kleinberg et al SIGKDD 2005). We have developed a nonparametricframework for probabilistic ergodic models of infinite unlabelled graphs (PNAS2009)andmade some connections with modularities arising in the physics literature and community models in the social sciences.Afundamental difficulty in implementing these procedures is computational complexity.We develop approaches which bypass these difficulties.
(This is joint work with Aiyou Chen and LizaLevina)
]]>SPEAKER: Julie Simmons Ivy, Associate Professor, Edward P. Fitts Department of Industrial and Systems Engineering, North Carolina State University
ABSTRACT:
Influenza pandemics are consideredone of the most significant and widely spread threats to public health. In thisresearch, we explore the relationship between local and state healthdepartments with respect to issuing alerts and responding to a potentialdisease outbreak such as influenza. We modeled the public health system as amulti-agent (or decentralized) partially observable Markov decision processwhere local and state health departments are decision makers. The model is usedto determine when local and state decision makers should issue an alert orinitiate mitigation actions such as vaccination in response to the existence ofa disease threat. The model incorporates the fact that health departments haveimperfect information about the exact number of infected people. The objectiveof the model is to minimize both false alerts and late alerts while identifyingthe optimal timing for alerting decisions. Providing such a balance betweenfalse and late alerts has the potential to increase the credibility andefficiency of the public health system while improving immediate response andcare in the event of a public health emergency. Using data from the 2009-2010H1N1 influenza outbreak to estimate model parameters including observations andtransition probabilities, computational results for near optimal solutions areobtained. In order to gain insight regarding the structure of optimalpolicies at the local and state levels, various model parameters includingfalse and late alerting costs are explored.
This research is a part of the NorthCarolina Preparedness and Emergency Response Research Center (NCPERRC) and wassupported by the Centers for Disease Control and Prevention (CDC) Grant 1PO1 TP000296-02.
]]>SPEAKER: Dimitris Bertsimas (Boeing Prof. of OR)
ABSTRACT:
In this presentation, we show a significant role that symmetry, afundamentalconcept in convex geometry, plays in determining the power of robustandfinitely adaptable solutions in multi-stage stochastic and adaptiveoptimization problems. We consider a fairly general class ofmulti-stage mixedinteger stochastic and adaptive optimization problems and propose agoodapproximate solution policy with performance guarantees that depend onthe
geometric properties such as symmetry of the uncertainty sets. Inparticular,we show that a class of finitely adaptable solutions is a goodapproximationfor both the multi-stage stochastic as well as the adaptiveoptimizationproblem. A finitely adaptable solution specifies a small set ofsolutions foreach stage and the solution policy implements the best solution fromthe given
set depending on the realization of the uncertain parameters in thepaststages. To the best of our knowledge, these are the first approximationresultsfor the multi-stage problem in such generality. (joint work with Vineet Goyal, Columbia University and Andy Sun, MIT)
Bio:
Dimitris Bertsimas is currently the Boeing Professor of OperationsResearch and the
codirector of the Operations Research Center at the Massachusetts Institute of Technology.
He has received a BS in Electrical Engineering and Computer Science at the National
Technical University of Athens, Greece in 1985, a MS in Operations Research at MIT in
1987, and a Ph.D in Applied Mathematics and Operations Research at MIT in 1988.
Since 1988, he has been in the MIT faculty.
SPEAKER: Professor Polly Biyu He
ABSTRACT:
We study the problem ofsetting nurse staffing levels in hospital operating rooms when there isuncertainty about the daily workload. We demonstrate in this healthcare servicesetting how information availability and choices of decision models affect anewsvendor's performance. We develop empirical models to predict the dailyworkload distribution and study how its mean and variance change with theinformation available. In particular, we consider different information setsavailable at the time of decision: no information, information on number ofcases, and information on number and types of elective cases. We use thesemodels to derive optimal staffing rules based on historical data from a USteaching hospital and prospectively test the performance of these rules. Ourempirical results suggest that hospitals could potentially reduce theirstaffing costs by an average of 39-49% (depending on the absence or presence ofemergency cases) by deferring the staffing decision until procedure-type informationis available. However, in reality, contractual and scheduling constraints oftenrequire operating room managers to reserve staffed hours several months inadvance, when little information about the cases is known. This motivates us toconsider the problem of adjusting the staffing level given information updates.Specifically, we develop decision models that allow the OR manager to adjustthe staffing level with some adjustment costs when he or she has betterinformation. We study how adjustment costs affect the optimal staffing policyand the value of having the flexibility to adjust staffing. We also demonstratehow to implement our adjustment policies by applying the optimal decision rulesderived from our models to the hospital data.
Joint work with Stefano Zenios, Franklin Dexter and Alex Macario.
]]>
SPEAKER: Reha Uzsoy
ABSTRACT:
Bio: Reha Uzsoy is Clifton A.Anderson Distinguished Professor in the Edward P. Fitts Department ofIndustrial and Systems Engineering at North Carolina State University. He holdsBS degrees in Industrial Engineering and Mathematics and an MS in IndustrialEngineering from Bogazici University, Istanbul, Turkey. He received his Ph.D inIndustrial and Systems Engineering in 1990 from the University of Florida. Histeaching and research interests are in production planning, scheduling, andsupply chain management. He is the author of one book, two edited books, andover eighty refereed journal publications. Before coming to the US he worked asa production engineer with Arcelik AS, a major appliance manufacturer inIstanbul, Turkey. He has also worked as a visiting researcher at IntelCorporation and IC Delco. His research has been supported by the NationalScience Foundation, Intel Corporation, Hitachi Semiconductor, HarrisCorporation, Kimberly Clark, Union Pacific, Ascension Health and GeneralMotors. He was named a Fellow of the Institute of Industrial Engineers in 2005,Outstanding Young Industrial Engineer in Education in 1997 and a UniversityFaculty Fellow by Purdue University in 2001, and has received awards for both undergraduateand graduate teaching.
TITLE: Graph-based methods for efficiently constructingfactorial designs
SPEAKER: Dr. Abhishek K. Shrivastava
ABSTRACT:
Fractional factorial designs are among the most popularclass of experimental designs among practitioners. Usually, a factorial designis selected after comparing all the designs in a catalog of designs for a givensize. Designs in these catalogs should be distinct under relabeling of factors,level labels and run order, i.e., they should be mutually non-isomorphic.Testing two designs for isomorphism is computationally hard, and constructionof non-isomorphic catalogs is even tougher with the large number of designsthat need to be compared for isomorphism. In this talk, I will present a newapproach for solving design isomorphism by representing designs as graphs. Theresulting graph isomorphism problem can be efficiently solved using methodsavailable in literature. Further, in the case of regular designs, I will showhow these graph representations can be exploited for speeding up the efficiencyof catalog construction by reducing the number of isomorphism tests. I willdemonstrate the gains from this approach by presenting results for 2-level regularfractional factorial and 2-level split-plot designs.
Bio:
Abhishek K. Shrivastava is an Assistant Professor in theDepartment of Manufacturing Engineering and Engineering Management in CityUniversity of Hong Kong, Hong Kong. He received his B. Tech. (Honors) inIndustrial Engineering from I.I.T. Kharagpur, India, in 2003, and his Ph.D. inIndustrial Engineering from Texas A&M University, College Station, USA, in2009. His research interests are in statistical modeling and analysis ofcomplex systems, design of experiments and rare event detection. He is a memberof INFORMS, IIE, ASA and IMS.
]]>In addition to this session, there is one on Thursday, August 19th from 1-2pm at the same location. Students only need to attend one of the sessions.
]]>]]>
SPEAKER: Dr. Anureet Saxena
ABSTRACT:
The underestimation of risk of optimized portfolios is a consistent criticism about risk models. Quantitative portfolio managers have historically used a variety of ad hoc techniques to overcome this issue in their investment processes. In this paper, we construct a theory explaining why risk models underestimate the risk of optimized portfolios. We show that the problem is not necessarily with a risk model, but is rather the interaction of expected returns, constraints, and a risk model in an optimizer. We develop an optimization technique that incorporates a dynamic Alpha Alignment Factor (AAF) into the factor risk model during the optimization process. Using actual portfolio manager backtests, we illustrate both how pervasive the underestimation problem can be and the effectiveness of the proposed AAF in correcting the bias of the risk estimates of optimized portfolios.
Speaker bio:
Dr. Anureet Saxena is a research associate at Axioma Inc. Prior to joining Axioma in 2008, he held research positions at Carnegie Mellon University (Egon Balas, mentor), Tata Institute of fundamental research (Narendra Karmarkar, mentor) and T.J. Watson Research Center (IBM).
