Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

TA03

Tuesday, 7:30AM - 9:00AM

n TA02 North Bldg 121B Robust Optimization – Dynamic and Discrete Considerations Sponsored: Optimization/Optimization Under Uncertainty Sponsored Session Chair: Dan Andrei Iancu, Stanford University,Stanford, CA, 94107, United States 1 - Distributionally Robust Linear and Discrete Optimization with Marginals Peng Xiong, National University of Singapore, Singapore, Singapore, Zhi Chen, Melvyn Sim We study linear and discrete optimization problems in which the objective coefficients are chosen randomly from a distribution, and the goal is to find robust bounds on the expected optimal value and the marginal distribution of the optimal solution. The set of joint distributions is assumed to be specified up to only the marginal distributions. With a primal-dual formulation, we generalize previous assumptions on the marginals to arbitrary marginals using optimal transport theory, and we identify a sufficient condition for polynomial time solvability using extended formulations. We conclude by exploring the implications on the solvability of problems in areas such as scheduling and max- flow. 2 - Adaptive Robust Optimization with AROMA Peng Xiong, National University of Singapore, SIngapore, Singapore, Zhi Chen, Melvyn Sim We present a tractable framework for adaptive robust optimization. The proposed framework could naturally unify classical stochastic programming, robust optimization, and more recent distributionally robust approaches with a new scenario-wise ambiguity set. Such an ambiguity set inspires a scenario-wise recourse approximation that provides tractable solutions to adaptive problems. Based on the new framework, we have developed a generic software package, named AROMA (Adaptive Robust Optimization Made Accessible) for a straightforward implementation of the proposed framework. Experiments are conducted to demonstrate the effectiveness of the proposed framework and software package. 3 - Dynamic Robust Optimization and Discrete Convexity Dan Andrei Iancu, Stanford University, 655 Knight Way, Stanford, CA, 94107, United States, Yehua Wei We discuss necessary and sufficient conditions for the optimality of specific classes of policies in dynamic robust optimization. We then focus on the specific case of affine policies, and show how our conditions can be used to recover and generalize several existing results in the literature. Our treatment draws interesting connections with the theory of discrete convexity (L-natural / M- natural convexity and multimodularity) and global concave envelopes, which may be of independent interest. Time permitting, we also discuss some related applications of the results in the context of a learning and stopping problem. 4 - Robust Dynamic Optimization for R&D Project Selection Aurelie Thiele, Southern Methodist University We investigate adaptive decision rules in multi-stage binary programming in the context of R&D project selection, where a manager can decide whether to continue or abandon projects, subject to future cash flow uncertainty represented using a robust optimization paradigm. The manager can choose from projects with various risk-return profiles. Information about cash flow uncertainty is updated over time. We study the computational tractability of the approach and present theoretical insights. Numerical experiments are encouraging. n TA03 North Bldg 121C Managing Innovation I Sponsored: Technology, Innovation Management & Entrepreneurship Sponsored Session Chair: Wenli Xiao, University of San Diego, San Diego, CA, 92110, United States 1 - The Impact of Social Orientation on Firm Innovation Xiaojin Liu, University of Virginia, Darden, Charlottesville, VA, 22903, United States, Raul Chao This study addresses the questions of whether and how firm social orientation influences firm innovation. In so doing, we propose mechanisms geared toward either proactive or reactive social orientation in project funding. Integrating large scale archival datasets, we empirically investigate the long term impact of social orientation on firm innovation.

n TA01 North Bldg 121A Data-Driven and Robust Optimization I Sponsored: Optimization/Optimization Under Uncertainty Sponsored Session Chair: Phebe Vayanos, University of Southern California, Los Angeles, CA, 90089, United States 1 - Distributionally Robust Strategic Queues Madhushini Narayana Prasad, The University of Texas at Austin, Austin, TX, 78712, United States, Grani Adiwena Hanasusanto, John Hasenbein We extend the classical Naor’s model for observable systems by relaxing the principal assumption of a deterministic arrival rate. While the Markovian assumptions still hold, we assume the arrival rate is uncertain and only the information about first-order moment is known to the strategic customers and managers. We derive the optimal threshold strategies of a revenue maximizer, a social optimizer and an individual customer, by maximizing the worst-case expected profit over the set of distributions satisfying the known information. We then compare and contrast our observations with the classical model results, and finally present our computational results. 2 - Robust Factored Markov Decision Process Man-Chung Yue, Imperial College London, London, United Kingdom Factored Markov decision process where the transition probabilities are uncertain is considered. We first provide a mathematical framework and discuss the difficulties of this class of problems. In particular, the optimization problem associated with the value function of any given policy has exponentially many constraints and variables and suffers from multi-linearity. Based on a novel rectangularity assumption, we construct a lifted MDP whose value function is the equivalent to that of the original factored MDP but do not suffer from multi- linearity. Efficient approximation is then obtained by combining this result with techniques such as linear apporixmation to reduce the exponentiality. 3 - Data-driven Approach for Demand Forecasting and Inventory Control of Slow-moving Items Sheng Bi, National University of Singapore, Singapore, 138600, Singapore, Long He, Chung-Piaw Teo We consider the demand forecasting and inventory control problem with intermittent usage. The data is often used to estimate the demand distribution if the item is ordered, and also the probability that an order will materialize in a period. However, errors in the estimation may skew the performance of this approach. In this paper, we use a portfolio approach to hedge the risk of inventory exposure between successive order arrivals, using as input an estimate of the joint distribution of demand and interarrival time. To account for errors in estimation, we propose a distributionally robust model using the Wasserstein uncertainty set to determine the state-dependent order-up-to levels. 4 - Robust Reinforcement Learning Huan Xu, ISyE Ga Tech, 755 Ferst Drive, NW,, Atlanta, GA, 30332, United States, Aurko Roy, Sebastian Pokutta We study reinforcement learning under model misspecification, where we do not have access to the true environment but only to a reasonably close approximation to it. We address this problem by extending the framework of robust MDPs to the model-free Reinforcement Learning setting, where we do not have access to the model parameters, but can only sample states from it. We define robust versions of Q-learning, SARSA, and TD-learning and prove convergence to an approximately optimal robust policy and approximate value function respectively. We scale up the robust algorithms to large MDPs via function approximation and prove convergence under two different settings.

243