Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

POSTER SESSION

Monday, 12:30PM - 2:30PM

6 - Multivariate Stochastic Approximation versus Design of Experiments for Learning Vector Quantization Hyperparameter Tuning Trevor Bihl, Air Force Research Laboratory, 209 George Wythe Way, Beavercreek, OH, 45434, United States, Daniel Steeneck The Generalized Relevance Learning Vector Quantization-Improved (GRLVQI), a Kohonen neural network classifier, is very accurate in the identification of communication devices in robust security applications. However, GRLVQI, like neural networks in general, has multiple hyperparameters which have been tuned via ad hoc and experiential methods. Herein, the authors consider two approaches to finding optimal GRLVQI hyperparameters: 1) factorial experimental design and 2) stochastic approximation. The results show that stochastic approximation has an improvement in accuracy over factorial designs while not constraining the design space. 7 - Machine Learning to Establish Confidence in Project Due Date Setting Yu Xia, Associate Professor, College of William and Mary, Mason School of Business, 101 Ukrop Way, Williamsburg, VA, 23185, United States, Jie Ding Erroneous setting of project due date can cause huge waste of resources. In project management, various postulated models are used to forecast project due date. However, project due date is influenced by many supply chain factors and some unexpected events. It is difficult to decide whether a postulated model is well-specified or not since they usually involve complex settings with nontrivial likelihood functions or vague priors on their parameters. In this paper, we works on establishing a quantitative index for confidence level of a due date setting. We also gather results from multiple models and come up with a self-learning self- adjusting final result to tackle the task. 8 - Atmospheric Monitoring Using Nano-satellites Jean-Pierre Ibitokun, Graduate Student, University of Arkansas at Little Rock, EIT South University Avenue, Little Rock, AR, 72204, United States, Yupo Chan, Edmond Wilson, Po-Hao Adam Huang Global warming concerns have led to environmental monitoring of our planet. The GOES satellite, a multi-billion platform, has been a key player in atmospheric sensing and weather forecast. We propose to use a constellation of CubeSat’s, or nano-satellites, to perform a similar function at a much lower cost. Signals from an onboard spectrometer will map atmospheric water vapor concentrations, C02, oxides of nitrogen, hydroxyl radicals and ozone concentrations, and recording solar spectra to establish trends between solar flux and atmospheric water vapor concentrations. This is to be validated against the patterns discerned from the GOES data using signal and image processing. 9 - Transforming Organizations Through the Use of Metadata Christy W. Goodnight, University of South Alabama, Mobile, AL, United States This poster examines the effect of the organizational structure of data of all types on a new process called the metadata transformational process. Adding a layer of metadata to all the data collected within our organization allows for us to use this knowledge in new and innovative ways that has the potential to transform the way we not only do business, but also the way our society functions. Thus, organizations that go through a metadata transformation will enable knowledge sharing which has the potential to facilitate innovation across the globe. 10 - Estimating Youden Index Under the Multivariate ROC Curve in the Presence of Missing Mass Malignant and Benign Biomarker Data Multivariate Youden index (J) and its corresponding optimal cut-point (c*) have become the most effective and efficient approach in multi-biomarker mass data analysis to distinguish diseased from the healthy in this digitized era. However, the existing statistical classical approach does not produce exact solutions in the presence of nuisance parameters and missing and group mass data. This proposed novel high-performing computational-based generalized Variable Method is an alternative procedure to produce exact statistical inferential results for J and c* in the presence of such constraints. 11 - Vehicle Flow Control Using Reinforcement Learning on Manufacturing Environment Younkook Kang, Seoul National University, Seoul, Korea, Republic of, Sungzoon Cho Automated Material Handling System (AMHS) which is responsible for movements between equipment is becoming an important factor of manufacturing competitiveness in semiconductor industry environment. In order to control traffic, traffic congestion should be reflected as a cost but it is hard to find proper cost. Our study focuses on setting traffic cost to maximize throughput at a key area. We suggest a method to adjust routing cost using reinforcement learning. It can help adapting traffic pattern changes without human interference. In the end, it can contribute improving manufacturing productivity. Lakmali Weerasena, University of Tennessee Chattanooga, 615 McCallie Ave, Chattanooga, TN, 37403, United States

