Introduction to the Special Issue for INFORMS Simulation Society (I-Sim) Workshop, 2021

In summer 2021, the INFORMS Simulation Society Workshop was held virtually and was jointly hosted by the Department of Industrial and Manufacturing Engineering (IME) and the Supply Chain and Information Systems (SCIS) Department at the Pennsylvania State University. The workshop was chaired by Dr. Uday V. Shanbhag (IME) and co-chaired by Dr. Russell R. Barton (SCIS), who organized the workshop together with the remainder of the following committee members: Drs. Saurabh Bansal (Pennsylvania State University), Güzin Bayraksan (Ohio State University), Henry Lam (Columbia University), Eunhye Song (Pennsylvania State University), Farzad Yousefian (Rutgers University), and Enlu Zhou (Georgia Tech). The workshop was centered around the theme of “From Data to Decision-making: Contending with Uncertainty and Non-stationarity in Simulation Theory” and was held between June 21 and June 23, 2021. The workshop was highlighted by three plenary talks and 24 talks by invited speakers. In addition, the workshop had a day-long simulation summer school covering a range of fundamental topics in simulation theory and optimization. Seven of the speakers built on their presentations to submit papers for a special issue in ACM TOMACS.

This special issue was handled by an editorial team comprising Drs. Russell R. Barton, Marvin K. Nakayama, Uday V. Shanbhag, and Eunhye Song. The articles in this issue span a diverse and timely set of topics, including the application of neural networks in the context of simulation, developing novel confidence interval procedures, the analysis of contextual ranking and selection, constrained Bayesian optimization, misspecified simulation optimization frameworks, and the role of stochastic approximation (SA) in analyzing the efficiency of Nash equilibria in uncertain environments.

In “NIM: Generative Neural Networks for Automated Modeling and Generation of Simulation Inputs,” Cen and Haas present Neural Input Modeling (NIM), a generative neural network framework that leverages data-rich environments to model simulation input processes and then generate samples from the model. This architecture combines a variational autoencoder architecture that learns the distribution of input data with Long Short-Term Memory components capable of capturing statistical dependencies across time. The authors proceed to show that modifications of this architecture allow for both exploiting a range of distributional properties and handling a breadth of processes including multivariate and categorical-valued processes.

In “Learning to Simulate Sequentially Generated Data via Neural Networks and Wasserstein Training,” Zhu, Liu, and Zheng propose a neural network-assisted sequentially structured simulator to model, estimate, and simulate a broad class of sequentially generated data. Given representative real data, the neural network parameters are estimated via a Wasserstein training process that does not necessitate restrictive distributional assumptions. In fact, this simulation can capture a range of elementary randomness and generate possibly heavy-tailed distributions. The authors proceed to derive consistency and rate guarantees for the estimation procedure.

In “Overlapping Batch Confidence Intervals on Statistical Functionals Constructed from Time Series: Application to Quantiles, Optimization, and Estimation,” Su, Pasupathy, Yeh, and Glynn present a confidence interval procedure (CIP) for statistical functionals built from data drawn from a stationary time series. By leveraging distribution-free analogues of the chi-squared and Student’s \(t\) random variables in the context of statistical functionals, this procedure is applicable in settings that include quantile estimation, gradient estimation, and estimation of rates of arrival processes. Akin to subsampling, the authors employ overlapping batches of time series data to estimate the underlying variance parameter. The correctness of the proposed CIP is proven, and extensive numerics suggest that when a suitable functional central limit theorem is in place, the resulting confidence intervals are often of significantly higher quality than those obtained from more generic methods like subsampling or the bootstrap.

In “Bayesian Optimisation for Constrained Problems,” Ungredda and Branke consider a Bayesian optimization framework for addressing practically occurring optimization problems based on complex black-box functions. The authors propose a generalization of the knowledge gradient acquisition function to the constrained regime. Convergence guarantees are provided in the infinite budget setting and empirical studies suggest the superior behavior of the scheme with competing Bayesian counterparts.

In “Contextual Ranking and Selection with Gaussian Processes and OCBA,” Cakmak, Wang, Gao, and Zhou consider the contextual ranking and selection problem in a finite-alternative-finite-context regime in which the goal is to find the best alternative in each context. By employing a distinct Gaussian process (GP) to model the reward for each alternative, the authors propose a sequential sampling policy referred to as GP-OCBA for which consistency is proven. Numerical studies suggest that the proposed scheme is competitive in terms of sampling efficiency with a lower computational overhead.

In “Stochastic Approximation for Estimating the Price of Stability in Stochastic Nash Games,” Jalilzadeh, Yousefian, and Ebrahimi approximate the price of stability (PoS) in stochastic Nash equilibria problems via stochastic approximation (SA) schemes. The PoS is a commonly employed metric for determining the efficiency of a Nash equilibrium and assumes relevance in the design of networked multi-agent systems. The authors propose a framework for estimating the PoS in uncertain settings by minimizing a convex function over the solution set of a monotone stochastic variational inequality problem. An efficient block-structured SA framework is presented with rate and complexity guarantees.

In “Stochastic Approximation for Multi-period Simulation Optimization with Streaming Input Data,” He, Shanbhag, and Song consider a simulation-optimization (SO) problem with continuous variables with the intent of optimizing an expected performance measure of a real-world system. The parameters of this system are unknown a priori and are re-estimated with higher precision at every period by utilizing periodically available streaming data. In addition, a decision-maker updates her decision at each period by solving an SO problem with increasingly accurate system parameters. The resulting multi-period SO problem is addressed via a multi-period SA framework that constructs a sequence of solutions by either restarting the stepsize sequence in SA (called ReSA) or warm starting the stepsize sequence (WaSA). Under suitable convexity and regularity conditions, both schemes achieve the best possible rate in expected sub-optimality when either an unbiased or a simultaneous perturbation gradient estimator is employed. Notably, the WaSA framework displays far more modest growth in computational effort as the number of periods increases.