INFORMS 2021 Program Book

INFORMS Anaheim 2021

ME11

different artificial neural networks to test their predictive performance empirically and improve valuations. 3 - Polyhedral Relaxations for Nonlinear Univariate Functions Kaarthik Sundar, Los Alamos National Laboratory, Los Alamos, NM, United States This talk will present linear programming and mixed-integer linear programming relaxations for the graph of a univariate, bounded, nonlinear function f(x) that converge to its convex hull and the its graph, respectively. Theoretical convergence guarantees and rates of convergence will be discussed. Efficacy of utilizing these relaxations in global optimization algorithms to solve Mixed Integer Non-Linear Programs (MINLPs) will be shown. Furthermore, we will also show techniques to use these relaxations to build convex relaxations for non- linear on-off constraints. 4 - Algorithms for Difference-of-Convex Programs Based on Difference-of-Moreau-Envelopes Smoothing Kaizhao Sun, Georgia Institute of Technology, Atlanta, GA, United States, Xu Andy Sun We consider minimization of a difference-of-convex (DC) function with and without linear constraints. We first study a smooth approximation of a DC function termed difference of Moreau envelopes (DME), which captures geometric properties of the original function and enjoys some growth conditions. Then we show that the gradient descent method and an inexact variant converge on the DME and deliver a stationary solution of the original DC function. Furthermore, when the minimization is constrained in an affine subspace, we proposed two variants of the augmented Lagrangian method based on DME, which allow a nonsmooth concave component in the objective compared to the literature. ME13 CC Room 201A In Person: Multi-agent Learning in Games General Session Chair: Chamsi Hssaine, Cornell University, Los Angeles, CA, 90025- 5692, United States 1 - Finite-time Last-iterate Convergence for Multi-agent Learning in Games Tianyi Lin, University of California, Berkeley, Berkeley, CA, 94720- 2502, United States, Zhengyuan Zhou, Panayotis Mertikopoulos, Michael Jordan We consider multiagent learning via online gradient descent in a class of games called cocoercive games, a fairly broad class of games that admits many Nash equilibria and that properly includes unconstrained strongly monotone games. We characterize the finite-time lastiterate convergence rate for joint OGD learning; further, building on this result, we develop a fully adaptive OGD learning algorithm that does not require any knowledge of problem parameter and show, via a novel double-stopping time technique, that this adaptive algorithm achieves same finite-time last-iterate convergence rate as non-adaptive counterpart. We extend OGD learning to the noisy gradient feedback case and establish last-iterate convergence results first qualitative almost sure convergence, then quantitative finite-time convergence rates all under nondecreasing step- sizes. 2 - Algorithmic Collusion: Supra-competitive Prices via Independent Algorithms Mallesh Pai, University of Pennsylvania, Philadelphia, PA, 19104, United States We study outcomes when competing sellers use machine learning algorithms to set prices while ignoring competitors’ prices. We show that the long-run prices depend on the informational value of price experiments: if low, these prices are consistent with the static Nash equilibrium of the full information setting. If high, these prices are supra-competitive—-the joint-monopoly outcome is possible. We show a novel channel: competitors’ algorithms end up running correlated experiments. Therefore, sellers’ misspecified models overestimate own price sensitivity, resulting in higher prices. ME14 CC Room 201B In Person: Data-Driven Smart Transportation and Shared Mobility General Session Chair: Amirmahdi Tafreshian, University of Michigan, Ann Arbor, MI, 48109-2125, United States

ME11 CC Room 304C In Person: Market Design and Advances in Auction Design General Session

Chair: Sasa Pekec, Duke University, Durham, NC, United States 1 - Revenue-Sharing Allocation Strategies for Two-Sided Media Platforms: Pro-Rata versus User-Centric Sasa Pekec, Duke University, Durham, NC, United States We consider a two-sided streaming service platform which generates revenues by charging users a subscription fee for unlimited access to the content, and compensates content providers (artists) through a revenue-sharing allocation rule. Platform users are heterogeneous in both their overall consumption and the distribution of their consumption over different artists. In addition to determining the platform optimal revenue allocation rule, we study two primary revenue allocation rules used by market-leading music streaming platforms — pro-rata and user centric. In the pro-rata rule artists are paid proportionally to their share in the overall streaming volume, while in the user-centric rule each user’s subscription fee is divided proportionally among artists based on the consumption of that user. We characterize when these two allocation rules can sustain a set of artists on the platform and compare them from both the platform and artists perspectives. In particular, we show that, despite the cross-subsidization between low and high streaming volume users, pro-rata rule can be preferred by both the platform and the artists. We further show that the platform’s problem of selecting an optimal portfolio of artists is NP-complete. However, building on duality theory, we develop a polynomial time algorithm which outputs a set of artists so that the platform profit is within a single artist revenue from the optimal profit. Furthermore, for a platform that uses pro-rata or user-centric rules, by establishing connections to Knapsack problem, we develop a Polynomial Time Approximation Scheme (PTAS) for the optimal platform profit. 2 - A Smarter Market to End Global Warming and How to Implement it John F. Raffensperger, RAND Corp., Santa Monica, CA, 90405, United States Existing emissions trading systems inhibit emissions but not warming per se. They give poor price signals, and require laws for implementation. To solve these problems, I propose (1) a single global double-sided “smart market” with constraints on warming for the next 150 years. Each auction gives prices for this range. Further, I propose (2) this market be started by a coalition of firms, with jurisdictions joining as they are ready. Firms must require their suppliers to join. The proposed rules and implementation path appear to incentivize a rush to join. This design appears to be closer to “first best,” with a lower cost of mitigation, than any in the literature, while lowering the chance of catastrophe. A market simulation shows the error of static prices and “global warming potential” used in existing ETSs. An implementation simulation supports the participation analysis. ME12 CC Room 304D In Person: Forecasting/Accounting and Nonlinear Programming Contributed Session Chair: Kaizhao Sun, Georgia Institute of Technology, Atlanta, GA, 30318, United States 1 - New Product Life-cycle Forecasting With Temporal Hierarchies Oliver Schaer, University of Virginia, Charlottesville, VA, United States, Nikolaos Kourentzes, Douglas Thomas Extending life-cycle curves to capture seasonality can substantially increase model complexity and complicate the estimation of model parameters. To address these issues, we suggest using temporal hierarchies that use optimal suited time-series models at each aggregation level to extract model structure and subsequently combine it to increase predictive accuracy. For example, fitting a diffusion model at the quarterly level, with a long term focus, and a seasonal exponential smoothing model at the weekly level, with a short term focus. Combining these hierarchically results in a prediction that retains both aspects. We evaluate our approach on life-cycle data from a computer manufacturer. 2 - Time Series Forecasting for Equity Valuation Using Accounting Data Lukas Benjamin Heidbrink, Bielefeld University, Bielefeld, Germany Companies and investors need to forecast accounting data in a variety of cases, such as equity valuation and cash flow forecasting, or due to regulators requirements. I evaluate modeling alternatives in equity valuation and profitability forecasting given the nonstationary nature of many variables, focusing on the value relevance of accrual and cash basis accounting measures. Thus, I utilize classic multivariate time series models, error correction models and

Co-Chair: Hao Yan, Tempe, AZ, 85281-3673, United States 1 - A Generalized Fluid Model of Ride-Hailing Systems

90