His research interests include mixed integer linear and non-linear programming, stochastic programming and quantitative finance. He has published four scholarly articles in Mathematical Programming and has delivered more than thirty presentations at various professionalmeetings. Dr. Saxena is the recipient of the 2008 Gerald L. Thompson doctoral dissertation award in management science for his thesis titled Integer Programming, a Technology. He holds MS and PhD in industrialadministration from Tepper School of Business (CMU), BTech in computer science and engineering from IIT Bombay and is currently a level 2 candidate in the Chartered Financial Analyst (CFA) program.
SPEAKER: Greg Fasshauer
ABSTRACT:
I will first present two scattered data approximation methods from a numerical analysis point of view: radial basis function or kernel interpolation and moving least squares approximation. Then I will introduce the idea of approximate moving least squares approximation and connect all three methods via a residual updating algorithm. In parallel I will attempt to point out connections to an analogous set of methods (kriging, local polynomial regression and higher-order kernels for density estimation) in statistics. The idea of residual updating will be illuminated both from a more analytical perspective and at the numerical linear algebra level where we have rediscovered an old algorithm due to Riley [1].
[1] J.D. Riley. Solving systems of linear equations with a positive definite, symmetric, but possibly ill-conditioned matrix. Mathematical Tables and Other Aids to Computation 9/51 (1955), 96–101.
Brief bio:
Greg Fasshauer
Professor, Associate Chair and Director of Undergraduate Studies
Illinois Institute of Technology
Department of Applied Mathematics
Chicago, IL 60616
Education and Positions
* Since 1997: Assistant, associate and full professor, Department of Applied Mathematics, IIT
* Ralph P. Boas Visiting Assistant Professor: Department of Mathematics, Northwestern University (1995-1997)
* Ph.D. (Mathematics): Vanderbilt University with Larry L. Schumaker (1995)
* M.A. (Mathematics): Vanderbilt University with Larry L. Schumaker (1993)
* Diplom (Mathematics) & Staatsexamen (Mathematics and English): University of Stuttgart with Klaus Höllig (1991)
Research Interests (currently supported by NSF)
* Meshfree approximation methods for multivariate approximation and their application
* Radial basis functions and positive definite kernels
* Approximation theory
* Computer-aided geometric design
* Spline theory
* Numerical solution of PDEs
Books and 40+ papers
* Meshfree Approximation Methods with MATLAB
Interdisciplinary Mathematical Sciences - Vol. 6 World Scientific Publishers, Singapore, 2007
* Progress on Meshless Methods (edited with A.J.M. Ferreira, E.J. Kansa, and V.M.A. Leitão)
Computational Methods in Applied Sciences, Vol. 11 Springer, Berlin, 2009
SPEAKER: Professor Steve Pollock
ABSTRACt:
Developing mathematical models to describe (in some usefulway) operational situation is a critical component of doing effectiveOperations Research. Yet, for most students, learning and internalizingthe craft (or art) of doing mathematical modeling is a difficult andoften mysterious journey. And this struggle can actually be made worse byusing a normal classroom lecture-reading-studying mode of teaching. Moreover,the pedagogical challenge to those who try to develop and hone students' abilitiesto model is complicated when measures of performance normally used to judge"success" in this educational process are either not clear orpossibly pernicious. Indeed, choosing measures of performance for justabout any operational problems is often problematic.
Thetalk will address these issues via personal experiences, anecdotes andopinions. Any formal mathematics presented or discussed will beincidental, elementary or purposely unnecessarily complicated.
Biosketch:
Stephen M. Pollock is Herrick Emeritus Professor ofManufacturing and Professor Emeritus of Industrial and Operations Engineeringat the University of Michigan. He has been involved in applyingoperations research and decision analysis methods to understand and influenceof a variety of operational phenomena involving: military search and detection,criminal recidivism, manufacturing process monitoring, sequential allocation ofresources, predictive and proactive maintenance, sports, networks of queues,the stochastic behavior of infectious disease epidemics and the optimization ofradiation oncology plans.
Afterreceiving his Ph.D. at M.I.T. in 1964 he was a member of the technical staff atArthur D. Little, Inc. before joining the faculty at the U.S. NavalPostgraduate School in 1965 and the University of Michigan in 1969. Hewas chair of the IOE Department from 1981 through 1990. In 1992 he wasthe recipient of the Stephen S. Attwood Award, the highest honor awarded to afaculty member by the College of Engineering. He has authored over 65 technicalpapers, co-edited two books, and has served as a consultant to over 30industrial, governmental and service organizations.
ProfessorPollock was Associate Editor and Area Editor of Operations Research,Senior Editor of IIE Transactions, Associate Editor of ManagementScience and on the editorial boards of other journals. He has servedon various advisory boards for the National Science Foundation, on the ArmyScience Board, and chaired and been member of many National Research Councilcommittees, boards and panels. He was President of the OperationsResearch Society of America in 1986 and awarded the 2001 INFORMS Kimball Medalfor contributions to operations research and the management sciences. Hea fellow of INFORMS and the AAAS and is a member of the National Academy ofEngineering.
SPEAKER: Prof. Vijay Nair
ABSTRACT:
Traditional reliability analysis is based on time-to-failure data. In this talk, we discuss two different directions that lead to more informative reliability inference. The first is multi-state models. Challenges in statistical inference based on interval-censored data will be described, and methods for doing likelihood-based inference in semi-Markov models will be discussed. The second part of the talk will provide an overview of degradation models, describe a class of non-homogeneous Gaussian processes and some inference issues. This talk is based on joint work with current and former students Yang Yang and Xiao Wang.
Bio: Vijay Nair is the Donald A. Darling Professor of Statistics and Professor of Industrial and Operations Engineering at the University of Michigan, Ann Arbor. He has been Chair of the Department of Statistics sine 1998. Previously, he was a Research Scientist in the Mathematical Sciences and Operations Research Centers at Bell Laboratories for fifteen years.
His research interests include engineering statistics, information technology, quality and process improvement, industrial experiments, reliability and risk analysis, neuro-informatics, and statistical methods in behavioral intervention research. He also has extensive practical experience in the automotive, semiconductor, and telecommunications industries.
Vijay has a Bachelor’s degree in Economics from the University of Malaya (Malaysia) and a Ph. D. in Statistics from the University of California, Berkeley. He has been editor of Technometrtics, International Statistical Review, coordinating editor of the Journal of Statistical Planning and Inference, and has served on the editorial boards of many other leading statistics and quality journals. He is currently Vice President of the International Statistics Institute, President-elect of the International Society for Business and Industrial Statistics, and Chair of the Statistics Division of the American Society for Quality. He has chaired or co-chaired several panels of the National Academies and is a former chair of the Board of Trustees of the National Institute of Statistical Sciences. He is a Fellow of the American Association for the Advancement of Science, the American Statistical Association, the American Society for Quality, and the Institute of Mathematical Statistics.
SPEAKER: Professor William Q. Meeker
ABSTRACT:
Usually the warranty data response used to make predictionsof future failures is the number of weeks (or another unit of realtime) inservice. Use-rate information usually is not available (automobilewarrantydata are an exception, where both weeks in service and number of milesdrivenare available for units returned for warranty repair). With newtechnology,however, sensors and smart chips are being installed in many modernproductsranging from computers and printers to automobiles and aircraftengines. Thusthe coming generations of field data for many products will provideinformationon how the product has been used and the environment in which it wasused. Thispaper was motivated by the need to predict warranty returns for aproduct withmultiple failure modes. For this product, cycles-to-failure/use-rateinformation was available for those units that were connected to thenetwork.We show how to use a cycles-to-failure model to compute predictions andprediction intervals for the number of warranty returns. We alsopresentprediction methods for units not connected to the network. In order toprovideinsight into the reasons that use-rate models provide betterpredictions, wealso present a comparison of asymptotic variances comparing thecycles-to-failure and time-to-failure models.
