Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

TB53

2 - A Column-generation Approach to Maximizing Angle Board Production Samir Bhetwal, Northern Illinois University, Dekalb, IL, United States, Christine Vi Nguyen The project focuses on providing a reliable solution to a paper production company producing angle boards. The company receives raw materials from the sister companies, and therefore has little control over the thickness of the paper that is used to create the angle boards. The column generation algorithm has been developed and a set of patterns are generated for each type of angle board product by which the production run can be maximized with the current available supply of raw materials. The model considers the current set of raw materials and its attributes in the production of good quality angle boards. 3 - Optimal Production and Inventory Planning with Inventory Based Financing Renato E. de Matta, University of Iowa, 2360 Mulberry Street, Coralville, IA, 52241, United States, Vernon Hsu We study a multi product, multi period production planning problem with restriction on the available working capital. We use cycle inventory as collateral to secure loans. We formulate the problem as a mixed integer programming model. The problem with one product is NP hard. We develop an efficient heuristic procedure to solve the multi product problem. We examine a variety of economic scenarios to show how the firm could significantly improve its profitability with the availability of inventory-based financing. We use real world data to validate our model and develop managerial insights. 4 - A Rhythm Wheel Approach for Production Planning and Batch Sizing Gokhan Memisoglu, LLamasoft, Inc., 201 South Division St., Ann Arbor, MI, 48104, United States, Mike Bucci In this study, we developed an optimization tool that includes line balancing, production cycle and sequencing. The intent of this tool is to create a high-level tactical solution which can guide a short-term production scheduling application. The tool uses Rhythm Wheel (RW) approach to estimate production cycles and creates a solution that balances setup and cycle stock costs. This tool has been used by a major pharmaceutical company in several project with great success. 5 - A Different Approach to Reduce Dimensionality in Planning Problems Carlos Monardes, Pontificia Universidad Catolica de Chile, Avenida Vicuna Mackenna 4860, Macul, Santiago, Chile, Alejandro Francisco Mac Cawley, Jorge R. Vera, Susan C. Cholette, Sergio Maturana Planning faces time dimensionality problems as instances size grows. In this work, we present a methodology to cope dimensionality. First, we defined a data structure to keep off time index in decision variables, which allowed us to construct a MIP model. Second, we solved the former problem using Constraint Generation. We tested this approach in winery industry, where model assists the assignment decision process of harvested grapes to fermentation tanks. This approach has shown interesting properties, increasing computational implementation efficiency. Joint Session SMA/Practice Curated: Medical Decision Support with Social Media Analytics Emerging Topic: Social Media Analytics Emerging Topic Session Chair: Qingpeng Zhang, City University of Hong Kong, Kowloon, 12180, Hong Kong 1 - Using Search Query Data to Predict HIV New Diagnoses Qingpeng Zhang, City University of Hong Kong, 83 Tat Chee Avenue, 6/F, Academic 1, Kowloon, 12180, Hong Kong Objectives: Internet data are important sources of abundant information regarding HIV epidemics and risk factors. A number of case studies found an association between Internet searches and outbreaks of infectious diseases, including HIV. In this research, we examined the feasibility of using search query data to predict the number of new HIV diagnoses in China. Design: We identified a set of search queries that are associated with new HIV diagnoses in China. We developed statistical models (negative binomial generalised linear model and its Bayesian variants) to estimate the number of new HIV diagnoses by using data of search queries (Baidu) and official statistics (for the entire country and for Guangdong province) for 7 years (2010 to 2016). Results: Search query data were positively associated with the number of new HIV diagnoses in China and in Guangdong province. Experiments demonstrated that incorporating search query data could improve the prediction performance in nowcasting and forecasting n TB52 North Bldg 231C

tasks. Conclusions: Baidu data can be used to predict the number of new HIV diagnoses in China up to the province level. This study demonstrates the feasibility of using search query data to predict new HIV diagnoses. Results could potentially facilitate timely evidence-based decision making and complement conventional programmes for HIV prevention. 2 - Social Support and User Churn Prediction for Online Health Communities – A Trajectory-based Deep Learning Method Xiangyu Wang, University of Iowa, Iowa City, IA, United States, Apoorva Joshi, Kang Zhao, Xun Wang Online Health Communities (OHCs) are a great source of social support for patients and their caregivers. Better predictions of user churns from OHCs can help to manage and sustain a successful OHC. We incorporate two methods into churn prediction for OHCs: identifying different types of social support activities from posts users have published via text mining, and using LSTM to learn from users’ trajectories in different types of social support activities. n TB53 North Bldg 232A Joint Session AMD/RMP: Incentives, Dynamics and Learning Sponsored: Auction and Marketing Design Sponsored Session Chair: Vasilis Syrgkanis, Microsoft Research, Cambridge, MA, 02139, United States 1 - Dynamic Pricing/Auctions by Multi-scale Online Learning Rad Niazadeh, Stanford University, 353 Serra Mall, Gate Bldg., Office 484, Stanford, CA, 94305, United States, Sebastien Bubeck, Nikhil Devanur, Zhiyi Huang In this talk, I explore dynamic pricing/auctions from the perspective of online adversarial learning, inspired by the cloud pricing application. The model we consider is simple: a seller sells an identical item in each period to a new buyer (or a new set of buyers) by posting a different price at each time (or by running one of the possible incentive compatible auctions). The goal is to extract a revenue that (almost) matches the revenue of best posted price (or auction) in hindsight. For the online posted pricing problem, we show regret bounds that scale with the best fixed price, rather than the range of the values (with a generalization to learning auctions). Moreover, we demonstrate a connection between the optimal regret bounds for this problem and offline sample complexity lower-bounds of approximating optimal revenue, studied in [Cole and Roughgarden, 2015]. Using this connection, we show our regret bounds are almost optimal as they match these information theoretic lower-bounds. Our online auctions and pricing are obtained by generalizing the classical learning from experts and multi-armed bandit problems to their “multi-scale versions”, where the reward of each action is in a different range. Here the objective is to design online learning algorithms whose regret with respect to a given action scales with its own range, rather than the maximum range. We show how a variant of online mirror descent solves this learning problem. 2 - Non-clairvoyant Dynamic Mechanism Design Song Zuo, Tsinghua University, Beijing, China, Vahab Mirrokni, Renato Paes Leme, Pingzhong Tang Despite their better revenue and welfare guarantees for repeated auctions, dynamic mechanisms have not been widely adopted in practice. This is partly due to their implementation complexity and unrealistic use of forecastings. We address the shortcomings and present a new family of dynamic mechanisms that are simple to compute and require no future distribution knowledge.We introduce the concept of non-clairvoyance in dynamic mechanism design, which means the allocation and pricing rule does not depend on future type distributions.We develop a framework to characterize the revenue extraction power of non-clairvoyant mechanisms against mechanisms with future distributional knowledge. 3 - Learning to Bid Without Knowing Your Value Chara Podimata, Harvard University, 33 Oxford St, Cambridge, MA, 02138, United States, Zhe Feng, Vasilis Syrgkanis We address online learning in complex auction settings, where the value of the bidder is unknown to her, evolving in an arbitrary manner and observed only if the bidder wins an allocation. We leverage the structure of the utility of the bidder and the partial feedback that they typically receive in auctions, in order to provide algorithms with regret rates that are exponentially faster in terms of dependence on the action space, than generic bandit algorithms. For that, we analyze a new online learning setting with outcome-based feedback, generalizing learning with feedback graphs. Lastly, we verify experimentally the performance of our algorithm and its robustness to noise in the feedback received.

293