2016 INFORMS Annual Meeting Program

SB16

INFORMS Nashville – 2016

4 - Lockers Network: A Solution To Last Mile Delivery Problem Guodong Lyu, National University of Singapore Business School, #B1-01, BIZ2 Building, 1 Business Link, Singapore, 117592, Singapre, 117592, Singapore, guodong.lyu@u.nus.edu Chung-Piaw Teo Lockers (Parcel Collection Point) are convenient for parcels collection. To maximize the coverage of failed delivery by using lockers, a novel locker location model is introduced and customer choice is considered as an input. We provide the mobile flow from stations to residence blocks, and use the mobility data to locate lockers. The mobile delivery flow is calculated based on the transportation data and failed delivery data. Customers’ parcel collection behavior is estimated from locker usage data. Furthermore, lockers location data from an Express are provided to validate our model. We demonstrate that the value of using mobility date for failed delivery coverage maximization is significant. SB16 105A-MCC Algorithms for Stochastic Programming Sponsored: Optimization, Optimization Under Uncertainty Sponsored Session Chair: Natashia L Boland, Georgia Institute of Technology, 755 Ferst Drive, NW, Atlanta, GA, 30332, United States, natashia.boland@isye.gatech.edu 1 - Scenario Set Partition Dual Bounds For Multi-stage Stochastic Programming Ilke Bakir, PhD Candidate, Georgia Institute of Technology, Atlanta, GA, 30309, United States, ilkebakir@gatech.edu Natashia Boland, Brian Dandurand, Alan Erera We propose expected partition (EP) bounds, a hierarchy of bounds based on partitions of the scenario set into subsets of (nearly) equal cardinality. Additionally, using the fact that solution of group subproblems for all subsets in a single partition also yields a dual bound, we establish that the best of even a very small number of such single partition bounds is likely to be better than the corresponding EP bound. By sampling partitions of the scenario set, we obtain an algorithm that computes confidence intervals for an EP bound, while also providing exact dual bounds. The practical value of these ideas is illustrated on benchmark problems, and the approach is compared to Sample Average Approximation. 2 - AMPL Representation And Solution Of Multiple Stochastic Programming And Robust Optimization Formulations Christian Valente, OptiRisk Systems, One Oxford Road, Uxbridge, UB9 4DA, United Kingdom, christian@optirisk-systems.com, Gautam Mitra, Christiano Arbex Valle, Robert Fourer Paradigms of ‘Uncertainty Models’, namely, recourse, chance constrained, integrated chance constrained, and robust optimization models, are introduced and represented through use case exemplars. Through our continued research and close collaboration with AMPL Optimization we have now developed AMPL templates and frameworks which can describe these classes of models and connect them to respective solvers, making them readily available to industry- based analysts and to the academic research community. We also describe an E-book (in preparation) which captures the use cases that underpin our goal to make these modelling and solving capabilities widely available to the OR community. 3 - Combining Progressive Hedging And Frank-Wolfe To Solve The Lagrangian Dual Of a Multistage Stochastic Integer Program Natashia Boland, Georgia Institute of Technology, natashia.boland@isye.gatech.edu, Jeffrey Christiansen, Brian Dandurand, Andrew Eberhard, James Luedtke, Jeffrey Linderoth, Fabricio Oliveira We present a new primal-dual algorithm for computing the value of the Lagrangian dual of a stochastic mixed-integer program (SMIP) formed by relaxing its nonanticipativity constraints. The algorithm relies on the progressive hedging method, but unlike previous progressive hedging approaches for SMIP, it converges to the optimal Lagrangian dual value. The key improvement is an inner loop of optimized linearization steps, as in the classical Frank-Wolfe method. Numerical results show that the new algorithm outperforms the standard progressive hedging approach.

4 - First Order Approximation Methods For Estimating Decision Covariance In Stochastic Optimization Sriram Sankaranarayanan, Johns Hopkins University, 3400, N. Charles St, Baltimore, MD, 21218, United States, ssankar5@jhu.edu, Felipe A Feijoo, Sauleh Ahmad Siddiqui We use first order methods to efficiently approximate the covariance matrix of the optimal solution vector. The idea is extended for the covariance of the solution to a complementarity problem by posing the complementarity problem as an unconstrained minimization problem using the Fischer Burmeister merit function. Having done this, we estimate the variability in the estimated Market equilibrium of a Natural gas model, owing to uncertainty in the parameters of the demand function. SB17 105B-MCC Optimization Challenges in Modeling Sponsored: Optimization, Optimization Under Uncertainty Sponsored Session Chair: Leobardo Valera, University of Texas El Paso, 500 W. University Ave, El Paso, TX, 79968, United States, lvalera@miners.utep.edu 1 - Problem Of Moments Over Log-concave Measures Christopher Thomas Ryan, The University of Chicago, Chicago, IL, United States, chris.ryan@chicagobooth.edu, Simai He, Bo Jiang, Teng Zhang We explore the classical discrete problem of moments in a setting where the underlying distribution are discrete log-concave. This situation arises in applied problems such as reliabilit The challenge is that with log-concavity constraints, the problem of moments becomes a non-convex optimization problem. Nonetheless, we are able to characterize optimal extreme point solutions to such problems. Our approach gives rise to improved bounds over the existing literature, as well as analytical insight into the structure of extreme point optimal solutions as geometric distributions. 2 - Interval Constraint Solving Techniques And Model Order Reduction To Enhance The Solution Of Dynamic Systems Leobardo Valera, The University of Texas at El Paso, lvalera@utep.edu, Martine Ceberio Many natural phenomena are dynamic systems, which can often be modeled as nonlinear systems of equations. Major issues with solving such nonlinear systems are that their dimension can be very large and that uncertainty, often present, is tricky to handle. Reduced-Order Modeling (ROM) can address dimension issues, via Proper Orthogonal Decomposition (POD) in particular. On the other hand, interval constraint-solving techniques (ICST) allow to handle uncertainty and ensure reliable results. In this presentation, we show how ICST can be embedded into POD to address dimension and uncertainty.

SB18 106A-MCC HPC in Optimization 1 Invited: High Performance Computing Invited Session

Chair: Kibaek Kim, Argonne National Laboratory, 9700 South Cass Avenue, Building 240, Lemont, IL, 60439, United States, kimk@anl.gov Co-Chair: Geoffrey Malcolm Oxberry, Lawrence Livermore National Laboratory, P. O. Box 808, Livermore, CA, 94551, United States, goxberry@gmail.com 1 - Asynchronous Parallelization Of Decomposition Methods Kibaek Kim, Argonne National Laboratory, kimk@anl.gov We present an approximate and incremental bundle method based on dual decomposition for stochastic mixed-integer programming (SMIP), where the Lagrangian dual function is incrementally updated by approximate subgradients as well as exact ones. We implemented the method in an open source parallel solver DSP, taking advantage of asynchronous communications on HPC cluster using MPI library. We present our computational results on several SMIP problem instances.

48