Bio: William Q. Meeker is a Professor of Statistics andDistinguished Professor of Liberal Arts and Sciences at Iowa State University.He is a Fellow of theAmerican Statistical Association (ASA) and the American Society forQuality(ASQ) and a past Editor of Technometrics. He is co-author ofthe books StatisticalMethods for Reliability Data with Luis Escobar (1998), and StatisticalIntervals: A Guide for Practitioners with Gerald Hahn (1991), sixbookchapters, and of numerous publications in the engineering andstatisticalliterature. He has won the ASQ Youdenprize four times and the ASQ Wilcoxon Prize three times. He wasrecognized bythe ASA with their Best Practical Application Award in 2001 and by theASQStatistics Division’s with their W.G. Hunter Award in 2003. In 2007 hewasawarded the ASQ Shewhart medal. He has done research and consultedextensivelyon problems in reliability data analysis, warranty analysis,reliability testplanning, accelerated testing, nondestructive evaluation, andstatisticalcomputing.
SPEAKER: David Goldberg
ABSTRACT:
Parallel server queues arise in many applications, ranging from call centers to national security and health care. Understanding how these systems behave when the number of servers is large and the service distribution is non-Markovian is a difficult problem. In this talk, we resolve several open questions related to a certain heavy traffic scaling regime (Halfin-Whitt) for parallel server queues, which has recently been used in the modeling of call centers. In particular, we derive the asymptotics of the steady-state queue length for a very general class of service distributions. We also bound the large deviations behavior of the limiting steady-state queue length, and prove that the associated critical exponent takes a particularly simple form in certain cases. Our main proof technique involves bounding the multiserver queue between two simpler systems. These systems exhibit an interesting duality, and yield bounds of a very general nature, which may be useful in answering a range of questions related to the modeling and optimization of queues.
]]>SPEAKER: Dr. Mou-Hsiung (Harry) Chang
ABSTRACT:
In this talk, we consider an infinite time horizon portfolio optimization problem in a market that consists of one savings account and one stock account whose unit price satisfies a nonlinear stochastic functional differential equation. Within the solvency region the investor is allowed to consume from the savings account and can make transactions between the two assets subject to paying capital-gain taxes as well as a fixed plus proportional transaction cost. The main objective is to seek an optimal consumption-investment strategy in order to maximize the expected utility from the total discounted consumption over the infinite time horizon. The portfolio optimization problem is formulated as a stochastic control problem that involves both the classical and impulse controls. A quasi-variational HJB inequality for the value function is derived and the verification theorem for the optimal investment consumption strategy is obtained. The value function is also shown to be the unique viscosity solution of the HJB inequality.
Bio: Dr. M. H. Chang is currently the manager of the Probability & Statistics Program and acting division chief of the Mathematical Sciences Division at the U. S. Army Research Office (ARO). Prior to joining ARO, Dr. Chang had been a tenured professor in the Department of Mathematical Sciences at the University of Alabama in Huntsville (UAH) for twenty-eight years and had served as Chair of the Mathematical Sciences Department for eight years. He received his B.S. in Applied Mathematics from National Chung-Hsing University (in Taiwan) and his M.S. and Ph.D. in Mathematics from the University of Rhode Island.
Dr. Chang publishes extensively in stochastic analysis, stochastic control, mathematical finance, and quantum Markov processes. He has published one research monograph “Stochastic Control of Hereditary Systems and Applications”, Volume 59 of the Stochastic Modeling and Applied Probability Series, Springer, New York, January 2008, ISBN 978-0-387-75805-3, more than 60 referred journal articles, and has made more than 80 invited presentations at professional conferences and universities. He is currently an associate editor for “Journal of Applied Mathematics and Stochastic Analysis” and “Stochastic Analysis and Applications”. He has also served as a referee for numerous mathematics and applied mathematics journals.
]]>SPEAKER: Yurii Nesterov
ABSTRACT:
In this talk we describe the new methods for solvinghuge-scale optimization problems. For problems of thissize, even the simplest full-dimensional vector operationsare very expensive. Hence, we suggest to apply anoptimization technique based on random partial update of
decision variables. For these methods, we prove the globalestimates for the rate of convergence. Surprisinglyenough, for certain classes of objective functions, ourresults are better than the standard worst-case bounds fordeterministic algorithms. We present constrained andunconstrained versions of the method, and its acceleratedvariant. Our numerical test confirms a high efficiency ofthis technique on problems of very big size.
The 8th Annual QCF Day Symposium will be held April 2, 2010. The symposium will take place in Atlanta, Georgia. It is hosted by the Quantitative and Computational Finance program at the Georgia Institute of Technology.
The purpose of this symposium is to showcase the newest and mostinnovative approaches to quantitative finance used today. In previousyears, over several hundreds of people have attended this symposium tolisten to both academic and non-academic invited experts. By and largethe talks concentrate on specific topics related to how Quantitativeand Computational Finance is being used today. The speakers in previousyears have included Senior Executives, Managing Directors, VicePresidents, Group Directors, Quantitative Associates, Senior Economistsand others from a wide range of organizations - Bank of America-MerrillLynch, Deutsche Bank, Goldman Sachs, INVESCO, ING InvestmentManagement, Morgan Stanley, NYFRB, Risk Metrics, S&P, SunGard,SunTrust Robinson Humphrey, UBS and Vanguard - as well as professorsfrom Columbia, Toronto, Chicago, Princeton, and New York University.
There will be many opportunities to speak with students andprofessionals about the work environment and activities in the field ofquantitative finance. The symposium is attended by students from theGeorgia Tech MS QCF Program and students from educational programsstrongly overlapping with the QCF area; by researchers and practitionersworking within the QCF area who want to share their experiences andlearn more from others in the QCF area; and by other individuals seekinginformation about the content of the very diverse QCF area.
To register for QCF Day, click HERE
For more information on the QCF program at Georgia Tech, please download ourprogram brochure.
]]>
SPEAKER: Professor S.M. Meerkov
ABSTRACT:
Production Systems Engineering (PSE) is an emerging branch of Engineering intended to uncover fundamental principles that govern production systems and utilize them for the purposes of analysis, continuous improvement, and design. In PSE, the machines areassumed to be unreliable and the buffers are finite. Under these assumptions, production lines are nonlinear stochastic systems. The study of their statics and dynamics is the goal of PSE.In this talk, the main problems of PSE and their solutions will be described along with several applications. In addition, the so-called PSE Toolbox, which implements themethods and algorithms developed, will be discussed. A demo of the toolbox can be found at http://www.ProductionSystemsEngineering.com
BIO: Semyon M. Meerkov received his MSEE degree from the Polytechnic of Kharkov, Ukraine, in 1962 and Ph.D. in Systems Science from the Institute of Control Sciences, Moscow, Russia, in 1966. He was with the Institute of Control Sciences until 1977. From 1979 to 1984 he was with the Department of Electrical and Computer Engineering, Illinois Institute of Technology, Chicago, IL. Since 1984 he has been a Professor at theDepartment of Electrical Engineering and Computer Science of the University of Michigan, Ann Arbor, MI. He has held visiting positions at UCLA (1978-1979), Stanford(1991), Technion, Israel (1997-1998 and 2008) and Tsinghua, China (2008). He was the Editor-in-Chief of Mathematical Problems in Engineering, Department Editor for Manufacturing Systems of IIE Transactions and Associate Editor of several other journals. His research interests are in Systems and Control with applications toProduction Systems, Communication Networks, and Theory of Rational Behavior.
]]>SPEAKER: Adrian lewis
ABSTRACT:
Concrete optimization problems, while often nonsmooth, are not pathologically so. The class of "semi-algebraic" sets and functions -those arising from polynomial inequalities - nicely exemplifies nonsmoothness in practice. Semi-algebraic sets (and theirgeneralizations) are common, easy to recognize, and richly structured, supporting powerful variational properties. In particular I will discussa generic property of such sets - partial smoothness - and its relationship with a proximal algorithm for nonsmooth compositeminimization, a versatile model for practical optimization.
Bio:
Adrian S. Lewis was born in England in 1962. He is a Professor at Cornell University in the School of Operations Research and Industrial Engineering. Following his B.A., M.A., and Ph.D. degrees from Cambridge, and Research Fellowships at Queens' College, Cambridge and Dalhousie University, Canada, he worked in Canada at the University of Waterloo (1989-2001) and Simon Fraser University (2001-2004). He is an Associate Editor of the SIAM Journal on Optimization, Mathematics of Operations Research, and the SIAM/MPS Book Series on Optimization, and is a Co-Editor for Mathematical Programming. He received the 1995 Aisenstadt Prize, from the Canadian Centre de Recherches Mathematiques, the 2003 Lagrange Prize for Continuous Optimization from SIAM and the Mathematical Programming Society, and an Outstanding Paper Award from SIAM in 2005. He co-authored "Convex Analysis and Nonlinear Optimization" with J.M. Borwein.
