2016 INFORMS Annual Meeting Program

WB16

INFORMS Nashville – 2016

2 - Detecting Anomalous Patterns Of Care Using Health Insurance Claims Satya Venkata Somanchi, University of Notre Dame, 344 Mendoza College of Business, Notre Dame, IN, 46556, United States, somanchi.1@nd.edu, Edward McFowland In this work, we propose methods for detecting heterogeneous treatment effects in observational data. We apply these methods to the patient health insurance claims to improve clinical practice by analyzing patterns across patients and providing actionable insights. Our goal is to analyze this complex patient care data in order to identify interesting patterns in patient care that have led to anomalous health outcomes. Specifically, we detect treatments in the outpatient patient care that have significantly deviated from the regular treatment process and have affected health outcomes either negatively or positively. This can further help improve patient health and reduce health care costs. 3 - FastMemory-efficient Anomaly Detection In Streaming Heterogenous Graphs Emaad Ahmed Manzoor, Carnegie Mellon University, Pittsburgh, PA, 19106, United States, emaad@cmu.edu We present StreamSpot, a method to continuously track anomalous graphs as they evolve from a stream of edges with node and edge types. We introduce a graph similarity function that captures both node/edge types and temporal order, and a constant-space sketch that can be maintained incrementally to approximate this similarity in constant time. We show that when applied to detect malicious software from a stream of system call traces of executing processes, StreamSpot scales to over 10,000 edges/second and demonstrates an average precision of over 95%. Project website: http://bit.ly/streamspot 4 - Sequential Goodness Of Fit Testing With False Discovery Rate Control Sangwon (Justin) Hyun, Carnegie Mellon, Pittsburgh, PA, United States, robohyun66@gmail.com, Max G’Sell We consider sequential goodness-of-fit testing for sequential model selection procedures. This leads to a multiple hypothesis testing setting where the hypotheses are ordered and one is only permitted to reject an initial contiguous block, H_1, ..., H_k, of hypotheses. A rejection rule in this setting amounts to a procedure for choosing the stopping point k, corresponding to a particular selected model. We will discuss a multiple hypothesis testing procedure for FDR control in this setting. We will also introduce recent results for goodness-of-fit testing for clustering and the graphical lasso as an illustration of this approach. WB16 105A-MCC Optimal Learning and Optimization via Simulation Sponsored: Optimization, Optimization Under Uncertainty Sponsored Session Chair: Peter Frazier, Cornell University, 410 Thurston Ave, Ithaca, NY, United States, pf98@cornell.edu 1 - Optimal Learning With Discrete Priors Weidong Han, Princeton University, whan@princeton.edu We consider an optimal learning problem with a random parameter following a discrete prior distribution. We formulate the problem into a dynamic program, and propose several applicable polices. We consider both finite-horizon and infinite-horizon problems, provide a lower bound on a well-known heuristic policy in the former case, and prove asymptotic convergence properties of the proposed policies in the latter case. We also present comparisons against an optimal learning policy for selected problem instances. Empirical experiments show that the proposed policies achieve near-optimal performances. 2 - Parallel Knowledge Gradient For Global Optimization Of Expensive Functions Jian Wu, Cornell University, Ithaca, NY, 14850, United States, jw926@cornell.edu, Peter Frazier In this talk, we will introduce parallel knowledge gradient (pKG) algorithm for batch Bayesian optimization. The method chooses to evaluate the one-step Bayes optimal batch of points. We demonstrate empirically that the method can find global minima significantly faster than previous batch Bayesian Optimization methods on both synthetic functions and tuning hyperparameters of complex machine learning algorithms. Especially, the method provides most value in noisy setting.

