Thesis: Ph. D., Massachusetts Institute of Technology, Sloan School of Management, Operations Research Center, 2018. / Cataloged from PDF version of thesis. / Includes bibliographical references (pages 191-205). / The term personalization typically refers to the activity of online recommender systems, and while product and content personalization is now ubiquitous in e-commerce, systems today remain relatively primitive: they are built on a small fraction of available data, run with heuristic algorithms, and restricted to e-commerce applications. This thesis addresses key challenges and new applications for modern, large-scale personalization. In particular, this thesis is outlined as follows: First, we formulate a generic, flexible framework for learning from matrix-valued data, including the kinds of data commonly collected in e-commerce. Underlying this framework is a classic de-noising problem called tensor recovery, for which we provide an efficient algorithm, called Slice Learning, that is practical for massive datasets. Further, we establish near-optimal recovery guarantees that represent an order improvement over the best available results for this problem. Experimental results from a music recommendation platform are shown. Second, we apply this de-noising framework to new applications in precision medicine where data are routinely complex and in high dimensions. We describe a simple, accurate proteomic blood test (a 'liquid biopsy') for cancer detection that relies on de-noising via the Slice Learning algorithm. Experiments on plasma from healthy patients that were later diagnosed with cancer demonstrate that our test achieves diagnostically significant sensitivities and specificities for many types of cancers in their earliest stages. Third, we present an efficient, principled approach to operationalizing recommendations, i.e. the decision of exactly what items to recommend. Motivated by settings such as online advertising where the space of items is massive and recommendations must be made in milliseconds, we propose an algorithm that simultaneously achieves two important properties: (1) sublinear runtime and (2) a constant-factor guarantee under a wide class of choice models. Our algorithm relies on a new sublinear time sampling scheme, which we develop to solve a class of problems that subsumes the classic nearest neighbor problem. Results from a massive online content recommendation firm are given. Fourth, we address the problem of cost-effectively executing a broad class of computations on commercial cloud computing platforms, including the computations typically done in personalization. We formulate this as a resource allocation problem and introduce a new approach to modeling uncertainty - the Data-Driven Prophet Model - that treads the line between stochastic and adversarial modeling, and is amenable to the common situation where stochastic modeling is challenging, despite the availability of copious historical data. We propose a simple, scalable algorithm that is shown to be order-optimal in this setting. Results from experiments on a commercial cloud platform are shown. / by Andrew A. Li. / Ph. D.

Thesis: Ph. D., Massachusetts Institute of Technology, Sloan School of Management, Operations Research Center, 2018. / Cataloged from PDF version of thesis. / Includes bibliographical references (pages 65-70). / In retail, there are predictable yet dramatic time-dependent patterns in customer behavior, such as periodic changes in the number of visitors, or increases in customers just before major holidays. The standard paradigm of multi-armed bandit analysis does not take these known patterns into account. This means that for applications in retail, where prices are fixed for periods of time, current bandit algorithms will not suffice. This work provides a framework and methods that take the time-dependent patterns into account. In the corrected methods, exploitation (greed) is regulated over time, so that more exploitation occurs during higher reward periods, and more exploration occurs in periods of low reward. In order to understand why regret is reduced with the corrected methods, a set of bounds on the expected regret are presented and insights into why we would want to exploit during periods of high reward are discussed. When the set of available options changes over time, mortal bandits algorithms have proven to be extremely useful in a number of settings, for example, for providing news article recommendations, or running automated online advertising campaigns. Previous work on this problem showed how to regulate exploration of new arms when they have recently appeared, but they do not adapt when the arms are about to disappear. Since in most applications we can determine either exactly or approximately when arms will disappear, we can leverage this information to improve performance: we should not be exploring arms that are about to disappear. Also for this framework, adaptations of algorithms and regret bounds are provided. The proposed methods perform well in experiments, and were inspired by a high-scoring entry in the Exploration and Exploitation 3 contest using data from Yahoo! Front Page. That entry heavily used time-series methods to regulate greed over time, which was substantially more effective than other contextual bandit methods. / by Stefano Tracà. / Ph. D.

