Informs Annual Meeting 2017

WB71

INFORMS Houston – 2017

WB70

4 - Blending of Computer Experiment Models with and with out Tuning Parameters Patxi Fernandez-Zelaia, Georgia Institute of Technology, Atlanta, GA, United States, pfz3@gatech.edu, Rui Tuo, V. Roshan Joseph This work focuses on the building of efficient statistical models from different classes of complex computer simulations and cheap analytical solutions. The computer experiments considered are deterministic yet contain numerical errors associated with real-valued tuning parameters. Additional errors corresponding to different computer code classes may be associated with the physical granularity or solution algorithm found in each code. Following recent work nonstationary Gaussian Process models are employed to blend predictions across tuning parameters. Blending of cross-class computer codes and analytical solutions is achieved utilizing unknown latent tuning parameters. 371D Optimization, Convex Contributed Session Chair: Tao Wu, Dow Chemical, Midland, MI, United States, danielwu9999@gmail.com 1 - Robust Optimization: Sensitivity to Uncertainty in Scalar and Vector Case Giovanni Paolo Crespi, Associate Professor, Universita’ degli studi dell’Insubria, Via Monte Generoso, 71, Varese, 21100, Italy, giovanni.crespi@uninsubria.it, Matteo Rocca, Daishi Kuroiwa How robust solutions react to changes in the uncertainty set? We prove their location w.r.t. the magnitude of a possible decrease in uncertainty, i.e. when the uncertainty set shrinks. Uncertainty may arise from incomplete information about people’s (stakeholders, voters, etc.) perception about a specific issue. When the decision maker (DM) has to look for approval, they might need to define the strategy that displeases the minority. Robust optimization allows such a solution. Nevertheless, such a solution might be far from the strategy that the DM wishes to uphold betraying the DM’s goals. Our results allow such a distance to be measured and to steer the robust solution closer to the desirable one. 2 - Advances in Pseudo-cut Strategies for Global Optimization Christopher Riley, Assistant Professor of Management, Delta State University, Cleveland, MS, United States, criley@deltastate.edu, Leon S.Lasdon, Fred W. Glover, Cesar Rego Pseudo-cuts are linear cutting planes that provide temporary, but possibly invalid, restrictions on the solution space of complex optimization problems, to guide a solution process toward a global optimum. We provide new pseudo-cuts for nonlinear optimization using scatter search and tabu tunneling mechanisms, and present computational results showing the merit of our approach. 3 - An Approach for Robust Multiobjective Optimization Through a Superiority and Robustness Tradeoff Framework Adham Mohammed, Research and Teaching Assistant, Cairo University, Faculty of Engineering at Cairo University, Cairo University Avenue, Giza, 12613, Egypt, adham@cu.edu.eg, Mohamed Gadallah, Sayed Metwalli This paper introduces an approach for Robust Multi-Objective Optimization that considers the two aspects of solution merit: performance superiority and robustness. A multi-objective formulation is proposed, combining the Performance Superiority Objectives with derived Performance Robustness Objectives. The multi-objective analysis is carried out and the Pareto frontier is obtained. This Pareto is then perturbed via Monte Carlo simulation and Feasibility Robustness is used for selecting a reduced Pareto. An Absolute Minimization Index is finally used to select the best single solution from this reduced Pareto. A results comparison with Goal Programming is provided through test problems. 4 - Surrogation Based Tight Bounds for Indefinite Separable Multiconstrained Quadratic Programs Jaehwan Jeong, Assistant Professor, Radford University, Department of Management, P.O. Box 6954, Radford, VA, 24142, United States, jjeong5@radford.edu, Chanaka Edirisinghe A linear-time bounding algorithm is developed for indefinite separable quadratic programs. It surrogates multiple constraints to form a single constraint within an iterative scheme to generate a sequence of indefinite knapsack sub-problems that are solvable based on our previous research. The computational analysis with two-constraint case shows the superiority of our method relative to CPLEX. 5 - Analytics Optimization for Lot Sizing Tao Wu, Data Scientist, Dow Chemical, 7109 Arrowhead Circle, Midland, MI, 48642-8110, United States, danielwu9999@gmail.com We develop an analytics optimization method for lot sizing where the information is utilized for an analytics-based branching and selection procedure to fix setup variables. Comparisons with other methods indicate that the optimization method is computationally tractable and can obtain better results. WB69

