Informs Annual Meeting 2017

TE04

INFORMS Houston – 2017

2 - A Computer Plays the Beer Game: A Deep Reinforcement Learning Algorithm for Inventory Optimization Lawrence V. Snyder, Lehigh University, Mohler Lab 200 West Packer Avenue, Bethlehem, PA, 18015-1582, United States, lvs2@lehigh.edu, Afshin Oroojlooy Jadid, Mohammadreza Nazari, Martin Takac We propose an algorithm to play the beer game with “teammates” that are humans or other algorithms. The beer game is a multi-agent serial supply chain in which agents attempt to minimize network cost while each agent only observes its local information. We develop a Deep Q-Network algorithm to solve this problem. We show that our algorithm outperforms approaches from literature and unlike the majority of them does not impose restrictions on the problem parameters, and also provides good solutions even if other agents do not follow rational policies. The algorithm can be extended to any decentralized multi-agent cooperative game with partial information, which is a common situation in supply chains. 3 - Stress on Operational Risk: Empirical Evidence from Commercial Bank Operational risk is now one of the three most important risks in the financial services industry. This paper studies how workload affects bank operational risk events frequency and severity. To achieve this goal, we use a unique operational risk event data set from a commercial bank that contains 1441 operational risk events in two years. We find that workload has a U-shape impact on operational risk frequency. We then proceed to discuss bank capital allocation on staffing level among branches so as to reduce operational risk losses. We compare our optimal staffing policy with bank’s original policy with both in-sample and out-of-sample studies. 320A Renewable Energy and Energy Efficiency Sponsored: Manufacturing & Service Oper Mgmt, Sustainable Operations Sponsored Session Chair: afak Yücel, safak.yucel@georgetown.edu 1 - Business Models for Off-Grid Energy Access at theBottom of the Pyramid Bhavani Shanker Uppari, INSEAD, 1 Ayer Rajah Avenue, Singapore, 138676, Singapore, BhavaniShanker.Uppari@insead.edu, Ioana Popescu, Serguei Netessine One in every five people people does not have access to electricity, relying mostly on kerosene for light. Solar technologies are healthier and offer greater value, yet they require significant one-time investments which are not affordable to people living on 2 USD per day. Even though cheaper rechargeable lighting technologies are available, their adoption is low and some consumers still use kerosene. We develop a consumer behavior model that accounts for income variability and liquidity constraints specific to impoverished markets, and investigate alternative business models, based on a case study in Rwanda. 2 - Investments in Renewable and Conventional Sources: The Role of Operational Flexibility Safak Yucel, Georgetown University, Georgetown University, 3700 O St. NW, 523 Hariri Building, Washington, DC, 20057, United States, safak.yucel@georgetown.edu, Gurhan Kok, Kevin Shang We study capacity investments of a utility firm in renewable and conventional energy sources with different levels of operational flexibility, i.e., the ability to quickly ramp up or down the output of a generator. We consider supply characteristics of conventional and renewable sources and derive the optimal capacity investment portfolio. We find that inflexible sources (e.g., nuclear energy) and renewables are substitutes; flexible sources (e.g., natural gas) and renewables are complements. 3 - Utility Ownership of Decentralized Combined Heat and Power Eric Webb, PhD Student, Indiana University, 2915 W. Winterberry Ct., Bloomington, IN, 47404, United States, eric.michael.webb@gmail.com, Owen Wu, Gilvan Souza Combined heat and power (CHP) plants generate electricity and capture excess heat for space heating, water heating, or industrial processes. CHP plants reach 70-80% efficiency, improving on the 40-50% efficiency of natural gas combined cycle or the 30-35% efficiency of coal. A CHP plant, located at a firm with large heating needs, may provide reliability benefits by continuing to operate during a grid outage. Despite these benefits, CHP plants are rarely owned by utilities in the United States. We study the economics of utility ownership of CHP, determining when it is valuable to the utility, to the on-site firm, and to society. We examine which regulatory policies encourage CHP ownership by utilities. Yuqian Xu, University of Illinois at Urbana-Champaign, Champaign, IL, United States, lillian.xyq@gmail.com, Fangyun Tan, Serguei Netessine TE04