Thesis (Ph. D.)--Massachusetts Institute of Technology, Sloan School of Management, Operations Research Center, 2006. / Includes bibliographical references (p. 271-284). / Recently, revenue management has become popular in many industries such as the airline, the supply chain, and the transportation industry. Decision makers realize that even small improvements in their operations can have a significant impact on their profits. Nevertheless, determining pricing and inventory optimal policies in more realistic settings may not be a tractable task. Ignoring the potential inaccuracy of parameters may lead to a solution that actually performs poorly, or even that violates some constraints. Finally, competitors impact a supplier's best strategy by influencing her demand, revenues, and field of possible actions. Taking a game theoretic approach and determining the equilibrium of the system can help understand its state in the long run. This thesis presents a continuous time optimal control model for studying a dynamic pricing and inventory control problem in a make-to-stock manufacturing system. We consider a multi-product capacitated, dynamic setting. We introduce a demand-based model with convex costs. A key part of the model is that no backorders are allowed, as this introduces a constraint on the state variables. We first study the deterministic version of this problem. / (cont.) We introduce and study a solution method that enables to compute the optimal solution on a finite time horizon in a monopoly setting. Our results illustrate the role of capacity and the effects of the dynamic nature of demand. We then introduce an additive model of demand uncertainty. We use a robust optimization approach to protect the solution against data uncertainty in a tractable manner, and without imposing stringent assumptions on available information. We show that the robust formulation is of the same order of complexity as the deterministic problem and demonstrate how to adapt solution method. Finally, we consider a duopoly setting and use a more general model of additive and multiplicative demand uncertainty. We formulate the robust problem as a coupled constraint differential game. Using a quasi-variational inequality reformulation, we prove the existence of Nash equilibria in continuous time and study issues of uniqueness. Finally, we introduce a relaxation-type algorithm and prove its convergence to a particular Nash equilibrium (normalized Nash equilibrium) in discrete time. / by Elodie Adida. / Ph.D.

Thesis: Ph. D., Massachusetts Institute of Technology, Sloan School of Management, Operations Research Center, 2016. / This electronic version was submitted by the student author. The certified thesis is available in the Institute Archives and Special Collections. / Cataloged from student-submitted PDF version of thesis. / Includes bibliographical references (pages 153-156). / The supply chain management for an online retailing business is centered around the operations of its fulfillment centers. A fulfillment center receives and holds inventory from vendors, and then uses this inventory to fill customer orders. Our research focuses on a new operating architecture of an order fulfillment system, enabled by new technology. We refer to it as the Semi-automated Fulfillment System. Different from the person-to-goods model in traditional warehouses, the semi-automated fulfillment system adopts a goods-to-person model for stowing and picking items from a storage field. In a semi-automated fulfillment system the inventory is stored on mobile storage pods; those mobile pods are then carried by robotic drives to static stations at which the operators conduct pick or stow operations. In the first chapter, we describe and identify three key operational decisions in the semiautomated fulfillment system, namely from which pods to pick the inventory needed (picking decision), where to return the pod to the storage field upon the completion of a pick or stow operation (storage decision), and to which pods to replenish the received inventory (stowage decision). We present a high-level capacity planning model for determining the number of robotic drives needed to achieve a given throughput level. This model highlights how the operational efficiency in this system depends on two key parameters, namely the travel time for an entire drive trip and the number of unit picks or stows per pod trip. In the second chapter, we focus on the storage decisions. The storage decision is to decide to which storage location to return a pod upon the completion of a pick or stow operation. We extend the academic results on the benefits of adopting velocity-based and class-based storage policies to the context of the semi-automated fulfillment system. We associate with each storage pod a velocity measure that represents an expectation of the number of picks from that pod in the near future. We then show that by assigning the high velocity pods to the most desirable storage locations, we can significantly reduce the drive travel time, compared to the random storage policy that returns the pod to a randomly-chosen storage location. We show that class-based storage policies with two or three classes, can achieve most of the benefits from the idealized velocity-based policy. Furthermore, we characterize how the performance of the velocity-based and class-based storage policies depend on the velocity variability across the storage pods; in particular we model how the benefits from velocity-based storage policies increase with increased variation in the pod velocities. In the third chapter, we build a discrete-time simulator to validate the theoretical models in the second chapter with real industry data. We observe a 6% to 11% reduction in the travel distance with 2-class or 3-class system, depending on the parameter settings. From a sensitivity analysis we establish the robustness of the class-based storage policies as they continue to perform well under a broad range of warehouse settings including different zoning strategies, resource utilization levels and space utilization levels. In the fourth chapter, we examine two stowage decisions, one at the zone level and the other at the pod level. The zone-level decision is to decide how to allocate the received inventory to multiple storage zones. The objective is to assure that the resulting picking workload for each zone is within its capacity. We show by simulation that a chaining-based allocation can be effective to balance the picking workload across different storage zones. The pod-level stowage decision is to decide on which pods to stow the inventory. We formulate a mixedinteger program (MIP) to find the optimal stowage profile that maximizes the number of unit picks per pod trip. We solve the MIP for a set of test cases to gain insight into the structure of optimal stowage policy. Motivated by these insights, we further propose a class-based stowage process that induces variability across the pod velocities. / by Rong Yuan. / Ph. D.

