Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

MC65

control schemes under the assumption that network physics constraints are completely known. In this talk, we explore an alternative approach that learns a control procedure by observing streams of measurements from the system. We show that a simple batch learning and optimal control procedure attains comparable performance to existing approaches without requiring any prior knowledge of the network and give some analytical guarantees. We then show how the procedure can be utilized to construct an online learning process for the problem. 2 - Exploration of Machine Learning Techniques in Time Series pH Prediction in Water Distribution System Xiushuang Li, Arizona State University, Tempe, AZ, 85281, United States, Daniella Saetta, Pitu B. Mirchandani, Treavor Boyer Accurate prediction of the system state is extremely important in controlling water distribution system. Machine learning appears to be a great way to estimate and predict the state of the dynamic system. We explored the application of Symbolic Regression (SR), Lasso Regression (LR) and Neural Network models (NN) to predict time series pH in a small water distribution system. All three methods can achieve comparable accuracy (R2 > 0.93), but LR and NN models take much less time to train than SR model. LR can also filter insignificant input variables in pH prediction by forcing their coefficients to be zero. Both LR and NN can be a useful tool to build the predictive model in water distribution system control. 3 - A Bayesian Forecasting Model of Electric Outages Luis J. Novoa, James Madison University, Harrisonburg, VA, United States, Babak Zafari, Goran Vojvodic, Refik Soyer As an aid for planning and preparation against severe weather events we propose a forecasting model for electric outages under a bayesian framework. We applied the model using real data. Chair: Evan Barlow, Weber State University, Goddard School of Business & Economics, 1337 Edvalson St, Ogden, UT, 84408, United States 1 - Thermodynamics of Markov Decision Processes Peeyush Kumar, University of Washington-Seattle, Seattle, WA, United States This research introduces information theoretic methods inspired by stochastic thermodynamics for MDPs with model uncertainty. Specifically, this research models the physical aspects of information into system dynamics. I will show an explicit link between the information entropy and the stochastic dynamics of a system coupled to an environment. We analyze various sources of entropy production: due to the decision maker’s uncertainty about the system- environment interaction characteristics, and the stochastic nature of system dynamics. This analysis provide a formulation that enables us to define the optimality criterion for MDPs with model uncertainty. 2 - If You Can’t Beat Them, Join Them: Collective Intelligence Outperforms Artificial Intelligence at Diagnosing Skin Cancer Erik P. Duhaime, Massachusetts Institute of Technology, 100 Main Street, e62-390, Cambridge, MA, 02142, United States A growing chorus of academics, thought leaders, and politicians warn that artificial intelligence will cause unprecedented job loss, even among highly trained knowledge workers such as doctors. At the same time, advances in information technology have enabled entirely new ways of organizing work, and research has shown that crowds of people oftentimes outperform even their most highly skilled individual members. Here, I leverage a dataset of the diagnoses of 1 state-of-the-art artificial intelligence system and 21 board-certified dermatologists in order to examine what happens when artificial intelligence is combined with - rather than compared to - human intelligence. 3 - Learning Hidden Markov Models with Regulated Initialization using Multinomial Regression Akash Gupta, Oklahoma State University, 40 South University Place, Stillwater, OK, 747075, United States, Tieming Liu, Christopher Crick, Dursun Delen Hidden Markov Models (HMMs) are widely used for developing disease progression models. Learning parameters of HMMs using Baum-Welch (BW) algorithm with random initialization is a computationally expensive task. In this work, we propose a regulated initialization framework for BW algorithm using multinomial logistic regression that allows early convergence. The preliminary results showed improvement of 68 - 90% compared with random initializations. The proposed method facilitates the development of disease progression model with high granularity in disease states. n MC65 West Bldg 104B Data Science and Artificial Intelligence I Sponsored: Data Mining Sponsored Session

n MC62 West Bldg 103A Deep Learning Methods for Manufacturing Processes Sponsored: Data Mining

Sponsored Session Chair: Chitta Ranjan

Co-Chair: Kamran Paynabar, ISyE, Georgia Institute of Technology 1 - Apply Deep Learning to Powder Bed Fusion Process Physics in 3D Printing He Luan, HP Labs, HP Inc., Palo Alto, CA, United States, Sam Stodder, Jordi Roca, Jun Zeng, Qiang Huang, David Murphy, Thomas Paula Informed by physical knowledge, the powerful learning ability of deep learning has huge potential to model and understand the complicated physical behaviors, such as fusing principles involved in powder bed fusion based 3D printing. Using HP’s Multi Jet Fusion (MJF) as an example, this research demonstrates how deep learning can be applied to model and explore process physics. In first application, deep neural network models were built to predict the voxel level fusing layer thermal behavior from multiple thermal influencers. The second application models the heterogeneous material thermal diffusivity as a deep neural network, to enable both in-situ thermal prediction and better understanding. 2 - Online and Active Learning-based Classifier for Non-scalable Media Data Geet Lahoti, ISyE, Georgia Institute of Technology, 1038 Mcmillan St. NW, Atlanta, GA, 30318, United States, Chitta Ranjan, Chuck Zhang Typically dimension reduction is performed prior to building a classification model for high-dimensional data. However, how to build a classifier for high- dimensional data when dimension reduction is not possible? Consider a problem where sheet metals are to be classified as defective or non-defective using their large images. In this situation, when the defect size is extremely small in comparison to the image, rescaling of the image often loses the defect signal. Such a problem is more pronounced in various video datasets. To address this, we develop a new approach using Active and Online Learning that uses a combination of strongly and weakly labeled data to build a defect classifier. 3 - Nonlinear Profile Monitoring via Variational Autoencoder Hao Yan, 699 S. Mill Ave, Tempe, AZ, 85281, United States Unsupervised learning of nonlinear profiles has attracted increasing attention among researchers as well as practitioners because of its applications in process monitoring and diagnostics. However, existing unsupervised learning techniques for profile monitoring are typically based on linear models such as Principal Component Analysis, which fails to model the distribution of nonlinear profiles. We present a unified framework for nonlinear profile monitoring based on variational autoencoder, which a testing statistics can be naturally derived. We then demonstrate how this framework can be used for denoising and profile monitoring with various simulations and case studies. 4 - A Novel Classification Deep Boltzmann Machine (C-DBM) for Supervised Learning Samaneh Ebrahimi, Georgia Institute of Technology, Atlanta, GA, 30318, United States, Kamran Paynabar The Restricted Boltzmann Machine (RBM) is a probabilistic model to model the distribution of a visible layer of features using one hidden layer. Deep Boltzmann Machines were developed as an extension of RBM with multiple hidden layers. These methods have been successfully applied for unsupervised learning. In 2009, Larochelle and Bengio proposed a new discriminative RBM for supervised learning known as Classification RBM (CRBM). Due to estimation intractability, an effective deep extension of CRBM has not been used in the literature. In this paper, we propose a new estimation approach for deep CRBM (C-DBM). Moreover, we implement this approach on a manufacturing process for validation. n MC64 West Bldg 104A Joint Session DM/Practice Curated: Data Science and Analytics to Manage Power and Water Sponsored: Data Mining Sponsored Session Chair: Durai Sundaramoorthi, Saint Louis, MO, 63130, United States 1 - Time Varying Optimization and Learning in Power Systems Jiafan Yu, Stanford University, Stanford, CA, United States, Siobhan Powell, Ram Rajagopal Traditional approaches to reactive power management in distribution networks have focused on designing stable and near-optimal centralized or distributed

207