INFORMS 2021 Program Book

INFORMS Anaheim 2021

MP01

3 - Double Tolerance Design or a Product Family Di Liu, Clemson University, Clemson, SC, United States We consider a production process with multiple types of products that are inspected on the same quality characteristic with target values are different for each product type. We use double tolerance sets to determine which products require reworking. The nonconforming products with measurements that are between two adjacent target values are separated by a shared outer tolerance. This outer tolerance is used to determine into which product type these products should be reworked. We develop a non-linear optimization model to identify the optimal locations of the shared outer tolerances to maximize profit given the trade-off among selling prices, processing costs and quality loss costs. 4 - Integrated Guaranteed Service Approach for Multi-echelon Inventory Optimization Victoria G. Achkar, INTEC (UNL-CONICET), Santa Fe, Argentina and Facultad de Ingeniería Química (UNL), Santa Fe, Argentina, Braulio Brunaud, Rami Musa, Carlos A. Méndez, Carlos A. Méndez, Ignacio E. Grossmann The purpose of Guaranteed Service Model (GSM) is to allocate safety stocks across the network so as to reach target service levels at the lowest cost, accounting for demand uncertainty. The main contribution of this work is the development of an extended integrated approach that accounts for stochastic lead times, raw materials and manufacturing centers management, non-nested review periods, batch sizes, external demand at intermediate nodes, and non-serial network topologies. We propose different NLP and MINLP formulations and evaluate their performance by solving several illustrative ase studies. 5 - Final Production Run and Trade-in Offers for High-tech Products under Warranty Erik Bertelli, University of California-Berkeley, Berkeley, CA, United States, Candace Arai Yano High-tech products have short life cycles but may have relatively long warranty periods that a firm must consider when deciding when to end production. However, warranty claims may be satisfied by new or refurbished items. We develop a model to dynamically optimize how long to continue contracting for production by a third-party manufacturer when the firm can later make a one- time trade-in offer to owners of the product to source refurbished inventory. MB46 CC Room 213D In Person: Optimization Contributed Session Chair: Hector Perez, Pittsburgh, PA, 15217, United States 1 - Multi-armed Bandits for Short-lived High-volume Contents Su Jia, CMU, Pittsburgh, PA, United States Consider the problem of recommending short-lived, high-volume content e.g. in content aggregation platforms and platforms with user-generated contents.We collaborated with Glance, a ``zero-screen’’ content platform with over 100 million daily active users and thousands of content pieces, called ``cards’’, produced daily.To recommend these cards to users, they currently deploy a DNN -based recommendation system that does not use online feedback. We show that our policy improves upon the current system by nearly 30\% in offline simulations based on historical data. A large field experiment to test the effectiveness of our policy is now underway. 2 - Training a Single Bandit Arm Eren Ozbay, University of Illinois at Chicago, Chicago, IL, United States, Vijay Kamble Motivated by the problem of optimizing job assignments to train novice workers of unknown quality in labor platforms, we consider a new objective in the classical stochastic multi-armed bandit setup. We consider the cumulative rewards earned from K arms at the end of T pulls, and aim to maximize the expected value of the highest cumulative reward across K arms. This corresponds to the objective of training a single, highly skilled worker using a limited supply of training jobs. We show that any policy must incur an instance-dependent asymptotic regret of Ω (logT) and an instance-independent regret of Ω (K1/3T2/3). We design an explore-then-commit policy achieving these bounds up to logarithmic factors.

3 - University Course Classroom Assignment During A Pandemic Mohammad Khamechian, PhD Candidate, University of Wisconsin Milwaukee, Milwaukee, WI, United States, Matthew Petering University course classroom assignment is a challenging real-world problem. This year the pandemic made this task harder due to reduced classroom capacities needed to implement social distancing. This study investigates university course classroom assignments with limited classroom capacities and the option for more than one classroom to be used for the same course at the same time (e.g. for exam scheduling). Distances between rooms assigned to a course and distances from instructors’ offices to assigned rooms are considered. A math model is developed, coded in C++, and solved with CPLEX. 4 - A Flexible Rolling Regression Framework for Time-varying Epidemiology Models Javier Rubio-Herrero, Assistant Professor, University of North Texas, Denton, TX, United States, Yuchen Wang We present a data-driven framework for describing the time-varying nature of an epidemiology model in the context of COVID-19. By embedding a rolling regression in a mixed integer bilevel nonlinear programming problem, our aim is to introduce a model that reproduces the observed changes in the number of infected, recovered, and death cases, while providing information about the time dependency of the parameters that govern our model. We propose this optimization model and a genetic algorithm to tackle its solution. Moreover, we test this algorithm with 2020 COVID-19 data and found that our results are consistent both qualitatively and quantitatively. 5 - A Digital Twin for Online Optimization of Business Processes in a Supply Chain Hector D. Perez, Carnegie Mellon University, Pittsburgh, PA, United States, John M. Wassick, Ignacio E. Grossmann We present a virtual replica of the order-to-cash process that mimics process behavior via a stochastic queueing network representation. Discrete event simulation is used to: 1) estimate order fulfillment dates, 2) forecast potential order delays, 3) identify and mitigate bottlenecks, 4) quantify the impact of design changes, and 5) test optimization policies. An optimization module in the digital twin provides heuristics and mixed-integer scheduling models that can be used to dynamically assign orders to queues and assign order priorities in those queues during process execution. MP01 CC - Ballroom E /Virtual Theater 1 Plenary: From Learning to Optimize to Learning to Explore Plenary Session 1 - Plenary: From Learning to Optimize to Learning to Explore Yoshua,Bengio, University of Montreal, Dept IRO, Montreal, QC, H3C 3J7, Canada We consider a discrete combinatorial space and a given objective function where the goal is not to find the maximum of the objective function but rather to dis- cover its main modes, which can be turned into the question of sampling values with probability proportional to the objective function. By taking a power of the objective function, that formulation can smoothly transform the problem of find- ing the leading modes (with more or less emphasis on the really larger ones) into focussing on just the argmax of the objective. This problem comes up in drug dis- covery and material discovery tasks, where the objective function is only a proxy (e.g. from a simulator, or imperfect assays) for what we really care about (e.g., more expensive assays, like with mice models, or even clinical trials). Finding a diversity of good solutions is therefore important, because the single argmax solution may not in the end be appropriate. Although MCMC methods can in principle be used for that, we present an alternative approach based on deep generative models seen as policies sampling a sequence of discrete actions and that has the potential to use the power of systematic generalization in order to guess the presence of isolated modes of the objective function. This avoids the mode mixing issue which often comes up with MCMC in high-dimensional spaces where local search methods get stuck and even annealing is not enough, but instead relies on the potential of machine learning to generalize out-of-distri- bution, a rapidly expanding area of research in deep learning. Monday, 9:45AM-10:45AM

57