INFORMS 2021 Program Book

INFORMS Anaheim 2021

MD30

2 - A Stochastic Variance-reduced Accelerated Primal-dualmethod For Finite-sum Saddle-point Problems Erfan Yazdandoost Hamedani, University of Arizona, State College, PA, 16801-4415, United States, Afrooz Jalilzadeh In this talk, we propose a variance-reduced primal-dual algorithm for solving convex-concave saddle-point problems with finite-sum structure and nonbilinear coupling function. This type of problem typically arises in machine learning and game theory. Compared with existing methods, our framework yields a significant improvement over the number of required primal-dual gradient samples to achieve an epsilon-accuracy of the primal-dual gap. We implemented our method for solving a distributionally robust optimization problem to show the effectiveness of the proposed algorithm. 3 - Using Deep Reinforcement Learning for Solving the Stochastic Capacitated Lot Sizing Problem Lotte van Hezewijk, Eindhoven University of Technology, Eindhoven, 3526 WD, Netherlands, Lotte van Hezewijk, ORTEC, Zoetermeer, Netherlands, Nico P. Dellaert, Noud Gademann We study a multi-item stochastic capacitated lot sizing problem. Inspired by industrial cases, we consider a limited production capacity, stochastic demand, and setup times. The objective is to determine the production quantities, while minimizing the costs of inventory, backorders and production. We use a Deep Reinforcement Learning (DRL) methodology to find solutions. Larger problem instances encounter some challenges, which we resolve by utilizing domain knowledge to support the DRL algorithm. MD29 CC Room 207C In Person: Distributionally Robust Optimization General Session Chair: Soroosh Shafieezadeh Abadeh, EPFL, Ecublens, 1024, Switzerland 1 - Two-stage Data-driven Distributionally Robust Optimization With Random Recourse Xiangyi Fan, U T-Austin, Austin, TX, United States We study two-stage data-driven stochastic optimization problems with random recourse where the adaptive decisions are multiplied with the uncertain parameters in both the objective and the constraints. We propose a scalable approximation scheme via piecewise linear and piecewise quadratic decision rules. The emerging decision rule problems can be reformulated as exact copositive programs, which admit tractable approximations in semidefinite programming. To address the inefficiency of solving large-size semidefinite programs, we design a decomposition algorithm where smaller-size subproblems can be solved in parallel. We further establish the performance guarantees of the proposed scheme and demonstrate its effectiveness through numerical examples. 2 - First-order Methods for Distributionally-Robust MDPs Christian Kroer, Columbia University, New York, NY, 10027-6623, United States, Julien Grand-Clement Markov decision processes (MDPs) are known to be sensitive to parameter specification. Distributionally robust MDPs alleviate this issue by allowing for ambiguity sets which give a set of possible distributions over parameter sets. The goal is to find an optimal policy with respect to the worst-case parameter distribution. We propose a framework for solving Distributionally robust MDPs via first-order methods and instantiate it for several types of Wasserstein ambiguity sets. By developing efficient proximal updates, our algorithms achieve convergence rates that are significantly better than existing value iteration methods. Numerical experiments show that our algorithm is significantly more scalable than state-of-the-art approaches across several domains. 3 - Optimal Transport Based Distributionally Robust Optimization Soroosh Shafieezadeh-Abadeh, ETH Zurich, Zurich, Switzerland, Liviu Aolaritei, Daniel Kuhn, John Lygeros, Florian Dorfler We show that the ordinary use of the Wassersyein type-p distance in DRO problems is not suitable even for simple loss functions. We then propose an optimal transport based DRO approach with a general transportation cost. In this general setting, the new DRO problem can be viewed as a zero-sum game. We prove that this zero-sum game admits a Nash equilibrium. We then proceed and explore the relationship between the distributional robustness and its regularization effect. In particular, we establish a link between the DRO setting and the use of high-order variation regularization, and then, we propose a simple dual formulation of the DRO problem for the class of generalized linear models using techniques in nonconvex optimization. This formulation enables us to both analyze the equivalency between the distributional robustness and its implicit/explicit regularization effect.

MD30 CC Room 207D In Person: Advanced Maintenance Models General Session Chair: Yisha Xiang, Texas Tech University, Lubbock, TX, 79409, United States Co-Chair: Ying Liao, Texas Tech University, Lubbock, TX 1 - Optimal Condition-based Maintenance for Assets Dispersed on a Graph Shadi Sanoubar, University of Pittsburgh, Pittsburgh, PA, 15207- 1176, United States, Bram de Jonge, Lisa M. Maillart, Oleg A. Prokopyev This talk is concerned with providing condition-based maintenance via a single maintenance resource to a set of geographically distributed assets. We use graph representation to model possible geographical locations, including idling and asset locations and the links between them. We formulate a Markov Decision Process to dynamically obtain the optimal positioning of the maintenance resource and the optimal timing of the interventions that the resource performs. We explore how the underlying graph structure impacts the maintenance thresholds and the locations most used for idling under the optimal policy, as well as the performance metrics such as resource utilization and asset downtime. 2 - Risk And Resilience-based Optimal Post-disruption Repair for Critical Infrastructures under Uncertainty Haitao Liao, University of Arkansas, Fayetteville, AR, 72703-9301, United States, Basem Alkhaleel, Kelly Sullivan Post-disruption restoration of critical infrastructures (CIs) often faces uncertainties associated with the required repair tasks and the related transportation network. In this paper, two-stage risk-averse and risk-neutral stochastic optimization models are proposed to schedule repair activities for a disrupted CI network with the objective of maximizing system resilience. Both models are developed based on a scenario-based optimization technique that accounts for the uncertainties of the repair time and the travel time spent on the underlying transportation network. An improved fast forward algorithm based on a wait-and-see solution methodology is provided to reduce the number of chosen scenarios. To assess the risks associated with post-disruption scheduling plans, a conditional value-at-risk metric is incorporated into the optimization models. 3 - Maintenance Optimization of an Offshore Wind Turbine Subject to Weather Conditions Morteza Soltani, Clemson University, Clemson, SC, United States, Jeffrey P. Kharoufeh, Amin Khademi We consider the maintenance optimization of an offshore wind turbine, where the feasibility of performing maintenance depends on the weather condition. The turbine’s degradation evolves as a Markov chain, and the objective is to minimize the sum of the expected total setup, replacement and downtime costs over a finite horizon. We devise a Markov decision process model and establish the existence of a threshold policy, as well as monotonicity of the value function and optimal policy. A novel approach for theoretical sensitivity analyses of key model parameters is also presented. 4 - Prognosis Analysis of Breast Cancer Based on Dirichlet Process Mixture Models Ying Liao, Texas Tech University, Lubbock, TX, 79415-5119, United States, Yisha Xiang, Di Ai, Ning Dong Breast cancer patients in a particular subgroup often have common disease progression pattern that leads to similar survival outcomes. It is of great importance to identify such subgroups because effective treatments can be developed for the patients based on their corresponding prognostic information. In clinical practices, the number of subgroups is generally unknown and it is also challenging to model the relationships between the group labels and various prognostic factors, such as age at diagnosis, estrogen and progesterone receptors status. In this work, we propose a novel clustering framework to probabilistically label the patients based on the Dirichlet process mixture models. Given the labels, we identify significant prognostic factors using advanced machine learning algorithms and provide some insights for clinical practitioners.

81