3 - Secant Tangents-averaged Stochastic Approximation (STAR-SA) Marie Chau, Virginia Commonwealth University, mchau@vcu.edu We present new theoretical results for Secant Tangents-AveRaged stochastic approximation (STAR-SA) and extend it to higher dimensions using simultaneous perturbation stochastic approximation. Under a setting where both direct and indirect gradients are available, STAR-SA and STAR-SPSA (for the multidimensional case) combine direct and indirect gradients using a convex weight. We derive deterministic weights to minimize the MSE of the gradient and of the estimate, both of which are lower than the classical Robbins-Monro and Kiefer-Wolfowitz SA methods. We prove convergence, show asymptotic normality, and investigate their empirical performance against well-known SA algorithms. 4 - Optimal Learning For Nonlinear Parametric Belief Models With Continuous Alternatives Xinyu He, Princeton University, Princeton, NJ, United States, xinyuhe@princeton.edu, Warren B Powell We consider the optimal learning problem for nonlinear parametric belief models, where our goal is to find the optimal alternative through a series of noisy and expensive measurements. This problem has been studied when the number of alternatives is finite. Our work presents an algorithm that extends the optimal learning framework to the case with continuous alternatives. Experiments show that our algorithm on the continuous framework exhibits significant performance improvements, especially in higher dimensions. Chair: Necdet Serhat Aybat, The Pennsylvania State University, 310 Leonhard Building, University Park, PA, 16802, United States, nsa10@psu.edu 1 - DQM: Decentralized Quadratically Approximated Alternating Direction Method Of Multipliers Aryan Mokhtari, University of Pennsylvania, 200 South 33rd Street, Room 203 Moore Building, Philadelphia, PA, 19104, United States, aryanm@seas.upenn.edu, Wei Shi, Qing Ling, Alejandro Ribeiro The decentralized alternating method of multipliers (DADMM) is a well- established iterative method for solving consensus optimization problems; however, implementation of DADMM requires solving an optimization subproblem at each iteration for each node which can be computationally costly. We propose a decentralized quadratic approximation of ADMM (DQM) that reduces the computational complexity of DADMM by minimizing a quadratic approximation of the objective function. We show that DQM converges to the optimal argument at a linear rate which is identical to the convergence rate of DADMM. Moreover, as time passes the coefficient of linear convergence for DQM approaches the one for DADMM. 2 - A Geometrically Convergent Method For Distributed Optimization Over Time-varying And Directed Graphs Wei Shi, Postdoctoral Research Associate, Coordinated Science Laboratory, University of Illinois at Urbana-Champaign, Urbana, IL, 61801, United States, wilburs@illinois.edu, Angelia Nedich, Alex Olshevsky In this presentation, we introduce a class of first-order algorithms, referred to as DIGing and its variants, for distributed optimization over time-varying and/or directed graphs. The algorithms employ fixed step-sizes and, yet, drives all the agents’ iterates to a global and consensual minimizer. Under the strong convexity assumption, we show that the algorithms all converge at some explicit global R- linear (geometric) rates as long as the step-sizes are no greater than some explicit bounds. We also show that the rate for DIGing scales polynomially in the number of agents. Numerical experiments demonstrate the efficacy of the introduced algorithms and validate our theoretical findings. 3 - Distributed Admm-like Methods For Cooperative Multi-agent Optimization Over Conic Constraints Erfan Yazdandoost Hamedani, Penn State University, Department of Industrial Engineering, 310 Leonhard Bldg. University Park, State College, PA, United States, evy5047@psu.edu, Necdet Serhat Aybat We consider the sharing problem over an undirected (dynamic) network of agents, where only those agents connected by an edge can directly communicate. The objective is to minimize the sum of agent-specific composite convex functions subject to a conic constraint coupling all agents’ local decisions. An ADMM-like primal-dual distributed algorithm is proposed. We examine the convergence rate in terms of suboptimality, and infeasibility; study the effect of underlying network topology on the rates; and show how to extend this method to handle communication networks with time-varying topology. WB17 105B-MCC Distributed Consensus Optimization Sponsored: Optimization, Nonlinear Programming Sponsored Session

400