Lewis' research concerns variational analysis and nonsmooth optimization, with a particular interest in optimization problems involving eigenvalues.
SPEAKER: Zuo-Jun Max Shen (Faculty Candidate)
ABSTRACT:
We postulate the existence of a “reverse bullwhip effect” (RBWE) that occurs during and immediately after supply disruptions. Whereas the classical bullwhip effect (BWE) describes an increase in demand/order volatility as one moves upstream in the supply chain, the RBWE describesthe opposite. We motivate our analysis by an example involving gasoline-buying patterns following Hurricane Katrina in 2005. We then present theoretical and empirical evidence for a RBWE that occurs in a variety of situations involving supply uncertainty. RBWE has great implication in service system design and management, so we examine causes of RBWE, discuss its impact, and suggest strategies for mitigating it.
Bio:
Zuo-Jun (Max) Shen received his Ph.D. from Northwestern University. He has been active in the following research areas: integrated supply chain design and management, market mechanism design, applied optimization, and decision making with limited information. He is currently on the editorial/advisory board for several leading journals. He received the CAREER award from National Science Foundation in 2003.
SPEAKER: Professor Enrique del Castillo
ABSTRACT:
We show how Statistical ShapeAnalysis, a set of techniques used to model the shapes of biological and otherkind of objects in the natural sciences, can be used also to model thegeometric shape of a manufactured part. We first review Procrustes-basedmethods, and emphasize possible solutions to the basic problem of havingcorresponding, or matching, labels in the measured ``landmarks", thelocations of the measured points on each part acquired with a CMM or similarinstrument. The analysis of experiments with shape responses is discussednext. The usual approach in practice isto estimate the form error of the part and conduct an ANOVA on the form errors.Instead, an F ANOVA test due to Goodall and a new permutation ANOVA test forshapes are presented. Real data sets as well as simulated shape data ofinterest in manufacturing were used to perform power comparisons for 2 and 3dimensional shapes. The ANOVA on the form errors was found to have poorperformance in detecting mean shape differences in circular and cylindricalparts. The ANOVA F test and the Permutation ANOVA test provide highest power to detectdifferences in the mean shape. It is shown how these tests can also be appliedto general "free form" shapes of parts where no standard definitionof form error exists in manufacturing practice. New visualization tools,including main effect and interaction plots for shapes and deviation fromnominal plots are presented to help interpreting the results of experimentswhere the response is a shape.
Bio:
Dr. Castillo isa Distinguished Professor of Industrial Engineering. Dr. Castillo also holds a joint appointmentwith the Department of Statistics. He isan author of over 85 journal papers and 2 textbooks and a former NSF CAREER awardee,Fulbright Scholar, and the Editor in Chief of the Journal of Quality Technology(2006-2009).
]]>Speaker: Professor Jeff Wu (ISyE,Gatech)
2. Title: Multi-LayerDesigns for Computer Experiments
Speaker: Professor Roshan J.Vengazhiyil (ISyE, Gatech)
Abstract: Computer experiments play a major role in the modern eraof scientific and technological development. In designing computer experiments,Latin hypercube designs (LHDs) are widely used. However, finding an optimal LHDis computationally cumbersome. On the other hand, although many optimal designsare well known for physical experiments, the redundancy of design points makethem undesirable for computer experiments. In this work, we present a new classof space-filling designs developed by splitting two-level full or fractionalfactorial designs into multiple layers. The method takes advantages of manyavailable results in designing physical experiments and therefore, the proposedMulti-layer designs (MLDs) are easy to generate. Moreover, our numerical studyshows that MLDs can have better space-filling properties than optimal LHDs.
3. Title: SomeNew Advances in Design and Modeling of Computer Experiments
Speaker: Professor Peter Z. G. Qian (Statistics, Wisconsin)
Abstract: Computer models are now becoming ubiquitous in nearly allfields of sciences and engineering. Design and modeling are two key aspects ofcomputer experiments. In this talk, I will report some recent advances in bothaspects. Specific topics include a new approach for emulation of computermodels with qualitative and quantitative factors; sequential space-fillingdesigns; Sudoku based space-filling designs; and sliced Latin hypercube designsfor ensembles of computer models.
4. Title: Analysisof Computer Experiments with Functional Response
Speaker: Professor Ying Hung (Statistics, Rutgers)
Abstract: Most existing methods for analyzing computer experimentswith single outputs such as kriging cannot be easily applied to functionaloutputs due to the computational problems caused by high-dimensionality of theresponse. In this paper, we develop an efficient implementation of kriging foranalyzing functional responses. Our methodology uses a two-stage model buildingprocedure with Kronecker products and an improved EM algorithm for estimation.The methodology is illustrated using a computer experiment conducted foroptimizing residual stresses in machined parts. This is a joint work with V.Roshan Joseph and Shreyes N. Melkote.
]]>SPEAKER: Professor Stephen Fienberg
ABSTRACT:
All too often academic statisticians think of their role as theproduction of new theory and methods. But at least as important is therole we can fulfill in support national activities and projectsrequiring statistical insights and rigor. Many (but far from all) ofthese are based at the National Research Council and its committees andpanels. I will illustrate the value and creativity of statisticalthinking in a diverse set of projects that span the past halfcentury---from the "National Halothane Study" through "The Polygraph andLie Detection." A crucial feature of several of these activities hasbeen the impetus for development of new research. I will also emphasizethe role I believe statisticians should be playing in such activitiesin the future.
]]>SPEAKER: Professor David Siegmund
ABSTRACT:
In a previous paper (Biometrics, 2006, pp. 22-32) we derived a Bayes Information Criterion (BIC) for determining the number of change-points in a sequence of independentobservations when the number $m$ of change-points is assumed to remain bounded as the number of observations increases. Here we generalize that result to include multiple aligned sequences with intervals of simultaneous change that occur in a fraction of the sequences and a total number of of change-points $m$ that can increase with the sample size; and we include in the criterion terms that increase at rate $m$. Stochastic terms that enter into the new criterion involve integrals and maxima of two-sided Brownian motion with negative drift. Examples involve segmenting aligned DNA sequences according to copy number variations that occur at the same position in a fraction of the sequences.
This is joint research with N. Zhang.
SPEAKER: Professor Spiridon Reveliotis
ABSTRACT:
The problemof liveness-enforcing supervision -- or, deadlock avoidance -- for complexresource allocation systems (RAS) is a well-documented problem in supervisorycontrol theory. Most of the past research on it has acknowledged the fact thatthe maximally permissive liveness-enforcing supervisor (LES) possessessuper-polynomial complexity for most RAS classes, and therefore, it hasresorted to solutions that trade off maximal permissiveness for computationaltractability. In the presented work, we distinguish between the"off-line" and the "on-line" computation that is requiredfor the effective implementation of the maximally permissive LES, and we seek todevelop representations of the maximally permissive LES that require"minimal" on-line computation.
The particular representation that we adopt is that of a compact classifierthat will effect the underlying dichotomy of the reachable state space into safeand unsafe subspaces. Through a series of reductions of the derivedclassification problem, we are able to attain extensive reductions in, both,(i) the computational complexity of the off-line task of the construction ofthe sought classifier, and (ii) the complexity involved in the on-lineclassification process itself. We formally establish completeness andoptimality properties for the proposed design procedures. We also offerheuristics that, if necessary, can alleviate the computational effort that isnecessary for the construction of the sought classifier. Finally, wedemonstrate the efficacy of the developed approaches through a series ofcomputational experiments. To the best of our knowledge, these experiments alsoestablish the ability of the proposed methodology to effectively computetractable implementations of the maximally permissive LES for problem instancessignificantly beyond the capacity of any other approach currently available inthe literature.
SPEAKER: Dr. Haipeng Xin
ABSTRACT:
Markowitz's celebrated mean-variance portfolio optimization theoryassumes that the means and covariances of the underlying asset returnsare known. In practice, they are unknown and have to be estimated fromhistorical data. Plugging the estimates into the efficient frontier
that assumes known parameters has led to portfolios that may performpoorly and have counter-intuitive asset allocation weights; this has beenreferred to as the ``Markowitz optimization enigma.'' After reviewingdifferent approaches in the literature to address these difficulties, weexplain the root cause of the enigma and propose a new approach to resolve
it. Not only is the new approach shown to provide substantialimprovements over previous methods, but it also allows flexible modelingto incorporate dynamic features and fundamental analysis of the trainingsample of historical data, as illustrated in simulation and empirical studies. This is a joint work with Tze Leung Lai (Stanford University) and Zehao Chen (Bosera Funds).