371E Data Mining Contributed Session Chair: Muhammed Sutcu, Abdullah Gul University, Kayseri, Turkey, muhammed.sutcu@agu.edu.tr 1 - Sentiment Analysis of Blog Text Based on CNN and OCC Model Peng Wu, Professor, Nanjing University of Science &Technology, No.200, Xiaolingwei Road, Nanjing, Jiangsu, 210094, China, wupeng@njust.edu.cn, Heng wang Liu, Si Shen This paper proposed a Sentiment Analysis model based on OCC (Ortony, Clore, and Collins) model and Convolution Neural Network(CNN)model from the perspective of cognition. A system of fuzzy inferential rules of blog users’ emotion is formulated based on OCC model. The blog texts are represented by vectors and are used for training a polarity classification model based on CNN model. The competence of the proposed model is assessed through experiments on both Chinese and English online user’ blog texts. The results shows the OCC model can improve the accuracy of the annotation the emotion tendency, and the analyzing results of CNN is significantly higher than the method of SVM(Support Vector Machine). 2 - A Doubly Adaptive Inferential Method for Monotone Graph Invariants Junwei Lu, Princeton University, Princeton, NJ, United States, junweil@princeton.edu, Matey Neykov, Han Liu We aim to infer some graph invariants in graphical models (e.g., the maximum degree, the number of connected subgraphs, the number of singletons). We propose a new inferential framework called skip-down for testing nested multiple hypotheses and constructing confidence intervals of the unknown graph invariants under undirected graphical models. Compared to perfect graph recovery, our methods require significantly weaker conditions. Theoretically, we prove that the length of the obtained confidence intervals are optimal and doubly adaptive. 3 - Flipping Out: Exploring Motion Classification Methodologies Destie Provenzano, Deloitte, Arlington, VA, 22209, United States, destie@destie.com, Joel Bock The classification of real-time movement from video data using empirically-based learning algorithms is difficult. This study explored methods to simplify the feature construction process, and accelerate learning to predict physical maneuvers from a database of gymnastics videos. Beyond classifying movements, assessment of how well a maneuver has been executed is enabled by this novel approach. Salient results from these numerical experiments will be presented, including increased precision and sensitivity as compared to previous methods. 4 - A New Rational Classification Approach by using the Novel Mixed Data Binarization Method Muhammed Sutcu, Abdullah Gul University, Endustri Muhendisligi Bolumu, Kayseri, 38060, Turkey, muhammed.sutcu@agu.edu.tr, Omer Batuhan Kizilisik In this study, we propose a new methodology for classifying the similar customers in same classes and then find the most suitable class for the new coming customer. First, we determine the classes from available data and then we classify the new instances into these predetermined classes by using the new novel data binarization approach. We show how the each step of this algorithm could be performed efficiently.

WB71

371F Optimization, Linear Programming Contributed Session Chair: Xianfei Jin, Sabre, Southlake, TX, United States, Xianfei.Jin@sabre.com

1 - Complete Enumeration of Non-isomorphic Minimally Aliased Response Surface Designs using Integer Programming an d High-throughput Computing Jose Nunez Ares, PhD candidate, KU. Leuven, Kasteelpark Arenberg 30 - bus 2456, Leuven, 3001, Belgium, jose.nunezares@kuleuven.be, Peter Goos, Jeffrey T. Linderoth Response Surface Designs (RSDs) are statistical designs widely used in process industries. We enumerate 3-level Minimally Aliased RSDs, which are {-1,0,1} matrices with desirable statistical properties. Our branch-and-prune algorithm exploits the inherent symmetry of the design matrices. The enumeration is parallelized to run on a large-scale, heterogeneous and non-dedicated computing environment. Our novel load-balancing scheme relies on phases of breadth-first search along with a statistical estimation of the size of the enumeration tree rooted at child nodes to achieve efficiency. We find many new statistical designs of size significantly larger than previously known.

499