Thesis: Ph. D., Massachusetts Institute of Technology, Sloan School of Management, Operations Research Center, 2015. / This electronic version was submitted by the student author. The certified thesis is available in the Institute Archives and Special Collections. / Cataloged from student-submitted PDF version of thesis. / Includes bibliographical references (pages 133-139). / Financing drug development has a particular set of challenges including long development times, high chance of failure, significant market valuation uncertainty, and high costs of development. The earliest stages of translational research pose the greatest risks, which have been termed the "valley of death" as a result of a lack of funding. This thesis focuses on an exploration of financial engineering techniques aimed at addressing these concerns. Despite the recent financial crisis, many suggest that securitization is an appropriate tool for financing such large social challenges. Although securitization has been demonstrated effectively at later stages of drug development for drug royalties of approved drugs, it has yet to be utilized at earlier stages. This thesis starts by extending the model of drug development proposed by Fernandez et al. (2012). These extensions significantly influence the resulting performance and optimal securitization structures. Budget-constrained venture firms targeting high financial returns are incentivized to fund only the best projects, thereby potentially stranding less-attractive projects. Instead, such projects have the potential to be combined in larger portfolios through techniques such as securitization which reduce the cost of capital. In addition to modeling extensions, we provide examples of a model calibrated to orphan drugs, which we argue are particularly suited to financial engineering techniques. Using this model, we highlight the impact of our extensions on financial performance and compare with previously published results. We then illustrate the impact of incorporating a credit enhancement or guarantee, which allows for added flexibility of the capital structure and therefore greater access to lower costing capital. As an alternative to securitization, we provide some examples of a structured equity approach, which may allow for increased access to or efficiency of capital by matching investor objectives. Finally, we provide examples of optimizing the Sortino ratio through constrained Bayesian optimization. / by David Erik Fagnan. / Ph. D.

Thesis: Ph. D., Massachusetts Institute of Technology, Sloan School of Management, Operations Research Center, 2018. / This electronic version was submitted by the student author. The certified thesis is available in the Institute Archives and Special Collections. / Cataloged from student-submitted PDF version of thesis. / Includes bibliographical references (pages 165-173). / Recent advances in machine learning and optimization hold much promise for influencing real-world decision making, especially in areas such as health care where abundant data are increasingly being collected. However, imperfections in the data pose a major challenge to realizing their full potential: missing values, noisy observations, and unobserved counterfactuals all impact the performance of data-driven methods. In this thesis, with a fresh perspective from optimization, I revisit some of the well-known problems in statistics and machine learning, and develop new methods for prescriptive analytics. I show examples of how common machine learning tasks, such as missing data imputation in Chapter 2 and classication in Chapter 3, can benet from the added edge of rigorous optimization formulations and solution techniques. In particular, the proposed opt.impute algorithm improves imputation quality by 13.7% over state-of-the-art methods, as averaged over 95 real data sets, which leads to further performance gains in downstream tasks. The power of prescriptive analytics is shown in Chapter 4 by our approach to personalized diabetes management, which identifies response patterns using machine learning and individualizes treatments via optimization. These newly developed machine learning algorithms not only demonstrate improved performance in large-scale experiments, but are also applied to solve the problems in health care that motivated them. Our simulated trial for diabetic patients in Chapter 4 demonstrates a clinically relevant reduction in average hemoglobin A1c levels compared to current practice. Finally, when predicting mortality for cancer patients in Chapter 5, applying opt.impute on missing data along with the cutting-edge algorithm Optimal Classication Tree on a rich data set prepared from electronic medical records, we are able to accurately risk stratify patients, providing physicians with interpretable insights and valuable risk estimates at time of treatment decisions and end-of-life planning. / by Ying Daisy Zhuo. / Ph. D.