Short bio: Haipeng Xing graduated from the Department of Statistics at Stanford University at 2005, and then jointed the Department of Statistics at Columbia University. In 2008, he moved the Department of Applied Maths and Statistics at SUNY, Stony Brook. His research interests include financial econometrics and engineering, time series modeling and adaptive control, and change-point problems.
SPEAKER: Prof. Kwok Tsui
ABSTRACT:
For public health surveillance, timely detection of a rate increase in disease incidence is very important. This talks reviews some popular methods for temporal surveillance and proposes a general framework for spatial and spatiotemporal surveillance based onlikelihood ratio statistics over windows of tests. We show that the CUSUM and other popular likelihood ratio statistics are special cases under such a general framework. We compare the efficiency of these surveillance methods in spatial and spatiotemporal cases for detecting clusters of incidence using both Monte Carlo simulations and a real example. We will also discuss the generalization of weighted likelihood ratio tests for detecting different shift magnitudes under homogeneous and non-homogeneous populations.
]]>SPEAKER: Professor Daniel Apley
ABSTRACT:
In statistical analysis and data mining of multivariate data sets, many problems can be viewed as discovering variation patterns in a set of N observations of n variables. The term "variation pattern" refers to the structured, interdependent manner in which the n variables may vary over the N observations. In a very general mathematical representation we view each multivariate observation as a vector in n-dimensional space. Then over the set of N observations, we assume the data consist of a structured component plus noise, where the structured component lies on a p-dimensional manifold withp << n. The objective is to learn, or discover, the manifold based only on the set of data, with no prior knowledge of what to expect. Discovery of the manifold is useful in many different contexts: Denoising noisy images and other multivariate data; dimensionality reduction of large data sets; extraction of important features for enhancing subsequent analyses; exploratory analyses for identifying and understanding relationships between variables; etc. In this talk, I will discuss the manifold learning problem, applications, and algorithms. Linear structured manifolds can be easily discovered with standard principal components and factor analyses. Consequently, this talk will focus on discovering nonlinear manifolds, which is a much more challenging and nuanced problem.
Bio: Daniel W. Apley is an Associate Professor of Industrial Engineering & Management Sciences at Northwestern University. His research interests lie at the interface of engineering modeling, statistical analysis, and data mining, with particular emphasis on manufacturing variation reduction applications in which very large amounts of data are available. His research has been supported by numerous industries and government agencies. He received the NSF CAREER award in 2001, the IIE Transactions Best Paper Award in 2003, and the Wilcoxon Prize for best practical application paper appearing in Technometrics in 2008. He currently serves as Editor-in-Chief for the Journal of Quality Technology and has served as Chair of the Quality, Statistics & Reliability Section of INFORMS, Director of the Manufacturing and Design Engineering Program at Northwestern, and Associate Editor for Technometrics.
]]>
SPEAKER: Barry Smith
ABSTRACT:
Beginning in 2002,Travelocity and the Sabre Research Group collaborated to develop theEnterprise Network Model (ENM). The ENM combines discrete choicecustomer modeling with simulation and large-scale optimization toimprove Travelocity's management of its on-line travel distributionbusiness. The ENM helped Travelocity become a more effective retailerand contributed $43 million per year to Travelocity's bottom line. This project was recognized in 2005 as an Edelman competition finalist.
]]>SPEAKER: Haipeng Xing
ABSTRACT:
Markowitz's celebrated mean-variance portfolio optimization theoryassumes that the means and covariances of the underlying asset returnsare known. In practice, they are unknown and have to be estimated fromhistorical data. Plugging the estimates into the efficient frontierthat assumes known parameters has led to portfolios that may performpoorly and have counter-intuitive asset allocation weights; this has beenreferred to as the "Markowitz optimization enigma.'' After reviewingdifferent approaches in the literature to address these difficulties, weexplain the root cause of the enigma and propose a new approach to resolveit. Not only is the new approach shown to provide substantialimprovements over previous methods, but it also allows flexible modelingto incorporate dynamic features and fundamental analysis of the trainingsample of historical data, as illustrated in simulation and empirical studies. This is a joint work with Tze Leung Lai (Stanford University) and Zehao Chen (Bosera Funds).
Short bio: Haipeng Xing graduated from the Department of Statistics at Stanford University at 2005, and then jointed the Department of Statistics at Columbia University. In 2008, he moved the Department of Applied Maths and Statistics at SUNY, Stony Brook. His research interests include financial econometrics and engineering, time series modeling and adaptive control, and change-point problems.
]]>]]>
SPEAKER: Dr. Qingyu Yang, ResearchFellow
ABSTRACT:
The wide deployment and application of distributed sensing and computersystems have resulted in multi-stream sensing data leading to both temporallyand spatially dense data-rich environments, which provides unprecedentedopportunities for improving operations of complex systems in both manufacturingand healthcare applications. However, it also brings out new researchchallenges on data analysis due to high-dimensional and complextemporal-spatial correlated data structure. In this talk, as an example of myresearch work, I will discuss a critical research issue on how to separateimmeasurable embedded individual source signals from indirect mixed sensormeasurements. In this research, a hybrid analysis method is proposed byintegrating Independent Component Analysis and Sparse Component Analysis. Theproposed method can efficiently estimate individual source signals that includeboth independent signals and dependent signals which have dominant components inthe time or linear transform domains. With source signals identified, it isfeasible to monitor each source signal directly and provide explicit diagnosticinformation.
Bio:
Dr. Qingyu Yang is currently a postdoctoral research fellow with theDepartment of Industrial & Operations Engineering at the University ofMichigan-Ann Arbor. He received a M.S. degree in Statistics and a Ph.D. degreein Industrial Engineering from the University of Iowa in 2007 and 2008,respectively. He also held a B.S. degree in Automatic Control (2000) and a M.S.degree in Intelligent System (2003) from the Universityof Science and Technology Universityof China (USTC, China).His research interests include distributed sensor system, information system,and applied statistics. He was the recipient of the Best Paper Award from IndustrialEngineering Research Conference (IERC) 2009.
]]>Stewart School of Industrial and Systems Engineering
404.385.3102
]]>SPEAKER: Qie He
ABSTRACT:
Despite the elegant theory of cuts for mixed-integer programs with two rows and two integer variables derived from triangle and quadrilateral integer-free bodies, there has been only limited computational evidence that these cuts are capable of giving better results than those derived from simple splits, i.e. Gomory cuts which are derived from a single row. In this paper, we show that under a certain probabilistic model, a cut coefficient derived from a simple split is likely to be stronger than that derived from a triangle and as strong as that derived from a quadrilateral. Furthermore, we use our probabilistic model to compare the volume cut off from the linear relaxation by a split cut and a type 1 triangle cut. We empirically demonstrate that the probability that the split cut cuts off more volume than the triangle cut is quite high when the number of continuous variables is large.
This is a joint work with Shabbir Ahmed and George Nemhauser.
SPEAKER: Ji Zhu
ABSTRACT:
In this talk, we propose a computationally efficient approach for selecting non-zero partial correlations under the high-dimension-low-sample-size setting. This method assumes the overall sparsity of the partial correlation matrix and employs sparse regression techniques for model fitting. We illustrate the performance of our method by extensive simulation studies. It is shown that our method performs well in both non-zero partial correlation selection and the identification of hub variables, and also outperforms two existing methods. We then apply our method to a microarray breast cancer data set and identify a set of "hub genes" which may provide important insights on genetic regulatory networks. Finally, we prove that, under a set of suitable assumptions, the proposed procedure is asymptotically consistent in terms of model selection and parameter estimation. This is joint work with Jie Peng, Pei Wang and Nengfeng Zhou.
]]>SPEAKER: ProfessorE. A. Elsayed
ABSTRACT:
Accelerated LifeTesting (ALT) is designed and conducted to obtain failure observations in ashort time by subjecting test units to severer than normal operating conditionsand use the data for reliability prediction. Many types of stress loadings suchas constant-stress, step-stress and cyclic-stress can be utilized whenconducting ALT. Extensive research has been conducted on the analysis of ALTdata obtained under constant stress loading. However, the equivalency of ALTexperiments involving different stress loadings has not been investigated. Inthis paper, we provide definitions for the equivalency of test plans, generalequivalent ALT plans and some special types of equivalent ALT plans are explored.For demonstration, a constant-stress ALT and a ramp-stress ALT for miniaturelamps are presented and their equivalency is investigated.
