INFORMS 2021 Program Book

INFORMS Anaheim 2021

POSTER COMPETITION

18 - Conjecture on the Design of First Come First Served Skilled Parallel Service Systems Gideon Weiss, University of Haifa, Haifa, Israel Customers of several types are served by servers of different skills, subject to a bipartite compatibility graph. Service is first-come-first-served, FCFS, assign- longest-idle-server,.ALIS. With general service distributions this is an intractable system, it is even impossible to determine its stability. We model this as a problem of FCF S. matching of two multi-Bernoulli sequences, for which we can calculate matching rates. Based on these matching rates we obtain designs of work force that achieves quality of service as well as high utilization of resources. This is based on the conjecture that large volume many server systems converge to independent Poisson processes. 19 - Hyperparameter Optimization of Deep Neural Networks with Applications to Medical Device Manufacturing Gautham Sunder, Carlson School of Management, Minneapolis, MN, United States, Christopher Nachtsheim, Thomas Albrecht Bayesian Optimization (BO), a class of Response Surface Optimization (RSO) methods for nonlinear functions, is a commonly adopted strategy for Hyperparameter optimization (HO) of Deep Neural Networks (DNNs). Through a case study at a medical device manufacturer, we empirically illustrate that, in some cases, HO problems can be well approximated by a quadratic function, and in such cases BO is less efficient than Classical RSO (C-RSO) methods. When there is uncertainty in the complexity of the response function, we propose a highly efficient three staged batch sequential RSO strategy which estimates the response function complexity and adopts the best suited strategy between BO and C-RSO. 20 - Optimizing Moving Company Routes with COVID Restrictions Mohamad Afkhami, Blend360, Columbia, MD, United States, Amir Nasrollahzadeh, Pip Courbois, Serhat Kecici It is essential for an interstate moving company to know the size of the cargo in advance. Traditionally, this information was obtained through an in-person visit. With the in-place COVID restrictions, the companies have switched to relying on customers’ estimate, which may cause last minute cancellation, either due to price difference from initial quote or the limited capacity of the truck. As a result, an additional source of uncertainty is introduced in the planning of the moving company. We propose a stochastic optimization framework that incorporates this uncertainty in the routing planning of the moving companies. 21- Optimal Experiment Designs for Marketing Mix Models Amir Nasrollahzadeh, Blend360, Columbia, MD, United States, Mohamad Afkhami, Serhat Kecici, Pip Courbois Marketing mix models optimize advertising spend across different offline media (e.g., TV) by simulating the return of spend at geographical levels using regression. Recently, these models have incorporated online attributions which measure the effect of online channels (e.g., web) on customer conversion as another input to the media mix model. However, this approach fails to capture marketing lag effects, diminishing returns, and channel interactions. We propose a reinforcement learning approach to marketing experiment designs which learns the underlying relationship between spend and customer behavior while optimizing the return on investment. 22 - Solving the Canadian Prize Collection Problem with Application to Assess The Impact of an Ongoing Humanitarian Disaster John Becker, University at Buffalo, Buffalo, NY, United States Rajan Batta We introduce the Canadian Prize Collection Problem (CPCP): a pathing problem from s to t on a graph G where the unknown ground truth is a subgraph of G. Next, we provide two approaches, prize collection and shortest path to prize collection, as heuristic methods. Then, we apply this research in the area of disaster relief and conduct computational testing. 23 - Using Simulation to Advance Branch and Bound Search: Example for TSP Rajan Batta, University at Buffalo (SUNY), Buffalo, NY, United States, John Becker, Moises Sudit First, we introduce a sampling procedure for evaluating the value of branch-and- bound nodes. We then describe several branching procedures which guide our branch-and-bound search. Finally, we test our heuristic on a myriad of problem classes and compare it with other heuristics incorporated in branch-and-bound such as Fischetti-Lodi local branching and A* algorithms. 24 - Predicting the Outcome and Overuse Of Invasive Mechanical Ventilation in the Intensive Care Unit Maryam Alimohammadi, University of Arkansas, Fayetteville, AR, United States, Shengfan Zhang, Heather Nachtmann Mechanical ventilation is one of the main interventions in intensive care units (ICUs) for patients with various diagnoses and conditions. Predicting the outcome of mechanical ventilation in patients admitted to ICU can help clinicians manage ventilation resources better and improve the patients’. In this research, we train multiple machine learning conditions models to predict the outcomes of invasive mechanical ventilation in ICU. This helps identify the critical observation windows for close monitoring of patients. Additionally, we develop a framework based on the random forest model, which outperforms other models in most cases, to decide when to stop ventilation.

