Chicago Operations Workshop 2025
Monday, June 9th at Gleacher Center, University of Chicago Booth School of Business
This is the second Chicago Operations Workshop, organized to bring together operations management researchers and PhD students from the Chicago area. Looking ahead, we plan to continue expanding participation to include more departments across the Chicago metropolitan area and the Midwest.
This year, the workshop featured presenters from the University of Chicago, the University of Illinois at Chicago, Northwestern University, and Purdue University.
Chicago Operations Workshop 2025 Details and Registration
Date:
Monday June 9th, 2025
Location:
Room 300
Gleacher Center, University of Chicago Booth School of Business
450 Cityfront Plaza Dr, Chicago, IL 60611
Organizing Committee:
- Rene Caldentey, Booth School of Business, University of Chicago
- Itai Gurvich, Kellogg School of Management, Northwestern University
- Selva Nadarajah, College of Business Administration, University of Illinois Chicago
Registration:
Deadline to register: May 30th, 2025
Register for the Chicago Operations Workshop
Workshop Schedule
Breakfast and Registration: 9:00-9:50am
Introduction: 9:50:10:00am
Session 1: 10:00am – 12:00pm
Speaker: Brad Sturt, Assistant Professor, University of Illinois Chicago
Title: Improving the Security of United States Elections with Robust Optimization
Abstract: For more than a century, election officials across the United States have inspected voting machines before elections using a procedure called Logic and Accuracy Testing (LAT). This procedure consists of election officials casting a test deck of ballots into each voting machine and confirming the machine produces the expected vote total for each candidate. In this talk, I will bring a scientific perspective to LAT by introducing the first formal approach to designing test decks with rigorous security guarantees. Specifically, we propose using robust optimization to find test decks that are guaranteed to detect any voting machine misconfiguration that would cause votes to be swapped across candidates. Out of all the test decks with this security guarantee, the robust optimization problem yields the test deck with the minimum number of ballots, thereby minimizing implementation costs for election officials. To facilitate deployment at scale, we developed a practical exact algorithm for solving our robust optimization problems based on mixed-integer optimization and the cutting plane method.
In partnership with the Michigan Bureau of Elections, we retrospectively applied our robust optimization approach to all 6928 ballot styles from Michigan's November 2022 general election; this retrospective study reveals that the test decks with rigorous security guarantees obtained by our approach require, on average, only 1.2% more ballots than current practice. Our robust optimization approach has since been piloted in real-world elections by the Michigan Bureau of Elections as a low-cost way to improve election security and increase public trust in democratic institutions.
Speaker: Bahar Taşkesen, Assistant Professor, University of Chicago Booth School of Business
Title: Statistical Equilibrium of Optimistic Beliefs
Abstract: We introduce the Statistical Equilibrium of Optimistic Beliefs (SE-OB) for the mixed extension of finite normal-form games, drawing insights from discrete choice theory. Departing from the conventional best responders of Nash equilibrium and the better responders of quantal response equilibrium, we reconceptualize player behavior as that of optimistic better responders. In this setting, the players assume that their expected payoffs are subject to random perturbations, and form optimistic beliefs by selecting the distribution of perturbations that maximizes their highest anticipated payoffs among belief sets. In doing so, SE-OB subsumes and extends the existing equilibria concepts. The player's view of the existence of perturbations in their payoffs reflects an inherent risk sensitivity, and thus, each player is equipped with a risk-preference function for every action. We demonstrate that every Nash equilibrium of a game, where expected payoffs are regularized with the risk-preference functions of the players, corresponds to an SE-OB in the original game, provided that the belief sets coincide with the feasible set of a multi-marginal optimal transport problem with marginals determined by risk-preference functions. Building on this connection, we propose an algorithm for repeated games among risk-sensitive players under optimistic beliefs when only zeroth-order feedback is available. We prove that, under appropriate conditions, the algorithm converges to an SE-OB. Our convergence analysis offers key insights into the strategic behaviors for equilibrium attainment: a player's risk sensitivity enhances equilibrium stability, while forming optimistic beliefs in the face of ambiguity helps to mitigate overly aggressive strategies over time. As a byproduct, our approach delivers the first generic convergent algorithm for general-form structural QRE beyond the classical logit-QRE.