Thesis (S.M.)--Massachusetts Institute of Technology, Sloan School of Management, Operations Research Center, 2002. / Includes bibliographical references (p. 57-58). / This thesis presents a Markov chain model of the transmission of financial crises. Using bilateral trade data and a measure of exchange market pressure, it develops a method to determine a set of transition probabilities that describe the crisis transmission dynamics. The dynamics are characterized by one month conditional crisis probabilities and the probability of a crisis occurring within one year. Calculations of the transition probabilities for a three country example suggest that minor trading partners can increase the likelihood of a crisis in the home country through their effect on major trading partners. / by Kevin Amonlirdviman. / S.M.

Thesis (Ph. D.)--Massachusetts Institute of Technology, Sloan School of Management, Operations Research Center, 2009. / This electronic version was submitted by the student author. The certified thesis is available in the Institute Archives and Special Collections. / Cataloged from student-submitted PDF version of thesis. / Includes bibliographical references (p. 108-112). / Classical methods of maximum likelihood and least squares rely a great deal on the correctness of the model assumptions. Since these assumptions are only approximations of reality, many robust statistical methods have been developed to produce estimators that are robust against the deviation from the model assumptions. Unfortunately, these techniques have very high computational complexity that prevents their application to large scale problems. We present computationally efficient methods for robust mean-covariance estimation and robust linear regression using special mathematical programming models and semi-definite programming (SDP). In the robust covariance estimation problem, we design an optimization model with a loss function on the weighted Mahalanobis distances and show that the problem is equivalent to a system of equations and can be solved using the Newton-Raphson method. The problem can also be transformed into an SDP problem from which we can flexibly incorporate prior beliefs into the estimators without much increase in the computational complexity. The robust regression problem is often formulated as the least trimmed squares (LTS) regression problem where we want to nd the best subset of observations with the smallest sum of squared residuals. We show the LTS problem is equivalent to a concave minimization problem, which is very hard to solve. We resolve this difficulty by introducing the maximum trimmed squares" problem that finds the worst subset of observations. This problem can be transformed into an SDP problem that can be solved efficiently while still guaranteeing that we can identify outliers. / (cont.) In addition, we model the robust ranking problem as a mixed integer minimax problem where the ranking is in a discrete uncertainty set. We use mixed integer programming methods, specifically column generation and network flows, to solve the robust ranking problem. To illustrate the power of these robust methods, we apply them to the mean-variance portfolio optimization problem in order to incorporate estimation errors into the model. / by Tri-Dung Nguyen. / Ph.D.

Thesis (Ph. D.)--Massachusetts Institute of Technology, Sloan School of Management, Operations Research Center, 2008. / Includes bibliographical references (p. 123-130). / Large population health insurance claims databases together with operations research and data mining methods have the potential of significantly impacting health care management. In this thesis we research how claims data can be utilized in three important areas of health care and medicine and apply our methods to a real claims database containing information of over two million health plan members. First, we develop forecasting models for health care costs that outperform previous results. Secondly, through examples we demonstrate how large-scale databases and advanced clustering algorithms can lead to discovery of medical knowledge. Lastly, we build a mathematical framework for a real-time drug surveillance system, and demonstrate with real data that side effects can be discovered faster than with the current post-marketing surveillance system. / by Margrét Vilborg Bjarnadóttir. / Ph.D.

Thesis (S.M.)--Massachusetts Institute of Technology, Sloan School of Management, Operations Research Center, 2010. / Cataloged from PDF version of thesis. / Includes bibliographical references (p. 123-124). / Combat outposts (COPs) are small, well-protected bases from which soldiers reside and conduct operations from. Used extensively during the "Surge" in Iraq, COPs are usually established in populated areas and are prevalent in the counterinsurgency operations in Afghanistan in 2010. This research models population security to determine combat outpost locations in a battalion area of operation. Population security is measured by level of violence, level of insurgent activity, and effectiveness of host nation security forces. The area of operation is represented as a graphical network of nodes and arcs. Operational inputs include pertinent information about each node. The model allows the commander to set various weights that reflect his understanding of the situation, mission, and local people. Based on trade-offs in patrolling and self-protection, the deterministic model recommends the size and locations for emplacing combat outposts and conducting patrols. We use piecewise linear approximation to solve the problem as a mixed-integer linear program. Results are based on two representative scenarios and show the impact of an area of operation's characteristics and commander's weights on COP size and locations. / by Scott B. Seidel. / S.M.