Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

SA22

4 - Formal Barriers to Proof of Stake Protocols Christos-Alexandros Psomas, Carnegie Mellon University, PA, United States, Jonah Brown-Cohen, Arvind Narayanan, Seth Matthew Weinberg The security of most existing cryptocurrencies is based on a concept called Proof of Work, in which users must solve a computationally hard cryptopuzzle to authorize transactions (“one unit of computation, one vote”). Proof of Stake is an alternative concept that instead selects users to authorize transactions proportional to their wealth (“one coin, one vote”). In this work we focus on incentive-driven deviations (participants deviate if doing so yields higher reward) instead of adversarial corruption (an adversary controls part of the network, but everyone else is honest) and show several formal barriers to designing incentive- compatible PoS cryptocurrencies. n SA20 North Bldg 129A Data Driven Models for Revenue Management and Pricing in Industry Sponsored: Revenue Management & Pricing Sponsored Session Chair: Ravi Kumar, PROS, Inc. World Headquarters, Houston, TX 1 - Methods of Handling Flexible Capacity in Hotel Revenue Management Matthew Maxwell, SAS Institute, Inc., 500 SAS Campus Drive, Cary, NC, 27513, United States One of the fundamental assumptions of revenue management is that capacity is fixed. In some situations this restriction is relaxed by using one type of capacity to fulfill requests for a different type of capacity. The most common practice of this concept is upgradingùwhere a request for a lower class of capacity is satisfied by a higher class of capacity. An alternate form of flexible capacity is that of component rooms, or virtual suites. In this scenario, a set of rooms can either be closed off and sold as separate individual units or combined and sold as a larger, more valuable suite. We explore methods for handling this type of flexible capacity. 2 - Reinforcement Learning for Revenue Management and Pricing Manu Chaudhary, PROS, Houston, TX, 77054, United States, Warren Scott We discuss two algorithms to automatically and dynamically generate optimal prices for non-negotiated settings. For the first algorithm, dynamic pricing problem is posed as a Multi Armed bandit problem and is solved using Beta- Bernoulli-Thompson Sampling (TS). The second algorithm is based on more traditional approach of calculating demand response models but the optimization method considers confidence intervals around mean demand to reduce risk. We Air Cargo is a highly volatile business, especially regarding uncertainty in the volume and weight of the shipments and capacity of the plane. Moreover, a full flight often consists of only a small number of shipments. We propose a revenue optimization model that deals with the uncertainty in two ways: first, demand is modeled as flexible products; and second, a robust optimization approach is proposed. 4 - Holistic Approach to Predicting the Wholesale Energy Market Prices Bokan Chen, CA, United States, Ana Radovanovic, Tommaso Nesti Electricity market price predictions enable energy market participants to shape their consumption or supply while meeting their economic and environmental objectives. By utilizing the basic properties of the supply-demand matching process performed by grid operators, we developed a method to recover energy market’s structure and predict the resulting prices as a function of generation mix and system load on the grid. Our methodology uses the latest advancements in compressed sensing and statistics to cope with the highly dimensional and sparse real grid topology graphs, underlying physical laws, as well as scarce and public market data. compare the performance of these algorithms using simulations. 3 - Flexible Products in Air Cargo Revenue Management Dirk Daniel Sierag, PROS, Houston, TX, United States

n SA21 North Bldg 129B Bridging Machine Learning and Revenue Management Sponsored: Revenue Management & Pricing Sponsored Session Chair: Xi Chen, NYU Stern School of Business 1 - Data Driven Algorithms for Assortment Planning Under Nested Logit Model Yuan Zhou, Indiana University at Bloomington, Xi Chen, Yining Wang Dynamic assortment planning concerns about the optimal displaying strategy to maximize total revenue over the selling season with no a priori information on consumers’ choice model parameters. Combining combinatorial techniques and the powerful lower-upper confidence bound (LUCB) method, we develop data- driven algorithms to simultaneously learn consumers’ model (the Nested Logit model) and optimize assortment selection decisions. Our algorithms’ performance guarantees surprisingly do not depend on the number of candidate products, which is particularly useful in settings such as fast fashion retailing and online advertising. 2 - Learning Preferences with Side-information: Near Optimal Recovery of Tensors Andrew Li, MIT, Cambridge, MA, United States, Vivek Farias Many recent problems in e-commerce can be cast as large-scale problems of tensor recovery. Thus motivated, we study the problem of recovering tensors from their noisy observations. We provide an efficient algorithm to recover structurally æsimple’ tensors given noisy observations of their entries; our version of simplicity subsumes low rank tensors for various definitions of tensor rank. Our algorithm is practical for massive datasets and provides a significant performance improvement over incumbent approaches to Tensor recovery. Further, we show a near-optimal recovery guarantee. Experiments on music streaming data demonstrate the performance and scalability of our algorithm. 3 - Attribute Based Dynamic Learning Approach to Assortment Selection Avadhanula Vashist, Columbia University, Shipra Agrawal, Vineet Goyal, Assaf Zeevi We consider an online assortment optimization problem, where in every round, the retailer offers an assortment of N substitutable products to a consumer, and observes the consumers response. We assume that the products are described by a set of attributes and the mean utility of a product is linear in the values of attributes. We model consumer choice behavior using the widely used multinomial logit (MNL) model, and consider the retailer’s problem of dynamically learning the model parameters, while optimizing cumulative revenues over the selling horizon T. We present an algorithm whose regret only depends on the number of attributes and is independent on the number of products. 4 - Dynamic Assortment Selection with Features Wang Yining, Carnegie Mellon University We consider the problem of dynamic assortment planning with features. More specifically, each item is associated with a known revenue parameter and a known feature vector (which changes over time), and the customers’ preferences are modeled by items’ feature vectors together with an unknown linear model. We discuss intuitive policies for dynamic planning (i.e., learning preference models and maximize expected revenues at the same time) and give near-optimal regret analysis. n SA22 North Bldg 130 Joint Session RMP/Practice Curated: Interface between RM and Market Analytics Sponsored: Revenue Management & Pricing Sponsored Session Chair: Ang Li, PROS, Inc., Houston, TX, 77002, United States 1 - Spot Scheduling for Cable Networks Xin Ma, Turner Broadcasting System, Inc., 1050 Techwood Dr NW, Atlanta, GA, 30318, United States, J. Antonio Carbajal Spot scheduling is a media planning process in which TV commercials are placed into buckets of commercial airtime for airing. In this presentation, we provide some background information of media planning, and review different types of deals, constraints, and requirements involved in this process. Then we introduce the spot scheduling problem, and discuss how we decompose it in a multi-stage optimization framework that realizes spot prioritization and best use of Turner’s commercial airtime.

9