]]>
SPEAKER: David Banks
ABSTRACT:
Classical game theory has been an unreasonable description for humanbehavior, and traditional analyses make strong assumptions about commonknowledge and fixed payoffs. Classical risk analysis has assumed thatthe opponent is non-adversarial (i.e., "Nature") and thus isinapplicable to many situations. This work explores Bayesianapproaches to adversarial risk analysis, in which each opponent mustmodel the decision process of the other, but there is the opportunityto use human judgment and subjective distributions. The approach isillustrated in the analysis of two important applications: sealed bidauctions and simple poker; some related work on counterbioterrorism isalso covered. The results in these three applications areinterestingly different from those found in previous work.
]]>SPEAKER: Professor Myong K. (MK) Jeong
ABSTRACT:
Support Vector Machine(SVM) is one of the most popular data mining tools for solving classificationand regression problems. Due to its high prediction accuracy, SVM has beensuccessfully used in various fields. However, SVM has the following drawbacks. First, it is not easy toget an explicit description of the discrimination (or regression) function inthe original input space and to make a variable selection decision in the inputspace. Second, depending on the magnitude and numeric range of the given datapoints, the resulting kernel matrices may be ill-conditioned, so learningalgorithms may be suffered from numerical instability even though data scalinggenerally helps to handle this kind of issues but may not be always effective.Third, the selection of an appropriate kernel type and its parameters can becomplex while the performance of the resulting functions is heavily influenced.
To overcome thesedrawbacks, this talk presents the sparsesignomial classification and regression (SSCR) model. SSCR seek a sparsesignomial function by solving a linear program to minimize the weighted sum ofthe ℓ1-norm of the coefficient vector of the function and the ℓ1-normof violation (or loss) caused by the function. SSCR can explorevery highdemensional feature spaces with less sensitivity to numerical values or numericranges of the given data. Moreover, this method give an explicit description ofthe resulting function in the original input space, which can be used forprediction purposes as well as interpretation purposes. We present a practicalimplementation of SSCR based on the column generation and explore sometheoretical properties of the proposed formulation. Computational study showsthat SSCR is competitive or even better performance compared to other widelyused learning methods for classification and regression.
]]>SPEAKER: J. Michael Harrison
ABSTRACT:
Motivatedby applications in financial services, we consider the following customizedpricing problem. A seller of some goodor service (like auto loans or small business loans) confronts a sequence ofpotential customers numbered 1, 2, … , T. These customers are drawn at random from apopulation characterized by a price-response function r(p). That is, if the seller offers price p, then the probability of a successfulsale is r(p). The profit realized froma successful sale is p(p) = p – c, where c > 0 is known.
Ifthe price-response function r(×) were also known, thenthe problem of finding a price p* tomaximize r(p)p(p) would be simple, and theseller would offer price p* to eachof the T customers. We consider the more complicated case where r(×) is fixed but initially unknown: roughly speaking, the seller wants tochoose prices sequentially so as to maximize the total profit earned from the T potential customers; each successivechoice involves a trade-off between refined estimation of the unknownprice-response function (learning) and immediate profit (earning).
*Joint work with Bora Keskin and Assaf Zeevi
]]>SPEAKER: Dr. Gerard Cornuejols
ABSTRACT:
his talk will be based on joint work with Borozan, Basu,Conforti and Zambelli. We extend a theorem of Lovasz characterizingmaximal lattice-free convex sets. This result has implications in integerprogramming. In particular we show a ono-to-one correspondance betweenthese sets and minimal inequalities.
]]>SPEAKER: Professor Edward Kaplan
ABSTRACT:
This article presents the first model developedspecifically for understanding the infiltration and interdiction of ongoingterror plots by undercover intelligence agents, and does so via novelapplication of ideas from queueing theory and Markov population processes. Theresulting "terror queue" models predict the number of undetectedterror threats in an area from agent activity/utilization data, and alsoestimate the rate with which such threats can be interdicted. The modelstreat terror plots as customers and intelligence agents as servers. Agentsspend all of their time either detecting and infiltrating new terror plots (inwhich case they are "available"), or interdicting already detectedterror plots (in which case they are "busy"). Initially we examine aMarkov model assuming that intelligence agents, while unable to detect allplots, never err by falsely detecting fake plots. While this model can besolved numerically, a simpler Ornstein-Uhlenbeck diffusion approximation yieldssome results in closed form while providing nearly identical numericalperformance. The transient behavior of the terror queue model isdiscussed briefly along with a sample sensitivity analysis to study how modelpredictions compare to simulated results when using estimated versus knownterror plot arrival rates. The diffusion model is then extended to allow forthe false detection of fake plots. Such false detection is a real feature ofcounterterror intelligence given that intelligence agents or informants canmake mistakes, as well as the proclivity of terrorists to deliberatelybroadcast false information. The false detection model is illustrated usingsuicide bombing data from Israel.
]]>
SPEAKER: Dr. Tao Tang
ABSTRACT:
The globaldemand for rail services continues to outpace the available capacity and the infrastructure. At themeantime, aging systems and traditional business practices provide significant limitations to effectively address these challenges.In order to accommodate the significant increase of demands in rail services,it is important to increase the capacity of the railway network. Another paramount issue in rail service is thesafety – It is essential to design and operate a system that can providereliable and efficient communication and control of railway signals andeliminate any potential dangers in the system.
The Communication Based Train Control (CBTC) system uses radio signal to transmit real time controlinformation between thetrain and the wayside. With the help of the CBTC system,the train will continually monitor its position and operation status, whichsignificantly improves the dependability and increasesthe capacity of a railway network. Currently, the CBTC system has been adopted bymore and more railway networks in the world. A Cyber-Physical Rail Systems based on CBTC, with more enhancedinformation collection, Internet-based communications, and safety andefficiency improvement, will be the future trend in the railway systemcommunication and safety.
This presentation will introduce some key technologies in the CBTCsystem. The framework, research challenges, andopportunities of cyber-physical rail system will be discussed also.
]]>SPEAKER: Andres Felipe Hernandez Moreno and Professor Martha Grover
ABSTRACT:
Gaussian process model is an empirical modeling approach that has been widely applied in engineering for the approximation of deterministic functions, due its flexibility and ability to interpolate observed data. Despite its statistical properties, Gaussian process models (GPM) have not been employed to describe the dynamics of stochastic complex system like nanoscale phenomena. This presentation describes the methodology to construct approximate models for multivariate stochastic dynamic simulations using GPM, combining ideas from design of experiments, spatial statistics and dynamic systems modeling. In particular, the effect of sampling strategies in the identification and prediction of the GPM is analyzed in detailed. The methodology is applied in the prediction of a dynamic size distribution during the synthesis of platinum nanoparticles under supercritical CO_2 conditions.
]]>SPEAKER: Lulu Kang
PhD Candidate (in the Statistics Program)
School of Industrial and Systems Engineering
Georgia Tech
ABSTRACT:
Computer experiments simulate the engineering systems by implementing the mathematical models governing the systems in computers. Recently, experiments having large number of input variables and experimental runs started to emerge. In the existing literature, kriging has been commonly used for approximating the complex computer models, but it has limitations for dealing with the large-scale experiments due to its computational complexity and numerical stability. In this work, I propose a new modeling approach known as regression-based inverse distance weighting (RIDW). The new predictor is shown to be computationally more efficient than kriging while producing comparable prediction performance. We also develop a heuristic method for constructing confidence intervals for prediction. I will also discuss extensions of RIDW and my future research directions on this exciting topic
Bio:
Lulu Kang is a Ph.D. candidate in the Statistics Program of the School of Industrial and Systems Engineering at Georgia Institute of Technology. She is working with Professor Roshan J. Vengazhiyil. Her research interests are in developing statistical theories and methodologies, as well as their applications in physical science and engineering.
