Informs Annual Meeting Phoenix 2018

INFORMS Phoenix – 2018

SB53

4 - Opinion Dynamics with Stubborn Agents David S. Hunter, Massachusetts Institute of Technology, Cambridge, MA, United States We consider the problem of optimizing the placement of stubborn agents in a social network in order to maximally impact population opinions. We assume individuals in a directed social network each have a latent opinion that evolves over time in response to social media posts by their neighbors. The individuals randomly communicate noisy versions of their latent opinion to their neighbors. Each individual updates his opinion using a time-varying update rule that has him become more stubborn with time and be less affected by new posts. The dynamic update rule is a novel component of our model and reects realistic behaviors observed in many psychological studies. We show that in the presence of stubborn agents with immutable opinions and under fairly general conditions on the stubbornness rate of the individuals, the opinions converge to an equilibrium determined by a linear system. We give an interesting electrical network interpretation of the equilibrium. We also use this equilibrium to present a simple closed form expression for harmonic influence centrality, which is a function that quanties how much a node can affect the mean opinion in a network. We develop a discrete optimization formulation for the problem of maximally shifting opinions in a network by targeting nodes with stubborn agents. We show that this is an optimization problem with a monotone and submodular objective, allowing us to utilize a greedy algorithm. Finally, we show that a small number of stubborn agents can non-trivially influence a large population using simulated networks. Joint Session AMD/RMP: Machine Learning and Optimization for Automated Mechanism Design Sponsored: Auction and Marketing Design Sponsored Session Chair: Ellen Vitercik, Carnegie Mellon University, Pittsburgh, PA, 15213, United States 1 - Automated Design of High-Revenue Combinatorial Auctions Tuomas W. Sandholm, Carnegie Mellon University, Gates Center for Computer Science, Carnegie Mellon University, Pittsburgh, PA, 15213, United States, Anton Likhodedov Designing revenue-maximizing combinatorial auctions is a key problem, unsolved even for two items. Automated mechanism design [Conitzer & Sandholm UAI- 02] uses computational techniques to design mechanisms. In this paper [Operations Research 2015; extends our AAAI-04 & 05 papers], we introduced two ideas to (automated) mechanism design: 1) search for a good mechanism in parametric families where all mechanisms in the family satisfy desirable properties (e.g. incentive compatibility), and 2) using samples of valuations as input rather than assuming a prior distribution (which can be doubly exponential, thus unrealistic). This begat deterministic mechanisms with highest known revenues. 2 - Mechanism Design for Correlated Valuations: Efficient Methods for Revenue Maximization Michael Albert, Duke University, 308 Research Drive, Durham, NC, 27708, United States, Peter Stone, Vincent Conitzer, Giuseppe Lopomo In this work, we provide a both computationally and sample efficient method to design mechanisms that can robustly incorporate an imprecise estimate of the distribution over bidder valuations, using samples from the true distribution, in a way that provides strong guarantees that the mechanism will perform at least as well as ex-post mechanisms, while also performing nearly optimally with sufficient information. We also demonstrate through simulation that this new mechanism design paradigm generates mechanisms that perform significantly better than traditional mechanism design techniques. 3 - Optimal Auctions through Deep Learning Zhe Feng, Harvard University, Cambridge, MA, United States, Paul D tting, Harikrishna Narasimhan, David C. Parkes In this work, we initiate the exploration of the use of tools from deep learning for the automated design of optimal auctions. The design objective is revenue optimal, DSIC auctions. We show that multi-layer neural networks can learn almost-optimal auctions for settings for which there are known analytical solutions (including results due to Myerson, Manelli-Vincent, Pavlov, Daskalakis et al., ). Moreover, this can be done without appealing to characterization results. We are also able to design essentially optimal auctions for poorly understood problems, as well as obtain state-of-the-art results for combinatorial settings that have been studied in the framework of automated mechanism design. n SB53 North Bldg 232A

4 - A General Theory of Sample Complexity for Multi-item Profit Maximization Ellen Vitercik, Carnegie Mellon University, Pittsburgh, PA, 15213, United States, Maria-Florina Balcan, Tuomas W. Sandholm The design of profit-maximizing multi-item mechanisms is a notoriously challenging problem. The mechanism designer’s goal is to field a mechanism with high expected profit on the distribution over buyers’ values. Unfortunately, if the set of mechanisms he optimizes over is complex, a mechanism may have high empirical profit over a small set of samples but low expected profit. How many samples are sufficient to ensure that the empirically optimal mechanism is nearly optimal in expectation? We uncover structure shared by a myriad of auctions that allows us to prove strong sample complexity bounds: for any set of buyers’ values, profit is a piecewise linear function of the mechanism’s parameters. n SB54 North Bldg 232B Interactions in Supply Chains Sponsored: Behavioral Operations Management Sponsored Session Chair: Michael Becker-Peth, Rotterdam School of Management, Erasmus University, Rotterdam, 3062 PA, Netherlands 1 - Behavioral Biases in Newsvendor Competition: An Indirect Evolutionary Approach Ilkka Leppanen, Loughborough University, School of Business and Economics, Epinal Way, Loughborough, LE11 4NZ, United Kingdom Recent literature has shown that behavioral biases can explain nonnormative decision making of competing newsvendors. We use an indirect evolutionary approach where players maximize biased utility functions but actual profits determine fitness to study 2 behavioral biases, reference effects based on profit comparisons and overconfidence, in newsvendor competition. We show that both kinds of behavioral biases can be evolutionarily stable and lead to increases in inventory levels from the profit-maximizing equilibrium quantities, in line with experimental findings reported in literature. 2 - How to Improve Forcast Perfomance - Are Two Humans Better than One? Michael Becker-Peth, Rotterdam School of Management, Erasmus University, Burgemeester Oudlaan 50, Rotterdam, 3062 PA, Netherlands We investigate how Markeitng and Supply Chain Department interact in sharing forecasting data in a real world setting. We find that when Marketing is in charge of editing the SAP forecast, they inflate substantialy. After inserting Supply Chain Department as middle man, the performance increases becasue SCM anticipates the forecast inflation and reduces the effect. 3 - Strategic Inventories in Dual Channel Yan Lang, University of Texas at Arlington, 701 S. Nedderman Dr,, Arlington, TX, 76019, United States, Kay-Yut Chen We study the effect of choosing to strategic withhold inventories in a perfect information dual channel environment, which contains one supplier and one retailer. We design a series of laboratory experiments, which subjects are facing different market power and inventory holding cost issues, to test our hypotheses of equilibrium strategies. 4 - An Empirical Analysis of Buyer Strategic Ordering Decisions through a Behavioral Experiment Minseok Park, Salisbury University, 1101 Camden Ave. Perdue Hall 334, Salisbury, MD, 21801, United States, Pelin Pekgun, Sriram Venkataraman, Manoj Malhotra Theoretical literature shows that certain allocation policies can reduce order inflation behavior of buyers, while empirical evidence is scarce. In this study, through behavioral lab experiments, we analyze the impact of a capacity allocation policy and different information disclosures on the buyer ordering decisions in a single-supplier, two-buyer supply chain.

50