My research focuses on the use of stochastic simulation for decision-making under uncertainty. This encompasses everything from the design and analysis of ranking-and-selection (R&S) procedures to the comparison of simulation-optimization (SO) algorithms to the development of new methods for simulation output analysis.
Plausible Inference Methods
In certain cases, SO problems possess structural properties that can be verified analytically, e.g., a bounded, Lipschitz-continuous objective function. I am studying ways to exploit such functional information to deliver statistical inference (e.g., confidence regions, screening), even at unsimulated systems.
Benchmarking SO Algorithms
Compared to deterministic optimization algorithms, SO algorithms present additional challenges when it comes to benchmarking. I am exploring ways to evaluate and compare the finite-time performance of SO algorithms. This effort has been driven through SimOpt – a growing testbed of SO problems and solvers – and includes the development of new experiment designs and analysis techniques for understanding the behavior of SO algorithms.
Ranking & Selection
Ranking-and-selection procedures select from one or more simulated alternatives from among a finite set and can provide either frequentist or Bayesian statistical guarantees. I am examining the interplay between the design of R&S procedures and the guarantees they deliver, including for SO problems with stochastic constraints and multiple objectives. A primary goal of this line of research is to discover new design principles for general-purpose R&S procedures.
Simulation Analytics
This line of research studies ways to integrate simulation modeling and statistical learning for input and output analysis, including uncertainty quantification, validation, prediction, and optimization. For example, I am studying how statistical learning techniques can be applied on time-dependent trace data that describe changes to the state of the real or simulated system over time for better prediction and control in digital twins.