INFORMS 2021 Program Book

INFORMS Anaheim 2021

TE28

of arm-heterogeneity in reducing both the sample- and communication- complexity of \texttt{Fed-SEL}. As a special case of our analysis, we show that for certain heterogeneous problem instances, \texttt{Fed-SEL} outputs the best-arm after just one round of communication. Our findings have the following key implication: unlike federated supervised learning where recent work has shown that statistical heterogeneity can lead to poor performance, one can provably reap the benefits of both local computation and heterogeneity for federated best-arm identification. As our final contribution, we develop variants of \texttt{Fed-SEL}, both for federated and peer-to-peer settings, that are robust to the presence of Byzantine clients, and hence suitable for deployment in harsh, adversarial environments. 2 - Model Aggregation in Federated Learning: Security and Efficiency Salman Avertimehr, University of Southern California, Los Angeles, CA, United States Model aggregation is a critical component of federated learning. In this talk, I will discuss several shortcomings in the state-of-the-art approaches for model aggregation, in terms of their security and efficiency, and propose new techniques to overcome those barriers. 3 - Exploiting Shared Representations for Personalized Federated Learning Liam Collins, University of Texas at Austin, Austin, TX, United States Neural networks have shown the ability to extract universal feature representations from data such as images and text that have been useful for a variety of learning tasks. However, the fruits of representation learning have yet to be fully-realized in federated learning (FL). In this talk, we propose FedRep: a novel federated learning framework and algorithm for learning a shared data representation across clients and unique local heads for each client. We prove that FedRep learns the ground-truth representation with per-user sample complexity that diminishes with the number of users in a linear setting, demonstrating that FedRep harnesses the benefits of collaboration in FL. Finally, we discuss experimental results showing that FedRep outperforms a variety of personalized FL methods on multiple data-heterogeneous FL benchmarks. TE28 CC Room 207B In Person: Decision-making under Multistage Uncertainty General Session Chair: Eojin Han, Southern Methodist University, Dallas, TX, 75205, United States 1 - Dynamic Capacity Management for Deferred Surgeries Kartikey Sharma, Zuse Institute Berlin, Berlin, 60208-0834, Germany, Eojin Han, Omid Nohadani, Kristian Singh The COVID-19 pandemic necessitated sweeping deferrals of elective surgeries. These deferrals led to deterioration of patients’ conditions due to delayed procedures and potential departures. Current policies are ad-hoc, i.e., either all surgeries are deferred or capacities are extended by pre-determined factors. We develop an optimization framework to optimally manage the expansion of surgical capacity under uncertain backlog. Given that the model contains nonlinear products of uncertainties, we provide tractable policies for realistic problems. Numerical experiments on claims data from a large fraction of US hernia patients demonstrate sizable improvements over competing methods. 2 - Optimal Transportation Mode Selection and Capacity Allocation under Uncertainty We consider the overseas supply chain of a manufacturing company with long lead-times and multiple transportation modes, where orders are placed using forecasted demand. The forecast error, which is the difference between forecasted and actual demand quantity, is considered an uncertain parameter. We also assume that the amount of excess inventory at the beginning of each period is uncertain. Order quantities for each transportation mode must be determined. We model this problem using a two-stage stochastic programming approach to minimize the overall expected order procurement, inventory holding, and backorder costs under demand and inventory uncertainty. Further, we use scenario decomposition based method, Progressive Hedging Algorithm, to solve the problem under consideration. We evaluate the performance of our solution algorithm via a numerical study. Avnish K. Malde, Graduate Research Assistant, Clemson University, Clemson, SC, 31405, United States, Tugce Isik

TE26 CC Room 206A In Person: Integer Programming and Combinatorial Optimization General Session Chair: David Bernal Neira, Carnegie Mellon University, Pittsburgh, PA, 15206-4367, United States 1 - Cross-dock Truck Scheduling with Workforce Constraint in Freight Transportation Ritesh Ojha, PhD Student, ISyE Georgia Tech, Georgia Institute of Technology, Atlanta, GA, 30318, United States, Alan Erera Freight transportation companies operate cross-docking terminals to enable freight transfer between trailers throughout their consolidation network. This research addresses the integrated truck and workforce scheduling problem at unloading doors in a cross-dock. The objective is to minimize the total violation of fixed deadlines at the loading doors. We develop an optimization-based methodology, equipped with an iterative time refinement exact algorithm, for creating a timed schedule of trailer unloading activities with worker assignments. The algorithm solves small integer programs in each iteration to yield the optimal solution. A computational study demonstrates the utility of the model and effectiveness of the algorithm to solve practical instances based on data, representative of a large cross-dock, provided by our research partner. 2 - A Chance-Constrained Two-Echelon Vehicle Routing Problem With Stochastic Demands Natasja Sluijk, Eindhoven University of Technology, Eindhoven, Netherlands, Alexandre Florio, Joris Kinable, Nico P. Dellaert, Tom Van Woensel Two-echelon distribution systems are often considered in city logistics to maintain economies of scale and satisfy the emission zone requirements in the cities. In this work, we formulate the two-echelon vehicle routing problem with stochastic demands as a chance-constrained stochastic optimization problem, where the total demand of the customers in each second-echelon route should fit within the vehicle capacity with high probability. We propose two efficient solution procedures based on column generation. To ensure that the chance constraints are met, we use statistical inference techniques. Additionally, we employ feasibility bounds on the stochastic demands to reduce the number of times we have to verify the chance constraints. The results show the value of the stochastic formulation in terms of improved solution cost and guaranteed feasibility of the routes. 3 - Easily Solvable Convex MINLP Derived from Generalized Disjunctive Programming using Cones David E. Bernal Neira, Carnegie Mellon University, Pittsburgh, PA, United States, Ignacio E. Grossmann We model problems where discrete choices enforce convex constraints via Generalized Disjunctive Programs (GDP). GDP can be solved as MINLP through reformulations, eg the Hull reformulation (HR). We derive a convex GDP representation by modeling constraints in disjunctions with conic sets. These problems’ reformulations can be efficiently tackled using solvers which take advantage of their conic structure. The HR of conic GDP is described exactly, leading to a tight formulation that avoids perspective function approximations. Our results, obtained from solving over 400 convex GDP arising from Process Systems Eng. and ML, show how the conic modeling of GDP leads to performance improvements. TE27 CC Room 206B In Person: Efficiency in Distributed ML Environments: Data Parallel, Model Parallel and Federated Learning Solutions General Session Chair: Liam Collins, Austin, TX, United States 1 - The Benefit of Heterogeneity in Collaborative Learning: Federated and Distributed Best-Arm Identification Hamed Hassani, University of Pennsylvania, Philadelphia, PA, 19104, United States We study a federated variant of the best-arm identification problem in stochastic multi-armed bandits: a set of clients, each of whom can sample only a subset of the arms, collaborate via a server to identify the best arm (i.e., the arm with the highest mean reward) with prescribed confidence. For this problem, we propose \texttt{Fed-SEL}, a simple communication-efficient algorithm that builds on successive elimination techniques and involves local sampling steps at the clients. To study the performance of \texttt{Fed-SEL}, we introduce a notion of arm- heterogeneity that captures the level of dissimilarity between distributions of arms corresponding to different clients. Interestingly, our analysis reveals the benefits

145