INFORMS 2021 Program Book

INFORMS Anaheim 2021

MC28

MC28 CC Room 207B In Person Technology Tutorial: High Performance Computing Capabilities in Purdue’s MS BAIM Program Technology Tutorial 1 - High Performance Computing Capabilities in Purdue’s MS BAIM Program Matthew A. Lanham, Purdue University, Lafayette, IN, 47905- 4803, United States This tech tutorial will showcase the HPC capabilities at Purdue University and how masters’ students in our top-ranked MS Business Analytics & Information Management (BAIM) program are leveraging our capabilities to achieve outcomes with industry partners. MC29 CC Room 207C In Person: Last Mile Science and OR/ML Practice at Amazon General Session Chair: Liron Yedidsion, Amazon, Redmond, WA, 98052, United States 1 - Multi-stage Newsvendor Problem: Applications to Contact Center Staffing Kevin A. Melendez, Amazon.com, Seattle, WA, 33613, United States, German Riano We introduce an extension of the classical Newsvendor Problem in which we consider heterogeneous suppliers for the same stochastic demand. Each supplier has different fixed and variable cost, and are used in a given order. The problem is to allocate the demand into the available suppliers to minimize cost. We develop a heuristic algorithm that can solve the problem to optimality if certain conditions are met. We show an application of our model to dynamically staff Amazon contact center network. To the best of our knowledge, this approach has not been used in this context before. 2 - A Practical Take on Decomposition Algorithms for Stochastic Programming Semih Atakan, Amazon.com, Seattle, WA, 98121, United States Typical operations research aims to develop the fastest methodology to solve established problems of value to the community. In practice, business problems can evolve constantly and demand a fast answer, or else the opportunity to influence the business may go away. In stochastic programming, some of the “fastest” methodology comes with restrictive assumptions, which may get violated as soon as the underlying problem changes deeming more methodological effort necessary. In this talk, we put on a practitioner’s hat, and describe (and justify) the algorithmic choices we made when building our stochastic program capabilities. Our talk will focus on versatility challenges, practical concerns to stochastic programming (which we cannot resolve easily), and some bonuses of having flexible decomposition algorithms in our toolset. 3 - Demand Forecasting for New Nodes in a Delivery Network Chinmoy Mohapatra, Research Scientist II, Amazon, Seattle, TX, United States, Rohit Malshe, Liron Yedidsion, Abhilasha Katariya, Jin Ye, Dipal Gupta Effective demand forecasting is a crucial element in the short-term and long-term planning process of any logistic network. We study the demand forecasting problem for new nodes that are added to an existing logistic network. Such nodes may have very limited historical data and may share limited time-invariant features with existing nodes in the network. We propose a hybrid optimization and machine learning based forecasting approach that considers both time- varying and static features of different nodes. We show the effectiveness of the proposed approach through a real-life case study from a logistics network.

MC30 CC Room 207D In Person: Spatial & Temporal Analytics and Applications I General Session Chair: Jian Liu, University of Arizona, Tucson, AZ, 85719-0505, United States 1 - Spatiotemporal Monitoring of Melt-pool Variations in Additive Manufacturing Siqi Zhang, United States Advancements in image sensing systems offer great opportunities for in-situ monitoring and control of melt-pool characteristics in Additive Manufacturing. However, prior efforts are more concerned about feature-based modeling and analysis of melt-pool imaging data. Little work has been done to leverage the tensor decomposition to transform the time-varying melt-pool imaging data into low-dimensional profiles, and then utilize the Gaussian process to model these low-dimensional profiles for in-situ monitoring of AM process. In this paper, we present a novel Gaussian process framework for statistical modeling and monitoring of melt-pool imaging data. Experimental results show that the proposed framework shows great potential for process monitoring and control of AM process. 2 - Online Nonparametric Monitoring for Asynchronous Processes with Serial Correlation Ziqian Zheng, University of Wisconsin-Madison, Madison, WI, United States With the development of modern sensor technology, more and more complicated data streams are involved in process monitoring. However, most of the existing studies assume that the sampling intervals of all the data streams are the same, and process observations at different time points are independent. In this paper, we propose a generic nonparametric MSPC scheme that can handle asynchronous process data with serial correlation. Specifically, we first propose a nonparametric method for the pairwise correlation function estimation. Then an asynchronous monitoring framework is proposed to monitoring the decorrelated process. The performance of the proposed method is evaluated based on both synthetic data and a real-world dataset. 3 - StressNet Deep Learning to Predict Stress with Fracture Propagation in Brittle Materials Xiaowei Yue, Virginia Tech, Blacksburg, VA, 24061, United States, Yinan Wang, Weihong Guo Accurate prediction of internal stress is critical to improving the fracture resistance and reliability of materials. To reduce computational cost of Finite- Discrete Element Model (FDEM), a deep learning model, StressNet, is proposed to predict the entire sequence of internal stress. Specifically, the Temporal Independent Convolutional Neural Network is designed to capture the spatial features like fracture path and spall regions, and the Bidirectional Long Short- term Memory is adapted to capture the temporal features. By fusing these features, the evolution in time of the internal stress can be accurately predicted. Moreover, an adaptive loss function is designed to reflect the fluctuations in internal stress. The proposed model can realize accurate multi-step predictions in about 20 seconds, as compared to the FDEM run time of 4h, with an average MAPE 2%. 4 - Spatial-Temporal Trip Demand Prediction Considering Trip Chaining Effect Fenglian Pan, University of Arizona, Tucson, AZ, United States, Jian Liu Effectiveness of traffic management relies on accurate prediction of trip demand. In daily life, people usually travel in a chain of trips, which influence each other spatially and temporally. Without explicitly considering such spatial-temporal (ST) interdependence, existing methods fall short in prediction accuracy. In this research, a Hawkes process model is proposed to predict trip demand, with trip interdependence represented as a ST triggering pattern in the form of structural kernel function. An algorithm is developed to enable the estimation of the model with latent triggering pattern parameters. The performance of proposed model is demonstrated in a real-world case study.

66