Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

TA40

2 - Sequential Search For The Best Alternatives David Brown, Duke University, Fuqua School of Business, 100 Fuqua Drive, Durham, NC, 27708, United States, Santiago Balseiro We consider a variation of the sequential search problem studied in Weitzman (1979), where a DM sequentially searches a given set of alternatives with unknown rewards, drawn from independent distributions. Search is costly but reveals the rewards of an alternative. The DM can select previously revealed alternatives and collect the associated rewards. We study a variation in which the DM has the capacity to select multiple alternatives, which significantly complicates the problem. We consider a simple reservation price rule, with reservation prices depending only on the remaining capacity and show that the policy is asymptotically optimal as the number of alternatives grows large. 3 - On Learning the C Rule Subhashini Krishnasamy, Tata Institute of Fundamental Research, Mumbai, 400005, India, Ari Arapostathis, Ramesh Johari, Sanjay Shakkottai We study the c rule for multi class queueing systems. When the service rates are known, the c rule is known to minimize the expected holding cost over any fixed time horizon in the single server setting. For a parallel server network, we demonstrate that the c rule does not ensure stability. We present sufficient conditions for the stability of the c rule for a general parallel server network and also necessary and sufficient conditions for a specific class of parallel server networks. When the service rates are unknown, we propose, both for single and parallel server networks, algorithms using empirically learned service rates that achieve a holding-cost regret that does not depend on the time horizon. n TA39 North Bldg 226A Non-stationary Stochastic Networks with Information Sponsored: Applied Probability Sponsored Session Chair: Jamol Pender, Cornell University, Ithaca, NY, 14850, United States 1 - New Perspectives on the Erlang-a Queue Andrew Daw, Cornell University, Jamol Pender The non-stationary Erlang-A queue is a fundamental model used to describe the dynamic behavior of large scale multi-server service systems that experience customer abandonments, such as call centers, hospitals, and urban mobility systems. In this talk, we develop novel approximations for the moments, moment generating function, and cumulant generating function. We provide precise bounds for the difference of our approximations and the true model. Moreover, we show that our approximations have explicit stochastic representations as shifted Poisson random variables. Our approximations and bounds also hold for non-stationary Erlang-B and Erlang-C queues under certain stability conditions. 2 - Asymptotic Methods for Queues with Delayed Information Sophia Novitzky, Cornell University, Ithaca, NY, United State With the advancement of online technologies, services often provide waiting time or queue length information to customers. This information allows the customers to determine whether to remain in line or in the case of multiple lines, which line to join. Unfortunately, there is usually a delay associated with the waiting time information: either the information is not provided in real time or it takes the customers travel time to join the service after having received the information. Recent empirical and theoretical work uses functional dynamical systems as models for queueing systems and shows that if information is delayed long enough, a Hopf bifurcation can occur and cause unwanted oscillations in the queues. However, it is not known how large the oscillations are when a Hopf bifurcation occurs. To answer this question, we model queues with functional differential equations and implement two methods for approximating the amplitude of the oscillations. The first approximation is analytic and yields a closed-form approximation in terms of the model parameters. The second approximation uses a statistical technique, and delivers highly accurate approximations over a wider parameter range than does the first method. 3 - The Impact of Smartphone App Information on Bike Sharing Networks Shuang Tao, Cornell University, Ithaca, NY, United States Many systems like Citibike and Divvy provide riders with information about the network via smartphone apps so that riders can find stations with available bikes. In this paper, we study the impact of the smartphone app and its power to guide riders to non-empty stations. By combining customer choice modeling and finite capacity queueing model, we prove a mean-field limit and a central limit theorem for an empirical process of the number of stations with k bikes. Our results illustrate that if we increase the proportion of customers that use smartphone app information, the entropy of the bike sharing network is reduced, the network has a higher throughput, and riders experience less blocking in the network.

4 - The Impact of Queues with Delay and Pollution Announcements Jamol Pender, Cornell University, 228 Rhodes Hall, Ithaca, NY, 14850, United States Travel time information has been estimated and provided to drivers to help them make better routing decisions and alleviate congestion. However, because of challenges in data collection and sensor delays, travel time information is often delayed and hence inaccurate. This can misguide drivers and result in unstable traffic patterns. To this end, we explore the potential of giving drivers real-time en-route air pollution information and we develop a new queueing model that considers the behavior of drivers provided with both travel time and air pollution information. Our results indicate that provision of real-time air pollution information to travelers may help stabilize traffic.

n TA40 North Bldg 226B Expert Elicitation Sponsored: Decision Analysis Sponsored Session

Chair: Erin Baker, Univ of Massachusetts-Amherst, Univ of Massachusetts-Amherst, Amherst, MA, 01003, United States Co-Chair: Claire Cruickshank, Plano, TX, 75093, United States 1 - An Anchoring Explanation of Over- and Under-confidence during Probability Elicitation Saurabh Bansal, Penn State University, 405 Business Building, State College, PA, 16801, United States Understanding how good individuals are at estimating probability distributions is critical to theory and practice of decision analysis. A large body of literature has addressed this issue. The findings appear contradictory: some article report that decision makers’ judgments exhibit overconfidence while some others report underconfidence. We reconcile these seemingly inconsistent findings by using an anchoring-and adjustment model. A series of laboratory studies confirm the theoretical development. 2 - Combining Point Forecasts into a Predictive Distribution Zhi Chen, INSEAD, 1 Ayer Rajah Avenue, Singapore, 138676, Singapore, Anil Gaba Experts provide forecasts for a variable of interest often in the form of point forecasts, i.e., single-valued predictions. However, a decision maker has to then take the additional step of converting the point forecasts into a predictive distribution for that variable, necessary for the decision at hand. Theoretical and practical issues on this additional step will be discussed, including accounting for dependence among the experts. 3 - Does the Elicitation Mode Matter? Comparing Different Methods for Eliciting Expert Judgement Claire Cruickshank, MS. Student, University of Massachusetts, Amherst, MA, 01003, United States, Erin Baker An expert elicitation is a method of eliciting subjective probability distributions over key parameters from experts. Traditionally an expert elicitation has taken the form of a face-to-face interview; however, interest in using online methods has been growing. Our research project compared online and face-to-face elicitations by examining the effect of elicitation mode on the central values, overconfidence, accuracy and satisficing. Our results indicated that, in instances where the elicitation modes were directly comparable, the differences between the modes was not significant. Consequently, a carefully designed online elicitation may be used to obtain accurate forecasts. 4 - Aggregating Non-expert Opinions: Mechanism Design of Grading Scheme in MOOCs Steve Yoo, University College London, London, United Kingdom, Dongyuan Zhan One issue of operating Massive Open Online Courses (MOOCs) is the scalability of grading. MOOCs resort to peer-grading system. We model student learning and grading efforts in two stages. We find that with proper incentive in peer-grading, students exert more efforts in grading, which motivate their learning. Our model presents empirically testable model for researchers, and objective metrics for MOOCs to tailor their grading mechanism for each course.

257