INFORMS 2021 Program Book

INFORMS Anaheim 2021

WD32

3 - Optimizing for Strategy Diversity in the Design of Video Games Will Ma, Columbia University, New York, NY, United States, Oussama Hanguir, Christopher Ryan A situation common to video games is that a player, with limited resources, must choose a “loadout” of weapons to spend the resources on to maximize firepower. As the video game designer, is it possible to create weapons so that different players, at different stages of the game and possessing different resources, end up wielding different weapons in their loadout, as opposed to some weapons having so much firepower or being so cost-efficient that they are ubiquitous? In this work we show that there is a mathematical limit to the diversity in loadout strategies, and introduce an optimized weapon design which approaches this limit. WD32 CC Room 208B In Person: Statistical Learning and Optimization in Revenue Management/Platform Strategies General Session Chair: Hao Ding, Emory University, Emory University, Atlanta, GA, United States 1 - A New Approach for Vehicle Routing with Stochastic Demand: Combining Route Assignment with Process Flexibility Hanzhang Qin, Massachusetts Institute of Technology, Cambridge, MA, 02142-1365, United States, Kirby Ledvina, David Simchi- Levi, Yehua Wei We propose a new approach for the vehicle routing problem with stochastic demands for the case in which customer demands are revealed before vehicles are dispatched. Our approach combines ideas from vehicle routing and manufacturing process flexibility to propose overlapped routing strategies with customer sharing. We characterize the asymptotic performance of the overlapped routing strategies under probabilistic analysis. Using the characterization, we demonstrate that our overlapped routing strategies perform close to the theoretical lower-bound derived from the reoptimization strategy, and significantly outperforms the routing strategy without overlapped routes. The effectiveness of the proposed overlapped routing strategies in non-asymptotic regimes is further verified through numerical analysis. 2 - Learning to Use Auxiliary Observations in Adaptive Sequential Experiments Yonatan Gur, Stanford University, Stanford, CA, 94305-7216, United States In many practical settings, including assortment selection, pricing, and healthcare, Performance of sequential experiments can be improved in the presence of auxiliary observations that can be mapped to information on mean rewards. When these mappings are a priori unknown, we characterize necessary and sufficient conditions under which auxiliary information allows performance improvement. We devise a policy based on two different upper confidence bounds and establish its near-optimality. 3 - Are Buyers Strategic in Their Reviews in the B2B Market Hao Ding, Emory University In the digital age, buyer-generated-reviews drive purchasing decisions for buyers and success for both sellers and platforms. Buyers contribute in reviews to reward good sellers and help other buyers make informed purchasing decisions. Are buyers always altruistic in leaving reviews? In this paper, we study whether buyers in the B2B market leverage reviews strategically to maintain their bargaining power and preserve capacity allocation. Our unique dataset contains 240,288 reviews and 1,261,278 transactions of 6,682 suppliers on the largest B2B trading platform in the world. Using a difference-in-difference design with an exogenous shock, we causally test whether buyers change their review behaviors based on sellers’ recent transaction level. We find that buyers are 8% to 16% less likely to leave a review for each additional transactions that sellers had. WD33 CC Room 209A In Person: New Directions in Revenue Optimization General Session Chair: Georgia Perakis, Massachusetts Institute of Technology, Cambridge, MA, United States 1 - Addressing High Dimensional Prediction Challenges in RM Applications Ioannis Spantidakis, Massachusetts Institute of Technology, Cambridge, MA, 02139, United States, Tamar Cohen-Hillel, Georgia Perakis, Leann Thayaparan High dimensional data is a blessing and a curse, often necessary for the most

interesting machine learning problems but bringing with it feature-correlation, noise, and long runtimes. This is especially true in retail, where a large amount of information is needed to make optimal decisions. In this research, we develop a new dimensionality reduction algorithm called Supervised Approach for Feature Engineering (SAFE), which is an alternative to Principal Component Analysis (PCA). SAFE finds uncorrelated, lower dimensional features that best explain differences in the dependent variable (e.g., sales) facilitating the prediction task. 2 - XSTrees: Extended Sampled Tree Ensembles for Classification And Regression Omar Skali Lami, MIT, Ashdown 4091B, Cambridge, MA, 2139, United States, Georgia Perakis, Divya Singhvi Extended Sampled Trees (XSTrees), is a novel tree ensemble method for classification and regression. Instead of learning a single decision tree like CART, or an independent collection of trees like Random Forests, XSTrees learns the entire probability distribution over the tree space. This approach results in good theoretical guarantees and a significant edge over other methods in terms of performance. Analytically, we prove that XSTrees converge to the true underlying tree model with rate O(log(n)/n), where n is the number of observations. Experimentally, we show on publicly available datasets, synthetic data and two real-world case studies that XSTrees are very competitive with the state-of-the-art predictive models, with an average accuracy between 2.5% and 50% higher than competitors for classification, and an average R2 between 2% and 85% higher for regression. WD34 CC Room 209B In Person: Transportation-Operations I Contributed Session Chair: Bekircan Kirkici, Auburn University, Auburn, AL, 36830, United States 1 - Modeling System-wide Energy in Urban Rail Transit Systems for Sustainable Strategy Discovery and Decision-making Zhuo Han, University of Massachusetts Amherst, Amherst, MA, United States, Sean Donaghy, Eleni Christofa, Eric Gonzales, Jimi Oke Rail transit is critical to mobility in dense urban areas. In 2019, the light and heavy rail systems of the Boston area served 151M rides and consumed 414 GWh of energy. We estimate an interpretable machine learning model to predict energy usage, based on train trajectories, ridership and weather. We then develop a framework for discovering strategies to reduce energy (and consequently, costs and emissions) under a variety of scenarios. Further, we analyze the impacts of COVID-19 on the system in 2020 (ridership fell to 52M with 385 GWh energy consumed). We expect our framework will serve as a viable decision tool for sustainable and resilient urban rail systems. 2 - Modeling a Mixed-fleet Of Electric and Diesel Buses Operational Characteristics and Charging Infrastructure Planning in a Public Transit Network Amirali Soltanpour, Doctoral Researcher, Michigan State University, East Lansing, MI, United States, Mehrnaz Ghamami Electric buses (EB) reduce fuel consumption and emission production, while have limited range and are costly. This study finds the optimal bus types for each route and locates charging infrastructure within a transit network. Due to the computational complexity of the optimization model, a metaheuristic algorithm is developed to find the minimum investment and operation cost of transit systems. This study shows that smaller EBs can be more cost-effective than diesel buses. Electricity rate and availability of distributed energy resources affects the optimum location of charging infrastructure. 3 - Performance Evaluation and Pooling Efficiency for On-demand Public Transportation: A Case Study Bekircan Kirkici, Auburn University, Auburn, AL, United States, Daniel F. Silva, Alexander Vinel Demand responsive transportation system experiments have been recently implemented in several localities, which has renewed academic interest in the topic. We investigate a real-world system which aimed to create inexpensive means of transportation for those with limited access to mass transit. Hence, the focus is not profit maximization but equity and efficiency. First, we analyze the existing system’s performance. We then focus on pooling efficiency and investigate different related metrics and pooling criteria. Naturally, there exists a trade-off between waiting times and pooling, which we quantify based on the real data for the system in question.

172