Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

MD66

feature representations from both simulated and real data sets. 2 - Time Series Dimensionality Reduction Using Convolutional Recurrent Neural Networks Maziar Kasaei Roodsari, Arizona State University, Tempe, AZ, United States, Sangdi Lin, George Runger As huge volume of time series data is collected from various wearable devices and sensors in real-life, we are facing challenges in the storage, analysis, and visualization of time series data. An effective dimensionality reduction method for time series data is an important step towards solving these challenges. In this presentation, we discuss a novel deep learning model for time series dimensionality reduction which utilizes both convolutional neural networks and recurrent neural networks. The proposed model not only achieves low reconstruction error, but also generates discriminative features for future decision making. 3 - Ensemble-based Fast Shapelet Approximation Berk Gorgulu, Bogazici University, Istanbul, Turkeyr, Mustafa Gokce Baydogan, Gonenc Yucel Nearest-neighbor (NN) classifiers are widely used for classification of time series (TS) because of their simplicity. Recent TS classification methods focus on discovering discriminative subsequences, namely shapelets, to avoid problems with NN classifiers. Instead of pairwise distance calculations between the whole TS, shapelet-based approaches map TS to a feature vector based on the existence of the shapelets. We propose Ensemble-based Fast Shapelet Approximation (EFSA) that utilizes ensembles of regression trees to learn a piecewise approximation for shapelet identification in a supervised manner. Experiments show that EFSA yields fast and competitive results on benchmark datasets. n MD64 West Bldg 104A Joint Session DM/Practice Curated: Data Science for Food and Agriculture Sponsored: Data Mining Sponsored Session Chair: Durai Sundaramoorthi 1 - Optimal Control in Dynamic Food Supply Chain Under Service Level Constraints Ashton Kappelman, Kansas State University, 2061 Rathbone Hall, 1701B Platt St., Manhattan, KS, 66506, United States, Ashesh Kumar Sinha We consider a dynamic food supply chain with multiple suppliers and service level constraints. We propose an integrated approach that involves developing a Bayesian network to analyze the interaction of process parameters of suppliers on the service levels of the end product, and involves stochastic optimization models to provide insights on the structure of the optimal policy comprising of supplier at each level their process parameters. 2 - Predictive Analytics and Vehicle Routing for an Urban Food Delivery Platform Alexander Hess, PhD Student, WHU - Otto Beisheim School of Management, Burgplatz 2, Vallendar, 56179, Germany This talk illustrates the application of predictive analytics in conjunction with vehicle routing for an urban food delivery platform. In particular, we show how more accurate forecasts achieved with machine learning algorithms lead to an improved routing. This in turn lowers the need for capacity in vehicles. n MD65 West Bldg 104B Data Science and Artificial Intelligence II Sponsored: Data Mining Sponsored Session Chair: Shengfeng Chen, Western Michigan University, Kalamazoo, MI, United States 1 - Target-driven Navigation for Multi-robot via Deep Reinforcement Learning Mengqi Hu, University of Illinois at Chicago, 842 W. Taylor Street, MC 251, 2039 ERF, Chicago, IL, 60607, United States, Zishun Yu Multiple unmanned ground vehicle (UGV) navigation has attracted greater attention. In this research, we propose a multi-agent deep deterministic policy gradient (DDPG) model for multi-UGV navigation where the UGVs should move to their target points as fast as possible without collision with obstacles. DDPG is a model-free actor-critic algorithm for continuous control problem. In this model, we directly define the system state as a 1,500x3 matrix using data collected from

LiDAR installed on UGVs and actions as the angular and linear velocities. The experimental results demonstrate that the DDPG algorithm shows superior performance compared to state-of-the-art navigation algorithms. 2 - Improving Berth Schedule Efficiency via Evolutionary Algorithms with Efficient Parameter Control Maxim A. Dulebenets, Florida A&M University-Florida State University, Tallahassee, FL, 32311, United States, Masoud Kavoosi, Junayed Pasha, Olumide Abioye This study presents a self-adaptive Evolutionary Algorithm for the berth scheduling problem, where the crossover and mutation probabilities are encoded in the chromosomes and evolve throughout the algorithmic run. Computational experiments are undertaken to assess performance of the developed algorithm against the alternative Evolutionary Algorithms, which rely on the deterministic parameter control, adaptive parameter control, and parameter tuning strategies respectively. Results show superiority of the self-adaptive Evolutionary Algorithm over the alternative algorithms. 3 - Production Planning Using Reinforcement Learning Karthick Gopalswamy, PhD Student, North Carolina State University, Raleigh, NC, 27606, United States We consider the problem of finding optimal release plans for a production system under congestion given deterministic demand and random processing time. The problem is formulated as a model-free simulation based Reinforcement Learning (RL) problem. We outline the methodology and provide computational results comparing the operational costs of RL based algorithm and a simulation optimization algorithm. 4 - An Image Based Deep Convolutional Neural Network for Industrial Product Defect Detection Shengfeng Chen, Western Michigan University, 1903 Western Michigan Avenue, Kalamazoo, MI, 49008, United States, Lin Cheng Machine vision system are increasingly being used in industrial product defect detection. Traditional methods work well only under certain conditions with many requirements. We propose a deep convolutional neural network (CNN) structure that can efficiently and reliably extract powerful features for defect detection. With only four deep convolution layers, the proposed model can detect tiny changes in the image with a small training dataset of less than 400 prior knowledge of images. Experimental results achieve a high accuracy of near 100 percent. Text Mining in IS Research Sponsored: Artificial Intelligence Sponsored Session Chair: Jingjing Li 1 - More than the Quantity: Estimating the Value of Editorial Review for UGC Platform Yipu Deng, Purdue University, 2826 Peachleaf Drive, West Lafayette, IN, 47906, United States, Jinyang Zheng, Warut Khern-am-nuai, Karthik Kannan This work studies the implication of an editorial review program where a review platform starts to supplement the user-generated content on its website with editorial reviews that are written by the platform. Our research question is whether platform-generated content (i.e., editorial reviews) influences subsequent user-generated content (i.e., user reviews) both in terms of the quantity and quality. Our analysis suggests that editorial reviews have a positive net effect on subsequent user reviews in general. Specifically, users post more reviews for restaurants that have editorial reviews. Moreover, these user reviews tend to be longer and resemble editorial reviews in respect of quality. 2 - Social CRM and Brand Crisis: A Natural Experiment from Airline Industry Ramah Al Balawi, University of Illinois at Chicago, 601 S. Morgan St, Chicago, IL, 60607, United States, Yuheng Hu, Liangfei Qiu This work aims to understand how brands change their engagement strategy in response to a crisis. By using a recent crisis of United Airlines, we collected Twitter data about United Airline customer service and find that United Airline strategically changes its social media policy when engaging with its customers. n MD66 West Bldg 105A

237