Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

SD39

3 - The Economics of Bitcoin Ciamac Cyrus Moallemi, Columbia University, Columbia Business School, 3022 Broadway Uris 416, New York, NY, 10027, United States, Gur Huberman, Jacob Leshno The Bitcoin payment system is a platform with two main constituencies: users, who make and receive payments; and profit-seeking miners, who maintain the system’s infrastructure. We seek to understand the economics of the system: How does the system raise revenue to pay for its infrastructure? How are usage fees determined? How much infrastructure is deployed? What are the implications of changing parameters in the protocol? To address these questions, we offer and analyze an economic model of a cryptocurrency system featuring user-generated transaction fees, and focus on Bitcoin as the leading example. The analysis leads to design suggestions for future cryptocurrencies.

optimal policies. Our analysis shows that the optimal policies are structured, and of threshold type. We also provide an application of our model and numerical examples in the context of electric vehicle charging.

n SD40 North Bldg 226B Stochastic Systems Sponsored: Applied Probability Sponsored Session Chair: Mark S. Squillante, IBM Thomas J. Watson Research Center, Yorktown Heights, NY, 10598, United States 1 - On Stochastic Gradient Descent for Distributionally Robust Optimization (DRO) Soumyadip Ghosh, IBM TJ Watson Research Center, 1101 Kitchawan Road, Route 134, Yorktown Heights, NY, 10598, United States, Mark S. Squillante, Ebisa D. Wollega Current approaches to DRO do not scale well because of the high dimensionality of the decision variable of the probability mass function (pmf) over the entire dataset. We propose a new SGD algorithm to efficiently solve these min-max formulations. In each iteration, we approximate the optimization over the uncertainty set by sub-sampling the support, where the size of the sub-sample is itself generated from another pmf. We develop asymptotic guarantees on how this procedure optimally balances, in a strong statistical sense, the computational effort with the required level of accuracy. 2 - Distributionally Robost Optimal Bidding in Online Advertising Jose Blanchet, Stanford University, CA, United States We study practical and easy to implement distributionally robust optimal bidding strategies for online advertising in the context of first price auctions in which the bidder has imperfect information. 3 - Functional Cumulant Moments William A. Massey, Princeton University, ORFE Department, Sherrerd Hall, Princeton, NJ, 08544, United States, Jamol Pender Given a specific measurable real-valued function, functional cumulant moments of a random variable, when appropriately defined, are a new but complementary variation on cumulant moments. We demonstrate their utility to Markovian queueing systems by applying these functional moments to the cumulant moments of birth-death process. These results have applications to the steady state analysis of some given time-homogeneous queueing models as well as the transient behavior of their time-inhomogeneous counterparts. 4 - On Density-dependent Population Processes with Time-varying Behavior Mark S. Squillante, IBM Research, 1101 Kitchawan Road, Yorktown Heights, NY, 10598, United States, Yingdong Lu, Chai W. Wu We consider density-dependent population processes and the control of such stochastic processes, both with time-varying parameters. Such behaviors are of mathematical interest in general, and can be especially important and arise often in many existing and emerging applications.

n SD39 North Bldg 226A Markov Decision Processes Sponsored: Applied Probability Sponsored Session

Chair: Daniel Silva, Auburn Univeristy, Auburn, AL 1 - Pooled Dynamic Matching in Freight Networks Ankur Mani, University of Minnesota, Saint Paul, MN, 55114, United States We study online matching of trucks to shipments over a geographic network. We evaluate the impact of pooling and flexibility of truck drivers on the size of the matching and total welfare. Surprisingly, by only allowing truck drivers to submit their top two preferences, nearly optimal matching can be guaranteed with very high probability. In the dynamic setting, when shipment demands arrive everyday, a small fraction of flexible drivers can increase the total welfare significantly. 2 - Reannealing for Exploration in Deep Q-networks Xing Wang, Auburn Univeristy,Auburn, AL, 36830, United States, Alexander Vinel Existing exploration strategies in reinforcement learning (RL) often either ignore the history or feedback of search, or are complicated to implement. There is also a very limited literature showing their effectiveness over diverse domains. We propose an algorithm based on the idea of reannealing, that aims at encouraging exploration only when it is needed, for example, when the algorithm detects that the agent is stuck in a local optima. The approach is simple to implement, and the experimental results have shown effectiveness of our exploration strategy on accelerating training procedure as well as obtaining a better policy on hard RL problems. 3 - Optimal Pricing for Tandem Queues with Finite Buffers Xinchang Wang, Mississippi State University, Marketing Department, 324C McCool Hall; Mailstop: 9582, Mississippi State, MS, 39762, United States, Sigrun Andradottir, Ayhan Hayriye We consider optimal pricing for a two-station tandem queueing system with finite buffers. The service provider quotes prices to incoming customers using either static or dynamic pricing. The objective is to maximize either the infinite-horizon discounted profit or the long-run average profit. We show that there exists an optimal dynamic policy that exhibits an interesting monotone structure, in which the quoted prices have greater dependency on the queue length at station 1 than at station 2. We show that the optimal static policy performs as well as the optimal dynamic policy for the long-run average problem when the buffer size at station 1 becomes large and the arrival rate is either small or large. 4 - Easy Decomposable Markov Decision Processes Jie Ning, Case Western Reserve University, Department of Operations, 328 Peter B. Lewis Building, Cleveland, OH, 44106, United States We characterize a special class of Markov decision processes (MDPs) called easy decomposable MDPs. These MDPs have vector-valued continuous endogenous states and actions and a set of feasible actions that are decomposable and are independent of the endogenous state. The expected single-period reward and dynamical equations have special structures that allow an analytical characterization of the value function and an optimal policy. Particularly, we show that the value function and optimal policy depend on the solution of a set of auxiliary equations which depend only on the exogenous state. Finally, we give examples of this class of easy decomposable MDPs. 5 - Optimal Routing in Loss Systems with Flexible Customers Avnish Malde, Clemson University, Clemson, SC, United States, Tugce Isik We study a loss system with two classes of servers and a multiple customer types. We assume that there is at least one class of flexible customers that can receive service from servers of either type. We formulate the problem as a Markov decision process with rewards that are dependent on the customer type. For small systems, we use our formulation to characterize the policies that maximize the long-run average reward. For larger systems, we study the structure of the

n SD41 North Bldg 226C

Making Good Decisions Sponsored: Decision Analysis Sponsored Session Chair: Johannes Siebert, MCI 1 - Explaining Proactive Decision Making Reinhard Kunz, Philipp Rolf

The Proactive Decision Making scale, which is based on the concepts of value- focused thinking and decision quality, measures proactive personality traits and cognitive skills in decision-making. Using SEM, we show that proactive cognitive skills can explain up to 36% of life satisfaction, i.e. proactive decision makers are more satisfied with their decisions and with their lives. Furthermore, we provide empirical evidence that proactive cognitive skills can be trained in a course on decision-making. We recommend schools, colleges, and universities to offer more courses on decision-making to enhance student`s proactive cognitive skills and satisfaction with their decisions and lives.

104