25 - Improved Competitive Ratios for the Secretary Problem with Biased Evaluations Kathryn Dullerud, University of Southern California, Los Angeles, CA, United States, R. Srikant We consider a variant of an algorithm introduced by Salam and Gupta for the secretary problem where the candidates’ evaluations are biased depending on the demographic group to which they belong. We present new competitive ratio results which improve existing bounds by a factor of e.

Poster Competition CC – Exhibit Hall B, Foyer In Person Poster Compettion Competition Poster Session 1 - Eigen-entropy: A Metric For Sampling Design

Jiajing Huang, Arizona State University, Tempe, AZ, United States Hyunsoo Yoon, Ojas Pradhan, Teresa Wu, Jin Wen, Zheng O’Neill Sampling is to identify a representative data subset capturing characteristics of the whole dataset. Existing sampling algorithms have some limitations including required assumptions on data distributions or models. In this study, a new metric, termed Eigen-Entropy, is proposed, derived based on eigenvalues extracted from correlation coefficient matrix on multivariate data. The performance of the proposed method is evaluated using real building case studies. Evaluation results indicate that the proposed method outperforms the methods from existing literature in terms of accuracy while maintaining smaller number of samples. 2 - Predicting the Outcome and Overuse of Invasive Mechanical Ventilation in the Intensive Care Unit Maryam Alimohammadi, University of Arkansas, Fayetteville, AR, United States, Shengfan Zhang, Heather Nachtmann Mechanical ventilation is one of the main interventions in ICU. Predicting the outcome of mechanical ventilation in patients admitted to ICU can help clinicians manage ventilation resources better and improve the patients’ conditions. In this study, we use multiple machine learning models to predict the outcome of invasive mechanical ventilation in ICU. The descriptive statistics of time- dependent variables are calculated based on multiple time windows during a patient’s stay in the ICU. We develop a framework based on the best model, which outperforms other models in most cases, to decide when to stop ventilation. 3 - Costly Active Sensing of Structured Partially Observable Markov Processes Xiaoqi Bi, University of Illinois, Urbana-Champaign, Champaign, IL, United States, Erik Miehling, Carolyn Beck, Tamer Basar Gathering information to learn a hidden state process is often costly in practice. To model such scenarios, we propose an active sensing model for partially observable Markov decision processes (POMDPs), with a belief-based reward that quantifies uncertainty of the latent state, and a cost for sensing actions. A core element of our model is the structured distributions dictating the uncertainties in the POMDP. We assume the prior state distribution is conjugate to the observation likelihood. Such structure ensures beliefs are of the same family of distributions as the prior. The proposed model has various real-world applications, including allocation of diagnostic tests in uncertain epidemics. 4 - A Python API for Accessing Forest Inventory and Analysis Database in Parallel Ashkan Mirzaee, University of Missouri, Columbia, MO, United States Forest Inventory and Analysis (FIA) Program of the U S. Forest Service provides the information needed to assess America’s forests. Many researchers rely on forest attribute estimations from the FIA program to evaluate forest conditions. The Python API is developed to collect large data from FIADB in parallel. In this project we used Python and Slurm workload manager to generate numerous parallel workers and distribute them across the cluster. The API is designed to scale up the query process such that by increasing processing elements the process expected to speedup linearly and can be set up and configured to be run on a single core computer or in a cluster for any given specifications. 5 - Identifying the Optimal Chronic Kidney Disease Screening Frequency Among Diabetics Chou-Chun Wu, University of Southern California, Los Angeles, CA, United States, Sze-chuan Suen Diabetes is a leading cause of chronic kidney disease (CKD), as 40% of diabetics will develop CKD in a lifetime. However, the rate of undiagnosed CKD among diabetics can be as high as 50%. We develop screening guidelines stratified by age, proteinuria status, and prior test history for diabetics by race and gender. We adopt a Partially Observed Markov Decision Process (POMDP) framework to identify the optimal action (screen or wait) every three months from ages 30-85 that maximizes a patient’s discounted lifetime net monetary benefit (NMB). The optimal policy suggests more frequent screening in all race and gender groups compared with the annual screening policy recommended in the status quo.

31