INFORMS 2021 Program Book

INFORMS Anaheim 2021

TB32

2 - Hogwild! Over Distributed Local Data Sets with Linearly Increasing Mini-batch Sizes Lam M. Nguyen, IBM Thomas J. Watson Research Center, Ossining, NY, 10562-6037, United States We consider big data analysis where training data is distributed among local data sets in a heterogeneous way and we wish to move SGD computations to local compute nodes where local data resides. The results of these local SGD computations are aggregated by a central “aggregator” which mimics Hogwild!. We show how local compute nodes can start choosing small mini-batch sizes which increase to larger ones in order to reduce communication cost. We improve state-of-the-art literature and show O(K^{0.5}) communication rounds for heterogeneous data for strongly convex problems, where K is the total number of gradient computations across all local compute nodes. For our scheme, we prove a tight and novel non-trivial convergence analysis for strongly convex problems for heterogeneous data which does not use the bounded gradient assumption as seen in many existing publications. TB28 CC Room 207B In Person: Robust Optimization for Machine Learning General Session Chair: Phebe Vayanos, University of Southern California, Los Angeles, CA, 90089, United States 1 - New Algorithms and Complexity Analysis for Distributionally Robust Multistage Convex Optimization Shixuan Zhang, Georgia Institute of Technology, Atlanta, GA, United States, Andy Sun We present a novel algorithmic study and complexity analysis of stagewise independent distributionally robust multistage convex optimization (DR-MCO). A new class of dual dynamic programming (DDP) algorithms for solving DR-MCO is proposed. The new algorithms generalize existing DDP-type algorithms by introducing the technique of regularization that enables the algorithms to handle fast growth of Lipschitz constants, and problems without relatively complete recourse. We then provide a thorough complexity analysis of the new algorithms, proving both upper complexity bounds and a matching lower bound. Numerical examples are given to show the effectiveness of the proposed algorithms. 2 - Competitive Pricing in Airline Revenue Management with Multi-agent Reinforcement Learning Shulu Chen, George Washington University, Washington D.C, DC, United States, Syed A. Shihab, Peng Wei We explore the application of multi-agent reinforcement learning (MARL) in airline revenue management (RM) in a competitive market. By extending our group’s research of DRL in airline pricing, this research uses multi-agent reinforcement learning to observe certain competitors’ information and make learning-based decisions under uncertainty. We model two competitive agents in MARL frame as two competitive airline companies. At each time step, two airline companies will choose their price points respectively, according to their own remaining seats, time to departure, and competitor’s price. We expect that the competitive MARL agents would have a better performance than the previous methods because of additional information and learning-based adaptive decision making. 3 - Optimal Robust Classification Trees Nathan Justin, University of Southern California, Los Angeles, CA, United States, Andres Gomez, Phebe Vayanos, Sina Aghaei In many high-stakes domains, the data used to drive machine learning algorithms is noisy (due to e.g., the sensitive nature of the data being collected, limited resources available to validate the data, etc). In this paper, motivated by the need for interpretability and robustness in these domains, we present an efficient MIP- based method for learning optimal classification trees that are robust to perturbations in the data features. We evaluate the performance of our approach on numerous publicly available datasets and show significant improvements over the state-of-the-art. TB30 CC Room 207D In Person: Fundamentals of Start-Ups Panel Session Chair: Arman Sabbaghi, Purdue University, West Lafayette, IN, 47907- 2067, United States 1 - Fundamentals of Start-Ups Arman Sabbaghi, Purdue University, West Lafayette, IN, 47907- 2067, United States Researchers and faculty members are increasingly engaging start-up opportunities that involve new and challenging issues in entrepreneurship, venture capital

funding, and intellectual property rights. The panelists in this session will discuss their own experiences in pursuing start-up opportunities and creating companies. 2 - Panelist Kamran Paynabar, ISyE Georgia Tech, Georgia Tech, H. Milton Stewart School Of Isye, Atlanta, GA, 30332-0205, United States TB31 CC Room 208A In Person: Data Markets General Session Chair: Azarakhsh Malekian, University of Toronto, Toronto, Canada 1 - Synthetic Interventions Anish Agarwal Consider a setting where there are N heterogeneous units (e.g., individuals, sub- populations) and D interventions (e.g., socio-economic policies). Our goal is to learn the potential outcome associated with every intervention on every unit (i.e., N x D causal parameters). Towards this, we present a causal framework, synthetic interventions (SI), to infer these N x D causal parameters while only observing each of the N units under at most two interventions, independent of D. This can be significant as the number of interventions, i.e, level of personalization, grows. Importantly, our estimator also allows for latent confounders that determine how interventions are assigned. Theoretically, under a novel tensor factor model across units, measurements, and interventions, we formally establish an identification result for each of these N x D causal parameters, and establish finite-sample consistency and asymptotic normality of our estimator. Empirically, we validate our framework through both experimental and observational case studies; namely, a large-scale A/B test performed on an e-commerce platform, a phase 3 clinical trial data from a pharmaceutical company, and an evaluation of mobility- restricting policies on COVID-19. We believe this has important implications for program evaluation and the design of data-efficient RCTs with heterogeneous units and multiple interventions. 2 - A Model of Behavioral Manipulation Ali Makhdoumi, Duke University, Durham, NC, 27708-9972, United States The default position among AI researchers is that the vast amounts of data collected by online platforms ultimately benefit users by providing them with more informative advertising, better-targeted products, and more personalized services. This talk raises and explores the possibility that this informational advantage may also enable platforms to engage in behavioral manipulation, which we define as the ability of platforms to modify the behavior of users in a way that is beneficial for the platform and costly for users. TB32 CC Room 208B In Person: mpi-sppy: Asynchronous Optimization under Uncertainty General Session Chair: David Woodruff, University of California Davis, CA, United States 1 - Asynchronous Projective Hedging: Introduction, Implementation, and Large-scale Computational Experiments Using mpi-sppy Jean-Paul Watson, Senior Research Scientist, Lawrence Livermore National Laboratory, Livermore, CA, United States, David L. Woodruff, Jonathan Eckstein, Bernard Knueven We describe a scenario-based decomposition algorithm Asynchronous Projective Hedging, or APH for multistage stochastic programming that resembles the progressive hedging method of Rockafeller and Wets, but is capable of asynchronous parallel operation without sacrificing theoretical convergence in the convex case. Perhaps more importantly, each iteration of the decomposition method may process only a subset of the possible scenarios. We discuss the implementation of APH in the mpi-sppy parallel library for stochastic programming, and detail large-scale computational experiments highlighting both the effectiveness of APH and the scalability (to tens of thousands of ranks) of mpi- sppy. 2 - Bounds and Confidence Intervals in mpi-sppy David L. Woodruff, University of California, Davis, Davis, CA, 95616, United States, Xiaotie Chen, Bernard Knueven, Jean-paul Watson mpi-sppy (https://github.com/Pyomo/mpi-sppy) is a software package to allow for optimization of Pyomo optimization models uncertainty. In thistalk we will overview design and performance considerations related to bounds and confidence intervals. Particular attention will be paid to issuesassociated with problems that have more than two stages and scenarios that do not exhibit stage- wise independence.

107