Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

TA67

3 - Forecasting Demand Distribution for New Products using Subjective Rankings Marat Salikhov, PhD Student, INSEAD, 1 Ayer Rajah Ave, Singapore, 138676, Singapore, Nils Rudi We present a framework for demand distribution forecasting for new products that combines historical data for similar products with subjective ranking inputs. The framework is based on the decomposition of a demand vector into aggregate, ordered proportion and ranking components, with historical data being used to forecast the first two components and subjective inputs for the last one. The component forecasts are then combined via a simulation-based method. We propose multiple specifications for each of the components. Finally, we evaluate the out-of-sample performance of the method using actual demand data and subjective inputs from a retail company. 4 - A Data Glove Based American Sign Language Interpreter Sara Masoud, University of Arizona, 1127 E. James E. Rogers Way Room 111, Tucson, AZ, United States, Bijoy Dripta Barua Chowdhury, Young-Jun Son Sign language recognition is a central research problem in computerized gesture detection for enabling hearing impaired people. In this work, a real-time hand gesture recognition framework is proposed to detect the 26 sign language alphabets. VMG 30 data gloves are utilized by which each hand gesture is reported as a 132-feature vector. Feature extraction is performed via a linear discriminant analysis to reduce the size of input vectors to a 25-feature vector and decrease the chance of overfitting. Decision tree models are developed to detect the 26 sign language alphabets based on the 25-feature vector values. The proposed framework has shown a success rate of 98.23% in a real-time environment. n TA65 West Bldg 104B Flash Session – Applied Probability Sponsored: Applied Probability Sponsored Session Chair: Rami Atar, Technion, Technion, Haifa, 32000, Israell Co-Chair: Harsha Honnappa, Purdue University, West Lafayette, IN, 47906, United States 1 - Session on Open Problems in Applied Probability Harsha Honnappa, Purdue University, 315 N. Grant Street, West Lafayette, IN, 47906, United States Short presentations on open problems in Applied Probability (broadly defined), followed by discussion of the problem. n TA66 West Bldg 105A Reinforcement Learning for Supply Chain and Inventory Optimization Sponsored: Artificial Intelligence Sponsored Session Chair: Afshin OroojlooyJadid, Lehigh University, 200 West Packer Ave, Bethelhem, PA, 18015, United States Co-Chair: Mohammadreza Nazari, Lehigh University, Bethlehem, PA, 18015, United States 1 - A Deep Q-network for the Beer Game: A Reinforcement Learning Algorithm to Solve Inventory Optimization Problems Afshin Oroojlooy Jadid, Lehigh University, 15 Duh Drive, Bethlehem, PA, 18015, United States, Mohammad Reza Nazari, Lawrence V. Snyder, Martin Takac Beer game is a multi-agent serial supply chain in which agents attempt to minimize the total cost of the network given that each agent only observes its own local information. We propose a Deep Q-Network to solve this problem. Our algorithm outperforms approaches from literature and unlike the majority of them does not impose restrictions on the problem parameters, and works well even if other agents do not follow rational policies. Moreover, we propose a transfer learning approach to quickly train new agents with new cost coefficients or action space. The algorithm can be extended to any decentralized multi-agent cooperative game with partial information, which is a common situation in supply chains.

2 - A Deep Reinforcement Learning Approach for Dual Sourcing Inventory Management Joren Gijsbrechts, KU Leuven, Leuven, Belgium, Robert Boute, Dennis Zhang, Jan Van Mieghem The optimal dual sourcing replenishment policy with non-consecutive lead times does not have a simple structure and is highly state dependent. We model the dual sourcing problem as a Markov Decision Process, and compare three machine learning algorithms based on deep reinforcement learning to solve this analytically intractable problem. Our results show that in the traditional settings our algorithms match (and slightly outperform) the current best-in-class heuristics. Additionally, we develop an experiment including more realistic constraints such as non-linear ordering costs and non-stationary demand showing how machine learning can be applied in a more generalized setting. 3 - Returns Optimization Ajay Deshpande, IBM Research, 1101 Route 134 Kitchawan Rd, Yorktown Heights, NY, 10598, United States, Brian Quanz, Ali Koc, Yada Zhu, Jae-Eun Park Growth in e-commerce has spurred growth in online returns. IBM Research developed a returns optimization solution which finds an optimal location in a network to maximize net expected recovery on returns. Using learning based optimization we demonstrated significant potential of incremental revenue and logistics cost savings for a large fashion retailer. n TA67 West Bldg 105B Natural Language Processing and User Generated Content Sponsored: Information Systems Sponsored Session Chair: Nicholas Sullivan, University of Utah 1 - Understanding User Engagement in Social Media Through Neural Network-Based Text Mining Michael Lee, University of Nevada Las Vegas, 1142 West 200 North, Unit A406, Centerville, UT, 84014, United States Abstract not avalable. 2 - How do Traditional Rivals Compete on Online Social Media Platforms? An Empirical Investigation Mikhail Lysyakov, University of Maryland, Robert H. Smith School of Business, College Park, MD, 20783, United States, Kunpeng Zhang, Siva Viswanathan The evolution of social media platforms has generated new channels for firms to compete with their rivals. There is very little research on how traditional rivals compete on social media and what the effects of such competition are for related outcomes. To address the gap, our paper examines how firms compete on Twitter. We focus on firms that are close competitors in the traditional context and examine whether these traditional rivals also compete closely (i.e. adopt similar strategies) on Twitter. We find evidence of both isomorphism and divergence among traditional rivals in their online social media content strategies. Interestingly, we find that divergence is linked to higher online engagement. 3 - Master, Jedi, or Guru? Normalizing Job Titles in an Ever-Changing Environment Nicholas Carson Sullivan, University of Utah, 5752 W. Kintail Ct, West Valley, UT, 84128, United States In this project, we explore different ways to address the issue of job title normalization. This task has been addressed previously as a way to assess different text classification approaches or advanced word-embedding methods. However, we approach the issue in an attempt to create workable data that can be used to create a more effective job recommender system. Such a system would need to be built on a highly standardized and normalized dataset. Job title normalization is a difficult problem not only because there are often different names for very similar jobs (i.e. Software Developer vs. Software Engineer), but also because it is becoming a common practice to intentionally make up creative job titles for the same jobs (i.e. Java Jedi, Software Guru). Unlike previous work addressing similar tasks, we take advantage of the fact that job titles do not exist in isolation. Throughout a career, an individual will hold many different job titles in a sequential order. For any given job title, valuable information can be learned by looking at the job titles held by the individual before and after the given title. We leverage this information to propose new methods to address the task of job title normalization.

267