Informs Annual Meeting 2017

SC82

INFORMS Houston – 2017

SC80

We consider the dynamic assignment of multiple workers to multiple classes of jobs. Workers are heterogeneous and become more effective at a job class with experience, and therefore the scheduler faces a tradeoff between matching workers with familiar jobs to minimize short-term cost versus matching workers with unfamiliar jobs to build expertise. A deterministic version of the problem has distinct structure that we leverage to derive price-directed approximations to stochastic versions featuring randomly arriving workers and jobs. 2 - Learning Symmetries for Faster Reinforcement Learning Theja Tulabandhula, UIC, Chicago, IL, 60607, United States, tt@theja.org, Anuj Mahajan We explore methods to exploit symmetries for achieving sample efficiency in reinforcement learning. This problem is motivated by recent advances in the use of deep networks for complex tasks that end up requiring large amounts of training data. We propose a method for detecting symmetries using reward trails observed during episodic experiences and provide a framework for incorporating these discovered symmetries in functional approximation based architectures. Finally we show that the use of potential based reward shaping is especially effective for our symmetry exploitation mechanism. Experiments show that our methods improve performance significantly when symmetry is exploited. 3 - Approximations to Stochastic Dynamic Programs via Information Relaxation Duality A common technique in the analysis of stochastic systems is the use of “hindsight bounds,” in which decisions are made after all uncertainties are revealed. In some applications, however, hindsight bounds may lead to very weak performance guarantees. We show how to obtain stronger guarantees by incorporating penalties that punish the use of additional information, and demonstrate the technique on several applications, including stochastic knapsack problems. 4 - Revisiting Approximate Linear Programming using Root Finding and First Order Methods Selvaprabu Nadarajah, College of Business, University of Illinois at Chicago, 601 South Morgan Street,, U. H.2406, Chicago, IL, 60607, United States, selvan@uic.edu, Qihang Lin, Negar Soheili Azad Approximate linear programs (ALPs) are well-known models for approximating high dimensional Markov decision processes (MDPs). Solving ALPs exactly remains challenging, for instance, in applications where (i) the MDP includes nonlinear reward and transition dynamics and/or (ii) rich basis functions are required to obtain a good VFA. We address this tension between ALP theory and solvability by proposing a novel ALP reformulation and solution approach that combines root finding and first order methods. We present convergence guarantees for our approach and test it on inventory control and energy storage applications. 382B New Paradigms in Optimization Under Uncertainty Sponsored: Optimization, Optimization Under Uncertainty Sponsored Session Chair: Grani Adiwena Hanasusanto, The University of Texas at Austin, ETC 5.120, Austin, TX, 78712, United States, grani.hanasusanto@utexas.edu 1 - “Dice”-sion Making under Uncertainty: When can a Random Decision Reduce Risk? Erick Delage, HEC Montreal, 3000 Ch de la Cote-Sainte-Catherine, Montreal, QC, H3T.2A7, Canada, erick.delage@hec.ca, Daniel Kuhn, Wolfram Wieseman Consider an Ellsberg experiment in which one can win by calling the color (red or blue) of the ball that will be drawn from an urn in which the balls are of unknown proportions. It is well known (yet rarely advertised) that selecting the color based on a fair sided coin completely eradicates the ambiguity about the odds of winning. In this talk, we explore what are conditions under which a decision maker that employs a risk measure should have his action depend on the outcome of an independent random device. Surprisingly, we show that for any ambiguity averse risk measure there always exists a decision problem in which a randomized decision strictly dominates all deterministic decisions. SC82 David Brown, Duke University, Fuqua School of Business, 100 Fuqua Drive, Durham, NC, 27708, United States, dbbrown@duke.edu, Santiago Balseiro

381C Energy System Optimization Sponsored: Energy, Natural Res & the Environment, Energy Sponsored Session Chair: Kai Pan, University of Florida, Gainesville, FL, 32611, United States, kpan@ufl.edu 1 - Design and Dispatch Optimization of a Solid-oxide Fuel Cell Assembly for Unconventional Oil and Gas Production We present a multi-objective design and dispatch optimization model of a solid- oxide fuel cell (SOFC) assembly for unconventional oil and gas production. Heat produced by SOFCs during electricity generation is used to pyrolyze kerogen into a mixture of oil, hydrocarbon gas and carbon-rich shale coke. The Geothermic Fuel Cell (GFC) system costs, heating and electrical efficiencies are optimized, subject to geology heating demands, auxiliary component electric power demands and the GFC system performance characteristics. The resulting optimal cost of oil and gas produced using the GFC technology is about $39/bbl which is comparable to that from other unconventional oil extraction techniques. 2 - The Community Microgrid Distribution System of the Future Lei Wu, Clarkson University, 8 Clarkson Avenue, P.O. Box 5720, Potsdam, NY, 13676, United States, lwu@clarkson.edu Community microgrids (CMG) connect critical loads and distributed energy resources of multiple owners for providing reliable and resilient electricity services through distribution lines owned by a local power utility company. Thus, different from current single-owner microgrids, CMGs present significant new design goals, operating constraints, and business models. This presentation discusses key characteristics, identifies operation challenges, and proposes potential solution methodologies for the operation and control of the future CMGs. A practical community resilience microgrid that is underway for serving Potsdam, NY will also be discussed. 3 - Pricing Electricity Considering Non-convexity and Uncertainty in the Day-ahead Pool Market Jianhui Wang, Southern Methodist University, Dallas, TX, United States, jianhui@smu.edu Jianhui Wang, Argonne National Laboratory, Argonne, IL, United States, jianhui@smu.edu A new market clearing mechanism for the day-ahead market is proposed in this paper, which considers both the non-convexity caused by the unit commitment decisions, and the uncertainty of real-time operation. By applying this method, sufficient flexibility can be reserved through proper commitment decisions. Meanwhile, every dispatched unit is guaranteed for adequate revenue to recover costs without additional side payments. A primal-dual approach associated with specific revenue adequacy rules is applied to formulate the problem. 4 - Optimization of Food and Biofuel Production with Economic and Environmental Impacts Halil Cobuloglu, Senior Analytics Fellow, McKinsey & Company, 404 wyman st Suite 300, Boston, MA, 02451, United States, halil.cobuloglu@gmail.com, Esra Buyuktahtakin Biofuel production from food crops leads to debates about the security of the food supply while biofuels offer positive environmental impacts. To solve this trade-off, we propose a multi-objective mixed-integer optimization model to investigate the economic and environmental impacts of biofuel and food production using switchgrass and corn. This model provides optimal decisions regarding land allocations to food and energy crops, seeding time, harvesting time and amount, and budget allocations to farm operations. Finally, we present a two-stage MIP model that maximizes the economic and environmental benefits of food and biofuel production. Alexandra M. Newman, Colorado School of Mines, 1500 Illinois St, Golden, CO, 80401, United States, anewman@mines.edu, Gladys Anyenya, Neal Sullivan, Robert Braun 382A Approximate Dynamic Programming for Stochastic Optimization Sponsored: Optimization, Optimization Under Uncertainty Sponsored Session Chair: Selvaprabu Nadarajah, College of Business, University of Illinois at Chicago, College of Business, University of Illinois at Chi, Chicago, IL, 60607, United States, selvan@uic.edu 1 - Dynamic Matching of Workers to Jobs with Learning Curves Daniel Adelman, University of Chicago, Booth School of Business, 5807 South Woodlawn Avenue, Chicago, IL, 60637, United States, dan.adelman@chicagobooth.edu, Adam J. Mersereau SC81

97