n Monday Poster Session Exhibit Hall Poster Monday Poster Session Poster Session Chair: Neng Fan, University of Arizona, 1127 E. James E. Rogers Way, P. O. Box 210020, Tucson, AZ, 85721, United States Co-Chair: Junming Yin, University of Arizona, Management Information Systems Department, McClelland Hall, Room 430BB, Tucson, AZ, 85721, United States Co-Chair: Burcu B. Keskin, University of Alabama, 4138 Meretta Lane, Tuscaloosa, AL, 35406, United States 1 - Comparative Analysis of the Machine Learning and Statistical Classifiers Ivan Belik, Assistant Professor, Norwegian School of Economics, Helleveien 30, Bergen, 5045, Norway Since machine learning - statistics gap has experienced an outstanding reduction in the last years, machine learning methods have been applied for classification purposes as an alternative to purely statistical methods. In the given research we analyze machine learning and statistical techniques for the classification purposes and compare the efficiency and classification accuracy of those techniques. Specifically, we build the Na ve Bayes Classifier, based on the Bayesian Statistics concept, and the Artificial Neural Network classifier, based on the functional aspects of the biological neural networks. We are running the given competing paradigms over the variety of real-world datasets. 2 - Dynamic Cloud Resource Optimization of B2C E-commerce Process Using Deep Reinforcement Learning Ming Gao, Associate Professor, Dongbei University of Finance & Economics, China, Jianshan Street 217, Shahekou District, Dalian, 116030, China Current research on the optimization of cloud resources deployment for e- commerce platform is less based on the business process level and real world scenarios. In order to dynamically optimize the cloud resources for specific service scenarios of e-commerce transaction process, this paper analyzes the characteristics of the B2C e-commerce transaction scenario from published real world transaction history data, and apply the deep reinforcement learning algorithm to generate cloud resource realtime optimization strategies. Our work provides a reference solution for e-commerce companies to bridge the gap between business process goals and cloud resources efficientcy. 3 - Recent Approaches to Creating and Analyzing Big Data in Technology Management and Commercialization Clovia Hamilton, Assistant Professor of Management, Winthrop University, 701 Oakland Avenue, 310 Thurmond Building, Rock Hill, SC, 29733, United States This is a study of three (3) technological advances related to Big Data: web scraping, natural language processing and machine learning. This research identifies ways in which these advances can become more relevant and useful to empirical researchers in the field of technology and innovation management. This research provides a brief look at some applications of these advances in research studies, and engenders a broader discussion around the relevance and impact of these advances to research in technology and innovation management. In particular, making use of these advances in empirical research related to technology commercialization was researched and is discussed. 4 - Benchmarking Stream Processing Frameworks Giselle van Dongen, Ghent University, Ghent, Belgium, Dirk Van den Poel Due to the increasing interest in real-time processing, many stream processing frameworks were developed. However, no clear guidelines have been established for choosing a framework and designing efficient processing pipelines. We fill this gap by establishing a benchmark methodology for fine-grained benchmarking of common operations on multiple metrics: latency, peak throughput, sustainable throughput, memory usage and cpu utilization. We implemented this benchmark for four popular stream processing frameworks: Spark, Storm, Flink and Kafka Streams. 5 - A Two-stage SMAA-DEA Model for Supply Chain Evaluation and Rank We integrate stochastic multi-criteria acceptability analysis technique (SMAA) and data envelopment analysis (DEA) methodology and propose a two-stage SMAA-DEA model for supply chain rank and efficiency evaluation with stochastic criteria values. Our study extends network DEA to address uncertain or stochastic measures. The developed model can be considered as an MCDM method with a two-stage additive DEA value function. An empirical study of 27 supply chains evaluation is presented to illustrate the proposed models. Sheng Ang, University of Science and Technology of China, 96 Jinzhai Road, Hefei, China, Yunxia Zhu, Feng Yang

177