Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

MA03

2 - Sequential BTST Algorithm for the Multiple War Problem Atsuo Suzuki, PhD, Nanzan University, Zvi Drezner We apply the BTST (Big Triangle Small Triangle) method to the multiple Weber problem with attraction and repulsion (WAR). We formulate a variation of the WAR problem where several existing facilities are located in an area. The customers access to the nearest facility. The problem is to find the location of a new facility so as to minimize the weighted sum of the distances from the customers to their nearest facilities. We apply the BTST method to the problem. By solving the problem sequentially, we obtain an approximate solution of the multiple WAR problem. 3 - Recent Results on Solving the Quadratic Assignment Problem using Graphics Processing Unit Clusters on the Blue Waters Supercomputer Rakesh Nagi, University of Illinois Urbana-Champaign, 8 Transportation Building, 104 S. Mathews Ave., Urbana, IL, 61801, United States, Ketan Date We discuss parallel implementations of RLT2 formulation and branch-and-bound algorithm for solving the Quadratic Assignment Problem (QAP). Our parallel architecture consists of NVIDIA Graphics Processing Unit (GPU) clusters on the Blue Waters supercomputer at the University of Illinois at Urbana-Champaign. We implement a “distributed Dual Ascent algorithm for the GPUs, which shows excellent parallel speedup, and can effectively solve some of the well-known QAPs from the literature. 4 - Computational Results for Solving the Euclidean Distance Min-Max Location Problem with Fixed Distances in N-Dimensions n MA01 North Bldg 121A Bayesian Machine Learning Sponsored: Optimization/Optimization Under Uncertainty Sponsored Session Chair: Matthias Poloczek, University of Arizona, Tucson, AZ, 85721, United States 1 - Continuous Multi-task Bayesian Optimization with Correlation Juergen Branke, University of Warwick, UK, Warwick Business School, Orms Group, Coventry, CV4 7AL, United Kingdom, Michael Pearce We consider the problem of simultaneously identifying the optima for a set of correlated tasks, where the performance of a particular input parameter on a particular task can only be estimated from noisy samples. We propose a general multi-task Bayesian optimization framework and two myopic sampling procedures that determine task and parameter values for sampling, in order to e?ciently ?nd the best parameter setting for all tasks simultaneously. 2 - Bayesian Optimization of Combinatorial Structures Ricardo Baptista, Massachusetts Institute of Technology, Cambridge, MA, 02139, United States, Matthias Poloczek The optimization of expensive-to-evaluate black-box functions over discrete domains is a ubiquitous task in machine learning and engineering. The combinatorial explosion of the search space and costly function evaluations pose challenges for current techniques. In this talk we will propose Bayesian optimization of combinatorial structures (BOCS). Our algorithm is based on an adaptive statistical model that identifies useful combinatorial structures even when data is scarce, and a scalable acquisition function that uses semidefinite programming. We will also present experimental results demonstrating that BOCS outperforms other methods from combinatorial and Bayesian optimization. 3 - Bayesian Optimization with Expensive Integrands Saul Toscano, Cornell University, Ithaca, NY, 14850-5645, United States, Peter Frazier We propose a Bayesian optimization algorithm for objective functions that are sums or integrals of expensive-to-evaluate functions, allowing noisy evaluations. These objective functions arise in multi-task Bayesian optimization for tuning machine learning hyperparameters, optimization via simulation, and sequential design of experiments with random environmental conditions. Our method is average-case optimal by construction when a single evaluation of the integrand remains within our evaluation budget. We also show consistency of our method for objective functions that are sums. In numerical experiments, our method performs as well or better across a wide range of problems. Mark Cawood, Clemson University, 220 Parkway Drive, Clemson, SC, 29634-0975, United States, P. M. Dearing Computational results for primal and dual algorithms are presented for the problem, also called the minimum covering Euclidean ball of a set of Euclidean balls in n-dimensions. Both algorithms search along a directed path that is either a ray or a hyperbola. The step size is computed explicitly at each iteration. Monday, 8:00AM - 9:30AM

n MA02 North Bldg 121B

Policy Search – Solving Stochastic Optimization by Searching Over Parameterized Classes of Policies Sponsored: Optimization/Optimization Under Uncertainty Sponsored Session Chair: Warren B. Powell, Princeton University, Princeton, NJ, 08544, United States 1 - Learning to Learn How to Bid: Efficient Meta-learning based on Policy Search in Adaptive Algorithms for Sponsored Search Auction Bidding Donghun Lee, Princeton University, 11 Lawrence Drive, Apt 503, Princeton, NJ, 08540, United States In online optimization problems where the cost of obtaining examples is accounted for, the need for better efficiency in learning algorithm becomes clear. Contrary to traditional big-O notation asymptotic sample complexity analysis, we take practical perspective to efficiency: learning algorithms become more efficient if its performance is improved over less iterations and samples. We propose a meta-learning framework that improves the practical efficiency of a given generic learning algorithm by finding better parameters for the algorithm. We evaluate the meta-learning framework in adaptive bidding algorithms deployed in sponsored search auctions such as Google Adwords platform. 2 - Structured Actor Critic for Managing and Dispensing Public Health Inventory Yijia Wang, University of Pittsburgh, 331 Devonshire Street, Pittsburgh, PA, 15213, United States, Daniel Jiang Public health organizations face the problem of dispensing medications to groups of affected populations during emergency situations, typically in the presence of complexities like demand stochasticity and limited storage. We formulate an MDP model with two levels of decisions: the upper-level decisions come from an inventory model that “controls” a lower-level problem that optimizes dispensing decisions. We then derive structural properties of the MDP model and propose an ADP algorithm that leverages structure in both the policy and the value space. 3 - Parametric Cost Function Approximations for Multistage Stochastic Programming Saeed Ghadimi, PhD, Princeton University, Princeton, NJ, 08544, United States, Warren B. Powell, Raymond Theodore Perkins The parametric cost function approximation represents an alternative to stochastic lookahead models that represent the foundation of stochastic programming. Parametric CFAs require some intuition into how uncertainty might affect the optimal solution. 4 - Adaptive Aggregation with Exogenous Transition Uncertainty Bing-Zhen (Amy) Zhang, Cornell University, Ithaca, NY, United States, Itai Gurvich, Huseyin Topaloglu Approximation is often used in dynamic programming due to the high computational cost for high-dimensional problems, incurring a loss in optimality. However when transition probabilities are inaccurate due to limited empirical data, even the exact model would have produced suboptimal results. We look at the interplay of the two when applying approximation, specifically state aggregation, in the presence of model uncertainty; we show that aggregation will only cause limited additional loss, if any, and propose selection procedure for an aggregation mechanism that reduces such loss. Advances in Crowdfunding Research Sponsored: Technology, Innovation Management & Entrepreneurship Sponsored Session Chair: Philipp Cornelius, Rotterdam School of Management, Rotterdam, Netherlands 1 - Designing Crowdfunding Platform Rules to Deter Misconduct Simone Marinesi, Assistant Professor of Operations Management, The Wharton School, University of Pennsylvania, Philadelphia, PA, 19103, United States, Elena Belavina, Gerry Tsoukalas Lacking credible rule enforcement mechanisms to punish entrepreneurial misconduct, existing reward-based crowdfunding platforms can leave campaign backers exposed to two sources of risk: the risk that entrepreneurs run away with backers’ money (funds misappropriation) and the risk of product misrepresentation (performance opacity). We examine the adverse consequences of both misconduct risks and propose practically implementable platform designs aimed at curbing their harmful effects. n MA03 North Bldg 121C

119