INFORMS 2021 Program Book

INFORMS Anaheim 2021

ME21

ME19 CC Room 203A In Person: Deep Learning and Mechanism Design General Session Chair: Michael Albert, Assistant Professor, University of Virginia, Charlottesville, VA 1 - Auction Learning as a Two Player Game Jad Rahme, Princeton University, Princeton, NJ, United States, Sami Jelassi, S. Matthew Weinberg Designing an incentive compatible auction that maximizes expected revenue is a central problem in Auction Design. While theoretical approaches have hit some limits, a recent research direction initiated by Duetting et al. (2019) consists in building neural network architectures to find optimal auctions. We propose two conceptual deviations from their approach. First, we introduce a time- independent Lagrangian inspired by recent results in auction design theory. Second, we amortize the optimization procedure used in previous work to compute optimal misreports with the introduction of an additional neural network. We show the effectiveness of our approach by learning competitive or strictly improved auctions compared to prior work. Both results together further imply a novel formulation of Auction Design as a two-player game with stationary utility functions. 2 - Deep Learning for the Automated Design of Two-sided Matching Markets Sai Srivatsa Ravindranath, Harvard University, Cambridge, MA, United States, Zhe Feng, Shira Li, Jonathan Ma, Scott Kominers, David C. Parkes Economic theory provides the celebrated deferred acceptance mechanism for the design of a stable, two-sided matching mechanism that is also strategy-proof on one side of the market. At the same time, it is provably impossible to achieve both stability and strategy-proofness simultaneously. But there is little understanding in regard to how to navigate the necessary tradeoffs between these two properties. In this talk, we demonstrate how deep neural networks can be used for the automated, data-driven design of matching mechanisms that strike new tradeoffs between stability and incentive alignment, expanding the efficient frontier and suggesting new targets for economic theory. 3 - Certifying Strategyproof Auction Networks Michael J. Curry, University of Maryland, College Park, MD, United States The design of strategyproof, revenue-maximizing auctions is a classic goal of mechanism design, but in multi-agent, multi-item settings, progress has been very limited. This has motivated attempts to train deep neural networks to approximate optimal auctions. These approaches work well empirically. However, there is no way to be completely sure that they are actually strategyproof. By drawing connections between strategyproofness and adversarially robust machine learning, we devise a modified auction architecture for which it is possible to compute exact, tight certificates on the degree to which strategyproofness is violated for a given bid profile. We find that our approach is effective, though scalability could be improved, and that gradient-based approximations do underestimate the extent of violations. 4 - Provable Lower Bounds for Black Box Mechanism Design Michael Albert, Assistant Professor, University of Virginia, Charlottesville, VA, 22903-1416, United States, Minbiao Han The field of mechanism design has had significant success in constructing optimal mechanisms to allocate resources when there are information asymmetries. However, there are many situations under which no optimal mechanism is known, leading to the adoption of black box optimizers, such as deep neural networks, to learn good mechanisms. However, these learned mechanisms only approximately satisfy traditional mechanism design guarantees, such as incentive compatibility. Given that these mechanisms fail traditional mechanism design guarantees, they cannot guarantee any lower bound on their performance. In this work, we present a procedure where by having sample access to a mechanism we can prove a lower bound on the performance. Moreover, we develop new techniques to construct mechanisms using deep neural networks that provide good lower bounds on the performance.

ME20 CC Room 203B In Person: OR for Equity in Health and Society General Session Chair: Caleb Bugg, University of California-Berkeley, Berkeley, CA, 94704, United States 1 - Applications Of Nonnegative Tensor Completion Via Integer Optimization Caleb Bugg, University of California-Berkeley, Berkeley, CA, 94720, United States There is an unresolved tension in the literature on tensor completion. One set of approaches has polynomial-time computation but requires exponentially more samples than the information-theoretic rate, whereas another set of approaches achieves the information-theoretic rate but requires solving NP-hard problems for which there are no known numerical algorithms to compute global minima. This paper resolves this tension for nonnegative tensors by developing a numerical algorithm that provably converges to a global minima in a linear (in numerical tolerance) number of oracle steps while achieving the information-theoretic rate. We then show the usefulness of our algorithm on health-related data. 2 - Persistent Inequities and Systemic Racism Jonathan W. Welburn, RAND Corporation, Santa Monica, CA, 90401, United States Persistent inequities and systemic racism have resulted in the current American disparity where Black households hold a tenth the wealth of their white peers. Calls for reparations from have regained significant attention as a solution for addressing the harms of slavery and of the de jure segregation and institutionalized discrimination that followed. We introduce a microsimulation model of intergenerational wealth to analyze the potential impact of policy interventions, from reparations to baby bonds, and their affect on the Black-white wealth gap. ME21 CC Room 204A In Person: Valuing External and Collaborative Innovation General Session Chair: Zeya Wang, PhD Student, Atlanta, GA, 30308-1192, United States 1 - Leveraging the Potential of Outsourcing and Offshoring in Complex Product Development Ole Frauen, Volkswagen AG, Wolfsburg, 38100, Germany, Arnd H. Huchzermeier, Jurgen Mihm Leveraging the potential of outsourcing and offshoring remains a major challenge in complex product development. It is a question about effectively decomposing and distributing work across geographical and organizational boundaries while providing high quality products. The decisions must clearly depend on the product’s characteristics and the emerging collaborative network. The conducted study is based on an extensive data set involving all development projects of one of the largest car manufacturers worldwide. We demonstrate that a precise distinction between outsourcing and offshoring, as well as the introduction of a network perspective, are crucial to evaluate the effects in more detail. 2 - An Entrepreneur’s Innovation Dilemma: Learning-financing Tradeoff at Lean Startups Onesun Steve Yoo, University College London, London, United Kingdom, K Sudhir Using a Bayesian learning model of lean startup and a Nash bargaining game between the investor and entrepreneur, this paper examines entrepreneur’s trade-off between optimizing learning for the startup’s success relative to the need to convey market potential by producing observable success signal for early-stage investors. We find that depending on the entrepreneur’s relative bargaining position, the entrepreneur may distort product development downwards or upwards to sacrifice learning relative to the one prescribed in the Lean startup method. The two types of distortions differently impact the innovation efficiency and innovation output. We examine how they could be mitigated and collectively improve the innovation economy.

93