Abstract
(To be announced)
SPEAKER: Professor Eberhard O. Voit
ABSTRACT:
The combination of high-throughput methods of molecular biology with advanced mathematical and computational techniques has propelled the emergent field of systems biology into a position of prominence. Unthinkable only a decade ago, it has become possible to screen and analyze the expression of entire genomes, simultaneously assess large numbers of proteins and their prevalence, and characterize in detail the metabolic state of a cell population. While very important, the focus on comprehensive networks of biological components is only one side of systems biology. Complementing large-scale assessments, and sometimes at risk of being forgotten, are more subtle analyses that rationalize the design and functioning of biological modules in exquisite detail. This intricate side of systems biology aims at identifying the specific roles of processes and signals in smaller, fully regulated systems by computing what would happen if these signals were lacking or organized in a different fashion. I will exemplify this type of approach with two examples. The first is a detailed analysis of the regulation of glucose utilization in /Lactococcus lactis/. This organism is exposed to alternating periods of glucose availability and starvation. During starvation, it accumulates an intermediate of glycolysis, which allows it to take up glucose immediately upon availability. This notable accumulation poses a non-trivial control task that is solved with an unusual, yet ingeniously designed and timed feedforward activation system. The elucidation of this control system required high-precision /in vivo/ data on the dynamics of intracellular metabolite pools, combined with methods of nonlinear systems analysis, and may serve as a paradigm for multidisciplinary approaches to fine-scaled systems biology. The second example describes our attempts to understand signal transduction in the human brain, along with perturbations in diseases like Parkinson’s disease and Schizophrenia.
*/References:/*
Voit, E.O.: /Computational Analysis of Biochemical Systems. A Practical Guide for Biochemists and Molecular Biologists/, xii + 530 pp., Cambridge University Press, Cambridge, U.K., 2000.
Voit, E.O., A.R. Neves, and H. Santos. The Intricate Side of Systems Biology. /PNAS/, 103(25), 9452-9457, 2006.
Qi, Z., G. W. Miller, and E. O. Voit: Computational analysis of determinants of dopamine dysfunction. /Synapse/* 63*: 1133-1142, 2009.
Wu, Jialiang, Z. Qi, and E.O. Voit: Investigation of delays and noise in dopamine signaling with hybrid functional Petri nets. In Silico Biol. 10, 0005 (2010).
SPEAKER: Dr. William Pulleyblank
ABSTRACT:
In2009, IBM launched a major activity in business analytics andoptimization. This includes operations research, data analytics,statistical analysis, business optimization and risk management. Iwill review this and discuss how it fits with our Smarter Planetsstrategic initiative. I will focus on the types of client problemsthat are being attacked as well as the types of activities that aretaking place. I will also discuss the disciplines that being combinedin his activity as well as some of the major challenges we face.
]]>SPEAKER: Professor Ben Van Roy
ABSTRACT:
When used to guide decisions, linear regression analysis typically involves estimation of regression coefficients via ordinary least squares and their subsequent use in an optimization problem. When features are not chosen perfectly, it can be beneficial to account for the decision objective when computing regression coefficients. Empirical optimization does so but sacrifices performance when features are well-chosen or training data are insufficient. We propose directed regression, an efficient algorithm that combines merits of ordinary least squares and empirical optimization. We demonstrate through computational studies that directed regression generates performance gains over either alternative. We also develop a theory that motivates the algorithm.
]]>SPEAKER: Professor Yinyu Ye
ABSTRACT:
A natural optimization model that formulates many online resource allocation and revenue
management problems is the online linear program (LP) where the constraint matrix is revealed
column by column along with the objective function. We provide a near-optimal algorithm for
this surprisingly general class of online problems under the assumption of random order of arrivaland some mild conditions on the size of the LP right-hand-side input. Our learning-based algorithm works by dynamically updating a threshold price vector at geometric time intervals, wherethe dual prices learned from revealed columns in the previous period are used to determine thesequential decisions in the current period. Our algorithm has a feature of learning by doing, and the prices are updated at a carefully chosen pace that is neither too fast nor too slow. Inparticular, our algorithm doesn't assume any distribution information on the input itself, thusis robust to data uncertainty and variations due to its dynamic learning capability. Applicationsof our algorithm include many online multi-resource allocation and multi-product revenue managementproblems such as online routing and packing, online combinatorial auctions, adwordsmatching, inventory control and yield management.
Joint work with Shipra Agrawal and Zizhuo Wang.
Abstract
(To be announced)
For program and registration information, visit the conference website at www.scl.gatech.edu/humlog2010/.
The conference will bring together speakers and attendees from non-government organizations (NGOs), industry, government, military, and academia focusing on various topics, including planning, preparing, and responding to disasters; recovery and mitigation; addressing major societal problems related to health; and short- and long-term humanitarian response. The main objectives of the conference are to articulate the opportunities and challenges in humanitarian response and world health, both from a humanitarian and a corporate/economic perspective, to identify important research issues, to create academic awareness for the research opportunities, and to establish priorities for NGOs, corporations, and the government in terms of their strategies, policies, and investments.
The Conference is hosted by the Center for Humanitarian Logistics (http://www.scl.gatech.edu/research/humanitarian/), which is under the umbrella of Georgia Tech's Supply Chain and Logistics Institute (SCL).
To access presentation slides, speaker biographies, and other conference information for the 2009 Conference, visit: http://www2.isye.gatech.edu/humlog09/
]]>Abstract
Our focus is on the development of provably efficient simulation algorithms for estimating large deviations probabilities (such as overflow probabilities) in the context of many server queues. These types of systems, which have been the subject of much investigation in recent years, pose interesting challenges from a rare event simulation standpoint, given their measure valued state descriptor. We shall explain a technique that has the following elements. First, it introduces a pivotal set that is suitable chosen to deal with boundary-type behavior, which is common in the analysis of queueing systems. Second, it takes advantage of Central Limit Theorem approximations that have been developed recently for these types of systems and third it use a novel bridge-sampling approach in order to describe an symptotically optimal (in certain sense) importance sampling scheme. This work provides the first systematic approach to develop provably efficient rare-event simulation methodology for these types of systems.
Bio
Jose Blanchet is a faculty member of the IEOR at Columbia University. Jose holds a Ph.D. in Management Science and Engineering from Stanford University. Prior to joining Columbia he was a faculty member in the Statistics Department at Harvard University. Jose is a recipient of the 2009 Best Publication Award given by the INFORMS Applied Probability Society and a CAREER award in Operations Research given by NSF in 2008. He worked as an analyst in Protego Financial Advisors, a leading investment bank in Mexico. He has research interests in applied probability and Monte Carlo methods. He serves in the editorial board of Advances in Applied Probability, Journal of Applied Probability, QUESTA and TOMACS.
Abstract
(To be announced)
Abstract
When used to guide decisions, linear regression analysis typically involves estimation of regression coefficients via ordinary least squares and their subsequent use in an optimization problem. When features are not chosen perfectly, it can be beneficial to account for the decision objective when computing regression coefficients. Empirical optimization does so but sacrifices performance when features are well-chosen or training data are insufficient. We propose directed regression, an efficient algorithm that combines merits of ordinary least squares and empirical optimization. We demonstrate through computational studies that directed regression generates performance gains over either alternative. We also develop a theory that motivates the algorithm.
Abstract
Motivated by applications in financial services, we consider the following customized pricing problem. A seller of some good or service (like auto loans or small business loans) confronts a sequence of potential customers numbered 1, 2, … , T. These customers are drawn at random from a population characterized by a price-response function Ï(p). That is, if the seller offers price p, then the probability of a successful sale is Ï(p). The profit realized from a successful sale is Ï€(p) = p − c, where c > 0 is known.
If the price-response function Ï(-) were also known, then the problem of finding a price p* to maximize Ï(p)Ï€(p) would be simple, and the seller would offer price p* to each of the T customers. We consider the more complicated case where Ï(-) is fixed but initially unknown: roughly speaking, the seller wants to choose prices sequentially so as to maximize the total profit earned from the T potential customers; each successive choice involves a trade-off between refined estimation of the unknown price-response function (learning) and immediate profit (earning).
* Joint work with Bora Keskin and Assaf Zeevi
]]>Abstract
(To be announced)
Abstract
(To be announced)
The H. Milton Stewart Advisory Board is comprised of distinguished professionals and community leaders, and serve as a sounding board for the school chair and assist with the School's development goals.
The meeting will be held at the Georgia Tech Hotel and Conference Center from 11:00 am - 5:00 pm with a working lunch and dinner to follow
]]>SPEAKER: Dr. Jose Blanchet
ABSTRACT:
Our focus is on the development of provably efficient simulation algorithms for estimating large deviations probabilities (such as overflow probabilities) in the context of many server queues. These types of systems, which have been the subject of much investigation in recent years, pose interesting challenges from a rare event simulation standpoint, given their measure valued state descriptor. We shall explain a technique that has the following elements. First, it introduces a pivotal set that is suitable chosen to deal with boundary-type behavior, which is common in the analysis of queueing systems. Second, it takes advantage of Central Limit Theorem approximations that have been developed recently for these types of systems and third it use a novel bridge-sampling approach in order to describe an asymptotically optimal (in certain sense) importance sampling scheme. This work provides the first systematic approach to develop provably efficient rare-event simulation methodology for these types of systems.