Speaker: Daniela Hurtado-Lange, Assistant Professor, Kellogg, Northwestern University
Title: The transform Method for Markov-Modulated Queues
Abstract: The Transform Method has been developed in the last four years for heavy-traffic analysis in a variety of queueing systems, including the single-server queue, load-balancing system, input-queued switch, generalized switch under complete resource pooling, and ride-sharing. These are studied in discrete time under i.i.d. arrival and potential service process. In this talk, I will introduce a generalization to the Transform Method, where we allow a Markov-modulated arrival and service process. The key is in carefully designing the test function, which includes the queue length and a correction due to the Markov-modulated nature of the arrival process. In this talk, I will present a simple case, but the methodology can be generalized to more complex queueing networks.
Speaker: Marty Lariviere, Professor, Kellogg, Northwestern University
Title: Who should use the self-checkout lane? A Slow Server Problem with Two Classes
Abstract: Over half of the transactions at American supermarkets are at self-checkout lanes. We examine which customers should use the self-checkout lanes and when.
Lunch: 12:00 – 1:30pm, Room 350 Lounge
Session 2: 1:30 – 3:00pm
Speaker: Levi DeValve, Assistant Professor, University of Chicago Booth School of Business
Title: Managing E-commerce Fulfillment Networks: Subset Selection and Approximate Submodularity
Abstract: Many challenging problems in managing e-commerce fulfillment networks can be posed as subset selection problems: which fulfillment arcs to select during the network design phase, where to place each item (SKU) in the network during the inventory planning phase, and which distribution centers to use for each order in the fulfillment phase. These impactful problems are computationally challenging due to their combinatorial structure, and both the academic and practitioner communities have recognized the need to develop effective heuristics. The academic literature has long understood submodularity to be an invaluable structural property for analyzing subset selection problems, for the intuitive reason that local changes in the objective can be used to bound global changes. Unfortunately, most subset selection problems arising in the e-commerce setting lack submodularity, due to the inherent ``two-sided” nature of the supply and demand networks under consideration. We overcome this technical challenge using a novel form of approximate submodularity to analyze local search heuristics for these problems, proposing a general framework that provides new constant factor approximation guarantees. In particular, we prove approximation guarantees of: (1-exp(-2))/2 (approx. 0.432) for cardinality constrained network design, 1/3 for network item placement, and (1-exp(-(1-exp(-1)))/2 (approx. 0.234) for order fulfillment problems. Further, numerical simulations guided by our worst-case analysis consistently provide near optimal performance, and our primal-dual analysis leads to efficient numerical implementations that significantly accelerate local search algorithms, a critical feature for deployment at scale required by modern e-commerce applications.
Speaker: Yichen Zhang, Assistant Professor, Purdue University Mitch Daniels School of Business
Title: Managing Inventory and Information in Supply Chains
Abstract: This work integrates time-series forecasting with supply chain management to evaluate the value of information sharing in a two-tier supply chain consisting of a single retailer and a single manufacturer. Surprisingly, we demonstrate that in a decentralized system, information sharing provides no value in reducing inventory related costs. The retailer's optimal replenishment policy, whether under full or no information sharing, remains invertible with respect to demand, allowing the manufacturer to recover the retailer's demand information organically through the construction of optimal forecasts. We compare the performance of an optimal inventory policy to a set of alternative benchmark policies that have been considered in the literature to study the value of order smoothing and information sharing in supply chains. Our analysis highlights the complex trade-offs in inventory replenishment policies and demonstrates that, compared to these benchmarks, a strategic retailer can benefit from inducing a controlled bullwhip effect by increasing the mean square forecast error (MSFE) of its orders to enhance the manufacturer's service level. Additionally, our analysis distinguishes between the standard bullwhip effect measure-defined as the ratio of order volatility to demand volatility-and an informationally (or forecast)-adjusted measure, defined as the ratio of the MSFE of the orders to the MSFE of the demand. We argue that the informationally adjusted bullwhip is a more appropriate measure, as it captures the propagation of volatility that cannot be forecasted away. This distinction is significant, as we demonstrate that these two measures do not always align; an inventory policy can exhibit a bullwhip effect under one measure while simultaneously dampening it under the other. This work is joint with René Caldentey (Chicago Booth) Avi Giloni (Yeshiva) and Clifford Hurvich (NYU Stern).