4 - Strategic Forward Trading and Technology

Heikki Peura, Imperial College Business School, South Kensington Campus, London, SW7 2AZ, United Kingdom, h.peura@imperial.ac.uk, Derek W. Bunn Motivated by the electricity industry’s transition towards renewable power, we analyze how operational factors of production, such as its flexibility and reliability, can influence market prices indirectly through altering the balance of spot and forward trading. We show, for example, that increasing the capacity of intermittent renewable electricity generation, despite its lower marginal production cost, may not necessarily reduce power prices. 320B Data-Driven and Dynamic Decision-Making for Healthcare Policies Sponsored: Health Applications Sponsored Session Chair: Alireza Boloori, Arizona State University, Tempe, AZ, 85283, United States, aboloori@asu.edu 1 - Optimal Scheduling of Medical Diagnostics Arkajyoti Roy, Northwestern University, Evanston, IL, 60208-3119, United States, aroy@bgsu.edu, Omid Nohadani In medical decision making, clinical information on patient is collected from a history or physical examinations. To refine and to track the progression of care, often one or more diagnostic tests are conducted during the treatment. The timing of these tests affects the understanding and influences the efficacy of decisions. In this work, we provide a framework for optimal schedules via robust optimization by modeling an evolving uncertainty set in time. We apply the approach to cancer radiation therapy. The findings are general and relevant to a broad range of applications, e.g., maintenance scheduling. 2 - Using a Branch-and-bound Scheme with Multi-fidelity Models to Optimize Hepatitis C Screening and Treatment Policies under a Budget Constraint We propose an efficient algorithm for Hepatitis C (HCV) screening and treatment allocation problem with disease progression, budget constraint, and population evolution. We first present a low fidelity model by approximating future disease progression with a stationary Markov model. We then design a combined rollout algorithm embedding the low fidelity model into a high fidelity model with disease progression to efficiently identify a good incumbent policy. Lastly, we incorporate the incumbent and a theoretical upper bound into a branch-and- bound scheme. 3 - Optimal Breast Cancer Diagnostic Decisions Based on Disease Subtypes to Reduce Overdiagnosis Sait Tunc, University of Chicago, Chicago, IL, United States, sait.tunc@chicagobooth.edu, Oguzhan Alagoz, Elizabeth S. Burnside Mammography, which reduces breast cancer mortality, has several negative effects such as overdiagnosis. Overdiagnosis may be reduced if indolent breast cancer subtypes can be identified and followed with noninvasive imaging. We develop a large-scale MDP to optimize the post-mammography diagnostic decisions based on disease subtypes. We present an efficient and easily implementable algorithm to find the exact optimal solution. We reduce the computational complexity of the algorithm by obtaining feasible upper bounds for the optimal decision thresholds using dimension reduction. 4 - Operational Policy Recommendations for Treatment Protocol Restrictions at a County Hospital: A Process-based Case Study Olga Bountali, Southern Methodist University, 4210 Fairmount Street, Apartment 3063, Dallas, TX, 75219, United States, obountali@smu.edu, Farnaz Nourbakhsh, Sila Cetinkaya, Vishal Ahuja County hospitals serving the indigent are subject to treatment protocol restrictions under which the uninsured are admitted only if the patient’s clinical condition is evaluated as life threatening. The underlying goal of such restrictions is improved oversight and efficiency. Motivated by clinical observations, we present a process-based case study and offer analytical and numerical results for a comparative analysis of the impact of protocol restrictions on patient- and systems-level metrics. Counterintuitive to the common belief, protocol restrictions may be myopic. TE05 TingYu Ho, University of Washington, Industrial and Systems Engineering, Box 352650, Seattle, WA, 98195-2650, United States, tyhotw@uw.edu, Shan Liu, Zelda B.Zabinsky

410