Informs Annual Meeting 2017

MD22

INFORMS Houston – 2017

MD20

optimization model is developed to seek contributing variables to the accuracy of the prediction model. We found promising results experimenting on actual central gulf coast outage data.

342B TIMES Distinguished Speaker Sponsored: Technology, Innovation Management

MD22

& Entrepreneurship Sponsored Session

342D Multi-armed Bandit and its Applications to Revenue Management Sponsored: Revenue Management & Pricing Sponsored Session Chair: Xi Chen, New York University, New York, NY, 10012, United States, xchen3@stern.nyu.edu 1 - Optimal Maximum GAP Estimation in Multi-armed Bandit Yuan Zhou, Indiana Unviersity at Bloomington, Bloomington, IN, 47401, United States, yzhoucs@indiana.edu, Chao Tao, Xi Chen In this paper, we study the problem of maximum gap estimation in multi-armed bandit. We propose a new PAC-learning algorithm named MaxGapEst, which provides an $\eps$-close estimate of the maximum gap with probability $1- \delta$. The algorithm achieves the optimal sample complexity $\Theta(\frac{n}{\eps^2} \ln \\delta^{-1})$ for sufficiently small $\eps$. As a subroutine of the MaxGapEst, we propose a new algorithm for cost-sensitive best- arm identification problem, which assumes that sampling different arms incurs different costs. This result generalizes the classical results on the problem of best- arm identification with uniform sampling cost, which is of independent interest. 2 - Deep Exploration in Reinforcement Learning via Randomized Value Functions Daniel Russo, 3170 N. Sheridan Road, Apt. 1216, Chicago, IL, 60657, United States, Dan.Joseph.Russo@gmail.com Reinforcement learning algorithms learn to optimize performance in complicated MDPs by using observed state transitions. A major challenge is the ability to collect the right training data, and even important breakthroughs have relied on highly naive exploration, or the imitation of humans who have mastered the control task. We study the use of randomized value functions to guide sophisticated deep exploration in RL. This offers an elegant means for synthesizing statistically and computationally efficient exploration with common approaches to value function learning. We develop RL algorithms that leverage randomized value functions, and present computational studies and regret bounds. 3 - Non-stationary Stochastic Optimization with Local Spatial and Temporal Changes Yining Wang, Carnegie Mellon University, 5000 Forbes Avenue, 8203 Gates-Hillman Center, Pittsburgh, PA, 15213, United States, ynwang.yining@gmail.com We study the non-stationary stochastic optimization problem with local temporal and spatial changes by imposing an L_{p,q} norm constraints on the changes of functions between consecutive time epochs. We derive matching upper and lower bounds for our considered model, which have a few notable differences from the classical non-stationary stochastic optimization model where no locality of changes is imposed. In particular, we show a curse of dimensionality under our setting, which means the optimization problem gets harder when the problem dimension increases. 4 - Thompson Sampling Approach for Dynamic Assortment Selection Vashist Avadhanula, Columbia University, Industrial Eng. and OR, 423 SW Mudd Building, New York, NY, 10027, United States, vavadhanula18@gsb.columbia.edu, Shipra Agrawal, Vineet Goyal, Assaf Zeevi We consider a sequential subset selection problem under parameter uncertainty, where at each time step, the decision maker selects a subset of cardinality K from N possible items (arms), and observes a (bandit) feedback in the form of the index of one of the items in said subset, or none. Each item in the index set is ascribed a certain value (reward), and the feedback is governed by a Multinomial Logit (MNL) choice model whose parameters are a priori unknown. The objective of the decision maker is to maximize the expected cumulative rewards over a finite horizon T, or alternatively, minimize the regret relative to an oracle that knows the MNL parameters. We refer to this as the MNL-Bandit problem. This problem is representative of a larger family of exploration-exploitation problems that involve a combinatorial objective and arise in several important application domains. We present an approach to adapt Thompson Sampling to this problem and show that it achieves near-optimal regret as well as attractive numerical performance.

Chair: Juliana Hsuan, Copenhagen Business School, Department of Operations Management, Solbjerg Plads 3 B 5.27, Frederiksberg, DK- 2000, Denmark, jh.om@cbs.dk 1 - What’s Industry Got to do with it?

John Paul MacDuffie, University of Pennsylvania, Wharton School, Philadelphia, PA, United States, macduffie@wharton.upenn.edu

Studies of technological change and the management challenges it can create for incumbents commonly strive (like most social science) for theoretical insights and empirical findings that can be generalized to all settings. My research experiences have, through a mix of good luck, choice, and inertia, given me the opportunity for deep and prolonged study of a single context, the global automotive industry. This “industry of industries” (to quote Peter Drucker) has, over more than a century, seen an evolution in production system paradigms from craft to mass to lean production and is now facing a time of tremendous ferment as new technologies combine with new business models in a broader ecosystem of mobility services. Many analysts observing this time of ferment are applying the lens of theories of industry change developed during the rise of ICT/digital tech companies, with expectations of rapid transformation. I will argue that the application of these lenses can often be misguided, creating distorted assessments and misleading predictions, and revealing the need for better theorizing about contingencies and mechanisms as well as different research designs and methods. I will extract from my immersion in “industry studies” (both my research and my participation as a founding board member of the Industry Studies Association) some lessons for TIMES scholars about the advantages of bringing deep contextual knowledge about industries to the study of technological change and ecosystem evolution. 342C Decision Analysis and Natural Hazards Invited: InvitedNatural Hazard Planning Invited Session Chair: Sara Shashaani, University of Michigan, sshashaa@umich.edu 1 - Modeling Uncertainty: the Cost of Wrong Assumptions in Infrastructure Disaster Risk Modeling Benjamin Rachunok, Purdue University, West Lafayette, IN, United States, brachuno@purdue.edu, Roshanak Nateghi Risk modeling is a stochastic endeavor which commonly requires selecting a probability distribution to represent uncertainty in model inputs. No empirical data truly follows a theoretical distribution, as such distributions impose additional assumptions whenever they are used. In this talk, we aim to assess the cost of making incorrect assumptions in infrastructure disaster risk analyses. We present examples using publicly available tropical cyclone data (from NHC) and severe-weather related power outages (from the DoE) to illustrate the consequences of making incorrect assumptions in modeling risk, that can lead to over/under investments in infrastructure resilience. 2 - Preemptive Decision-making to Support Power System Resilience Andrea Staid, Sandia National Labs, 412 Morningside Dr NE, Albuquerque, NM, 87108, United States, astaid@sandia.gov, Jean-Paul Watson, Michael Bynum, Bryan Arguello The electric power system is vulnerable to natural hazards; consequences can have huge financial, social, and health impacts. To improve power system resilience, we must first understand the critical threats, potential damage, and how consequences can propagate through the system. Here we focus on improving resilience to by identifying optimal actions to take in advance of adverse weather. We use data from a large electric utility company to develop realistic scenarios of transmission system outages for use in a stochastic programming formulation using a DC optimal power flow model. We evaluate several mitigation strategies and characterize the impact on system resilience. 3 - Hurricane Power Outage Prediction with Out of Bag Feature Selection Approaches Sara Shashaani, University of Michigan, 1777 Plymouth Rd Apt 2C, Ann Arbor, MI, 48105, United States, sshashaa@umich.edu, Seth Guikema Predicting hurricane power outages facilitates natural hazards response decisions. However the spatial data is largely zero-inflated with a sizable number of explanatory variables, and finding statistical models that can provide reliable predictions remains to be a challenge. We study a feature selection approach by using out of bag performance of the predictors for such datasets. A combinatorial MD21

243