INFORMS 2021 Program Book

INFORMS Anaheim 2021

MC07

needed. The order of policies, whether RR should be first or second, gives rise to two dispatching policy classes, RATS and STAR. We show that the two-level STAR policy always outperforms RATS, and often outperforms any single-level policy. Moreover, STAR policies are robust across a range of parameter values and distributions for inter-arrival times and job sizes. 2 - On the Convergence Rate of Entropy-regularized Natural Policy Gradient with Linear Function Approximation Semih Cayci, Urbana, IL, United States, Niao He, R. Srikant We study the convergence rate of entropy-regularized Natural Policy Gradient (NPG) algorithms with linear function approximation. We show that NPG exhibits linear convergence within an approximation error and O(1/T) convergence to the optimal value function under standard assumptions on the distribution mismatch and the representation power of the feature vectors. 3 - Lower Bounds On Information Requirements for Causal Network Inference Xiaohan Kang, University of Illinois at Urbana–Champaign, Urbana, IL, 85281, United States, Bruce Hajek Recovery of the causal structure of dynamic networks from noisy measurements has long been a problem of intense interest across many areas of science and engineering. Many algorithms have been proposed, but there is no work that compares the performance of the algorithms to converse bounds in a non- asymptotic setting. As a step to address this problem, this paper gives lower bounds on the error probability for causal network support recovery in a linear Gaussian setting. The bounds are based on the use of the Bhattacharyya coefficient for binary hypothesis testing problems with mixture probability distributions. Comparison of the bounds and the performance achieved by two representative recovery algorithms are given for sparse random networks based on the Erd s-Rényi model. MC10 CC Room 304B In Person: Behavior-Aware Modeling of Service Systems General Session Chair: David Cho, Woodbury University, Burbank, CA, 91504-1052, United States 1 - Silent Abandonment in Contact Centers: Estimating Customer Patience from Uncertain Data Antonio Castellanos, Technion – Israel Institute of Technology, Haifa, Israel, Galit B. Yom-Tov, Yair Goldberg Contact centers are one of the favorite channels of communication with companies. However, they face operational challenges common proxies for customer experience are subject to information uncertainty. A main source of such is silent abandonment by customers. These customers leave the system while waiting for a reply, but give no indication for doing so. As a result, agent capacity is wasted. In two case studies we show that up to 70% of the abandoning customers abandon silently, and that such behavior reduces system efficiency by up to 15%. We develop methodologies to identify silent abandonment and to estimate customer patience. We show how accounting for silent abandonments in a queueing model improves the estimation accuracy of key measures of performance. Finally, we suggest strategies to operationally cope with the phenomenon. 2 - On Two Models of Choice Between an Observable and an Unobservable Queue with Heterogeneous Servers Jonathan Milo, Tel Aviv University, Tel Aviv, Israel, Refael Hassin We consider a queueing system where customers arrive according to a Poisson process and select one out of two servers with exponentially distributed service durations. Customers observe the first queue length and make an irrevocable decision on whether or not to enter it, without observing the second queue. We analyze two models where the first server is slower than the second. In both models there is no queue in front of the first server. In one model there is also no queue at the second server and customers who reach it when it is busy are lost. In the second model, there is a queue in front of the second server. We characterize the equilibrium behavior and investigate the relation between the equilibrium and optimal strategies, including the price-of-anarchy. 3 - Behavior Aware Service Staffing David D. Cho, Woodbury University, Burbank, CA, 91504-1052, United States, Kurt M. Bretthauer, Kyle D. Cattani, Alex Mills Empirical studies of service systems have shown that workers exhibit different service rates depending on their assigned workload. We model two commonly observed behavioral effects, speedup and slowdown, then incorporate the model into a multi-period workforce staffing problem to study their joint impact on service staffing. Our results show that a workload that maximizes the service rate is typically not optimal. We also find that the effectiveness of the widely practiced single-ratio workload staffing policy depends on the strength of the speedup and slowdown effects.

MC07 CC Room 201B In Person: Diversity, Equity, and Inclusion: Challenges and Opportunities in Social Media Analytics General Session Chair: Jorge Mejia, Indiana University, Bloomington, IN, 47401-5931, United States Chair: Christopher Dalton Parker, American University, Washington, 20011, United States 1 - Dancing to the #challenge: The Effect of TikTok on Closing the Artist Gender Gap Yifei Wang, University of Maryland, College Park, MD, United States, Jui Ramaprasad, Anand Gopal While creative industries have a reputation for being liberal and tolerant, this has not historically translated into greater actual inclusivity. In this study, we examine how technology platforms may impact on artist success on social media and digital music streaming platforms, and importantly may particularly help women getting their work noticed. More specifically, we explore the impact of the TikTok hashtag dance challenge on artists’ popularity across two types of social media platforms: Instagram and Spotify, and investigate if the effect varies between male and female artists. Our findings shed new light on social media marketing and artist self-promotion, especially making the music industry more inclusive and attractive to female music artists. MC08 CC Room 303C In Person: spORts II General Session Chair: Brayden Park, Colorado School of Mines, Golden, CO, United States 1 - Clinch and Playoff Magic Numbers for the English Premier League Mark Husted, Colorado School of Mines, Golden, CO, 80401, United States The English Premier League (EPL) is the top-level soccer league in England composed of 20 teams. At the end of the regular season, highly ranked teams qualify for one of two international tournaments, the more prestigious Champions League and the Europa League, and low ranked teams are relegated to the second tier for next year’s EPL season. An integer-programming model determines when a team has guaranteed its position, or, conversely, when it has been eliminated from international play or relegation before the completion of the regular season. 2 - Improving the Crystal Ball in Professional Sports Brayden Park, Colorado School of Mines, Golden, CO, United States, Sam King, Eli Olinick, Alexandra Newman So-called magic numbers capture the attention of fans across a variety of professional sports, and provide information regarding when a team has clinched or been eliminated from a playoff spot, and, additionally, when a team has captured or lost the opportunity for a first-place final standing prior to post- season play. The RIOT sports project determined these numbers exclusively for major league baseball in the United States. Using marketing research, we demonstrate how and why we have expanded to other leagues. We illustrate with a newly remodeled website and behind-the-scenes computations. In Person: APS Special Session on ‘High-dimensional Statistics, Algorithms, and Algorithmic Intractability’ General Session Chair: Daniela Hurtado Lange, Georgia Institute of Technology, Atlanta, GA, United States 1 - Star And Rats: Multilevel Dispatching Policies Rhonda L. Righter, Professor, University of California-Berkeley, Industrial Engineering And Ops Research, Berkeley, CA, 94720- 1777, United States, Esa Hyytia We consider how to improve dispatching decisions (routing jobs to servers) in large computing systems by combining basic assignment policies into two levels: the first level dispatcher assigns jobs to a set of second level dispatchers, each with their own pool of servers. At each level the decision is made by a static (STA) policy (such as random routing or routing based on job size) or by a Round-Robin (R) policy. Such policies are fast and scale well as only local information is MC09 CC Room 303D

60