ABSTRACT

This Faculty Early Career Development Program (CAREER) grant advances the national health, prosperity and welfare by creating a robust decision-making framework for data-driven simulation. Due to its flexibility in capturing system randomness, simulation has been a popular tool to support decision-making problems that arise in manufacturing, healthcare, defense, finance, and other domains. However, simulation analysis is subject to ?model risk? of drawing an incorrect statistical inference due to discrepancy between the real system and the simulation model. Failure to account for such risk may lead to poor quality decisions made on the basis of these models. This research focuses on ?input model risk? that arises when the probability distribution functions driving randomness in a simulation model are estimated based on the available data. The project will study methods to quantify, reduce, and ensure robust decisions under input model risk. In particular, a new robust decision-making framework will be studied to balance a practical user input on acceptable suboptimality and robustness to the statistical error in the simulation model. The education mission of this grant is to train current and next-generation STEM workforce to make model risk a central focus of simulation analysis and equip them with computational tools to employ.

This research will enable input model risk quantification for complex simulated systems that are here-to-fore practically infeasible due to computationally complexity. A minimum-cost simulation experiment design will be obtained by applying the likelihood ratio method and solving a bilevel optimization problem. Moreover, a Gaussian process (GP) metamodel will be created to predict the simulation output mean as a function of both parametric and nonparametric input models. This GP metamodel will serve as a vehicle to design a comprehensive framework for all three steps of the robust simulation analysis life cycle: (1) risk quantification, (2) robust optimization, and (3) risk reduction. The concept of ?practically robust? optimality will be newly defined by accounting for the user-specified practical optimality gap of interest. This framework will reduce conservatism of existing methods while achieving the level of robustness the user desires. To find a practically robust optimum, an efficient simulation optimization algorithm, which sequentially allocates simulation effort guided by GP inference, will be created. Finally, an actionable guidance to reduce input model risk will be provided by optimizing the data collection plan to attain a stronger statistical performance guarantee for the practically robust optimum.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

Please report errors in award information by writing to: awardsearch@nsf.gov.