Speaker: Jan Van Mieghem, Professor, Kellogg, Northwestern University
Title: Inventory Control under Supply Yield Uncertainty
Abstract: I will discuss random yield to model of supply yield risk. Random yield is surprisingly subtle and complicates mathematical tractability. Even in the single-period model the optimal policy is tricky and non-linear (so order-up-to policies are typically not optimal). To the best of my knowledge, there are no multi-period optimality results published. We will proffer new dynamic results.
Poster Session: 3:00-4:00pm, Room 350
Session 3: 4:00 – 5:30pm
Speaker: Vijay Kamble, Associate Professor, University of Illinois Chicago
Title: Machine Learning Techniques for Causal Identification in Multi-task and Meta-learning
Abstract: Developing large-scale pre-trained models for basic prediction problems pervading across industries, such as demand learning, customer choice modeling, risk prediction, etc., is a natural next step, given the profound success of foundation models for language generation and computer vision. By enabling transfer learning, such models can overcome local data scarcity that hinders effective decision-making in individual organizations or decision-making units. However, there is a crucial difference between the machine-learning models used for prediction and those used for decision-making. In that, models used for decision-making must accurately capture the causal relationship between decisions and their outcomes. Several confounding factors in the data may hinder the learning of such causal relationships while using multi-task learning or meta-learning, the approaches that underlie the development of large-scale pre-trained models. In this work, we highlight some of these challenges and develop new machine-learning techniques for effective causal identification in multi-task and meta-learning. Through new theory and extensive experiments, we demonstrate that these techniques significantly reduce biases and improve causal identification compared to vanilla multi-task learning and meta-learning approaches.
Speaker: Negar Soheili, Associate Professor, University of Illinois Chicago
Title: Revisiting Model Selection for Sequential Decision-Making Approximations
Abstract: Approximating Markov decision processes (MDPs) is critical for solving large-scale sequential decision-making problems, where exact solutions are computationally infeasible. These approximations often involve trade-offs in model selection, feature representation, and solution methods. While the broader field of sequential decision-making includes diverse approaches, math-programming-based approximate dynamic programming (ADP) methods, such as approximate linear programming (ALP) and pathwise optimization (PO), serve as an illustrative case. These methods are notable for providing both policies and performance bounds by solving large-scale linear or convex programming models.
Recent advances in first-order solution methods have enhanced the solvability of these models, yet the interplay between model selection and feature representation remains underexplored. For example, with machine learning techniques like random features, both ALP and PO can achieve asymptotic optimality as the number of features increases. This raises new questions about how to select and design models effectively. To address these, we introduce a novel ALP relaxation that improves both solution quality and solvability for fixed features. When dominance between the ALP relaxation and PO is unclear, the choice of solution methods and model structure becomes critical. Notably, our ALP relaxation exhibits a separability structure, making it computationally advantageous over PO when using block-coordinate descent methods. This work highlights the need for integrated model selection strategies that consider feature architectures, model formulations, and solution methods collectively. Although illustrated using math-programming-based ADP, the insights have broader implications for approximations of large-scale MDPs in sequential decision-making contexts.
Speaker: X.Y. Han, Assistant Professor, University of Chicago Booth School of Business
Title: Using Embeddings To Understand AI Decision Making
Abstract: While many recent works in OR/OM have begun exploring /how/ to better train AI algorithms for decision making, less attention have been given to a complementary question: When we /do/ manage to train successful AI predictors, what rules did the AI actually learn? In this talk, we show how the geometry of a network’s learned embeddings could shed light on this question. We first start by describing the Neural Collapse phenomenon in image classification deep nets, where the optimality of the deep net can been seen from its embeddings converging to a Simplex Equiangular Tight Frame structure. Then, we describe new results exploring whether similar geometries emerge in the embeddings of neural net backbones of deep RL algorithms.
Register for Chicago Operations Workshop 2025
Registrations close on May 30th, 2025, please register before that time.