(Joint work with P. Glynn and H. Lam)
Bio:
Jose Blanchet is a faculty member of the IEOR at Columbia University. Jose holds a Ph.D. in Management Science and Engineering from Stanford University. Prior to joining Columbia he was a faculty member in the Statistics Department at Harvard University. Jose is a recipient of the 2009 Best Publication Award given by the INFORMS Applied Probability Society and a CAREER award in Operations Research given by NSF in 2008. He worked as an analyst in Protego Financial Advisors, a leading investment bank in Mexico. He has research interests in applied probability and Monte Carlo methods. He serves in the editorial board of Advances in Applied Probability, Journal of Applied Probability, QUESTA and TOMACS.
The H. Milton Stewart Advisory Board is comprised of distinguished professionals and community leaders, and serve as a sounding board for the school chair and assist with the School's development goals.
]]>SPEAKER: Sheldon M. Ross
ABSTRACT:
Suppose there are r gamblers, with gambler i initially having a fortune of ni. In our first model we suppose that at each stage two of the gamblers are chosen to play a game, equally likely to be won by either player, with the winner of the game receiving 1 from the loser. Any gambler whose fortune becomes 0 leaves, and this continues until there is only a single gambler left. We are interested in the probability that player i is the one left, and in the the mean number of games played between specified players i and j. In our second model we suppose that all remaining players contribute 1 to a pot, which is equally likely to be won by each of them. The problem here is to determine the expected number of games played until one player has all the
funds.
SPEAKER: Associate Professor SagarKamarthi
ABSTRACT:
This talk presents a vision for holistic product lifecyclemanagement via embedded sensors. The main goal of the proposed productmonitoring framework is to cut down the product repair/service costs, boost therecovery of end-of-life products, and curtail the product disposal rate. Thetalk focuses on personal computers to demonstrate the approach; however theconcepts and methods presented herein are equally applicable to a wide varietyof consumer products. It is in this context that the talk presents thetechnical details of the hard disk drive prognostics and a novel featureextraction method through discrete wavelet transforms.
]]>William N and Marie A Professor of Management Sciences,
Yale School of Management;
Professor of Public Health,
Yale School of Public Health;
Professor of Engineering,
Yale School of Engineering and Applied Science
Abstract
This article presents the first model developed specifically for understanding the infiltration and interdiction of ongoing terror plots by undercover intelligence agents, and does so via novel application of ideas from queueing theory and Markov population processes. The resulting "terror queue" models predict the number of undetected terror threats in an area from agent activity/utilization data, and also estimate the rate with which such threats can be interdicted. The models treat terror plots as customers and intelligence agents as servers. Agents spend all of their time either detecting and infiltrating new terror plots (in which case they are "available"), or interdicting already detected terror plots (in which case they are "busy"). Initially we examine a Markov model assuming that intelligence agents, while unable to detect all plots, never err by falsely detecting fake plots. While this model can be solved numerically, a simpler Ornstein-Uhlenbeck diffusion approximation yields some results in closed form while providing nearly identical numerical performance. The transient behavior of the terror queue model is discussed briefly along with a sample sensitivity analysis to study how model predictions
compare to simulated results when using estimated versus known terror plot arrival rates. The diffusion model is then extended to allow for the false detection of fake plots. Such false detection is a real feature of counterterror intelligence given that intelligence agents or informants can make mistakes, as well as the proclivity of terrorists to deliberately broadcast false information. The false detection model is illustrated using suicide bombing data from Israel.
Bio
Professor Kaplan's research has been reported on the front pages of the New York Times and the Jerusalem Post, editorialized in the Wall Street Journal (article), recognized by the New York Times Magazine's Year in Ideas, and discussed between the covers of Time (article), Newsweek (see article), US News and World Report, Consumer Reports and the New Yorker, and in person on NBC's Today Show, the Cronkite Report, and National Public Radio
(transcript).
The author of more than 100 research articles, Professor Kaplan received both the Lanchester Prize and the Edelman Award, the two top honors in the operations research field. An elected member of both the National Academy of Engineering and the Institute of Medicine of the US National Academies, he has also twice received the prestigious Lady Davis Visiting Professorship at the Hebrew University of Jerusalem, where he has investigated AIDS policy issues facing the State of Israel.
Kaplan’s current research focuses on the application of operations research to problems in counterterrorism and
homeland security.
Abstract:
Familiar statistical estimates such as batting averages, political polls, and medical trial results are obtained by direct observation of cases of interest. Sometimes, though, we can learn from the experience of "others": for instance there may be information about Player A's batting ability in the observed averages of Players B, C, and D. In his presentation, Professor Efron will present several examples showing how this works in practice, indicating some of the surprising theoretical ideas involved. The talk is mainly descriptive in nature, and is intended for a general scientific audience.
About Dr. Efron:
Bradley Efron is the Max H. Stein Professor of Statistics andBiostatistics at Stanford University's School of Humanities and Sciences and the Department of Health Research and Policy with the School of Medicine. He completed his undergraduate work in mathematics at the California Institute of Technology, and earned his doctorate in statistics from Stanford in 1964, joining the Stanford faculty that same year. He was Associate Dean for the School of Humanities and Sciences from 1987 to 1990, served a term as Chair of the Faculty Senate as well as three terms as Chair of the Department of Statistics, and continues as Chairman of the Mathematical and Computational Sciences Program. He has served as president of the American Statistical Association and ofthe Institute of Mathematical Statistics. He is a past editor of the Journal of the American Statistical Association and is presently the founding editor of the Annals of Applied Statistics.
Among the numerous honors that Efron has received are Fellowships of the American Academy of Arts and Sciences, the American Statistical Association, the Institute of Mathematical Statistics, the Royal Statistical Society, the International Statistical Institute and the MacArthur Fellows Program of the John D. and Catherine T. MacArthur Foundation. He is a member of the U.S. National Academy of Sciences, a recipient of the Ford Prize of the Mathematical Association of America and of both the Wilks Medal and the Noether Prize of the American Statistical Association. Efron was awarded the 1998 Parzen Prize for Statistical Innovation by Texas A&M University, and the first-ever Rao Prize for outstanding research in statistics by Pennsylvania State University in 2003. He received the 2005 National Medal of Science "for his contributions to theoretical and applied statistics, especially the bootstrap sampling technique; for his extraordinary geometric insight into nonlinear statistical problems; and for applications in medicine, physics and astronomy."
]]>SPEAKER: Vijay Nair
ABSTRACT:
Traditional reliability analysis is based on time-to-failure data. In this talk, we discuss two different directions that lead to more informative reliability inference. The first is multi-state models. Challenges in statistical inference based on interval-censored data will be described, and methods for doing likelihood-based inference in semi-Markov models will be discussed. The second part of the talk will provide an overview of degradation models, describe a class of non-homogeneous Gaussian processes and some inference issues. This talk is based on joint work with Yang Yang, Yves Atchade and Xiao Wang.
*Bio*: Vijay Nair is the Donald A. Darling Professor of Statistics and Professor of Industrial and Operations Engineering at the University of Michigan, Ann Arbor. He has been Chair of the Department of Statistics sine 1998. Previously, he was a Research Scientist in the Mathematical Sciences and Operations Research Centers at Bell Laboratories for fifteen years.
His research interests include engineering statistics, information technology, quality and process improvement, industrial experiments, reliability and risk analysis, neuro-informatics, and statistical methods in behavioral intervention research. He also has extensive practical experience in the automotive, semiconductor, and telecommunications industries.
Vijay has a Bachelor’s degree in Economics from the University of Malaya (Malaysia) and a Ph. D. in Statistics from the University of California, Berkeley. He has been editor of Technometrtics, International Statistical Review, coordinating editor of the Journal of Statistical Planning and Inference, and has served on the editorial boards of many other leading statistics and quality journals. He is currently Vice President of the International Statistics Institute, President-elect of the International Society for Business and Industrial Statistics, and Chair of the Statistics Division of the American Society for Quality. He has chaired or co-chaired several panels of the National Academies and is a former chair of the Board of Trustees of the National Institute of Statistical Sciences. He is a Fellow of the American Association for the Advancement of Science, the American Statistical Association, the American Society for Quality, and the Institute of Mathematical Statistics.