The research objective of this award is to develop a robust and tractable methodology for assessing the impact of input model error, with minimal assumption placed on the parametric form of the input model. The discrepancy between model assumption and reality, or input model error, constitutes an important concern in many stochastic estimation problems, since it can affect the accuracy of outputs in various magnitudes. The key ingredient is a class of sensitivity estimators that are taken with respect to nonparametric statistical distances, derived via a new line of infinitesimal analysis on optimization problems over probability space. These sensitivity estimators are robust in that they automatically capture model discrepancy along the worst-case directions in the model space, and they are computationally tractable via efficient Monte Carlo methods. On the theoretical side, they will help provide fundamental understanding of model risk in stochastic computation, and on the practical side, they can be widely used to perform stress tests on the reliability of model assumptions.
If successful, the results of this research will provide a robust and implementable method to assess the effect of model risk in computations that arise in many applications. Industries such as manufacturing, communication, call centers, and financial risk management all need to carry out calculations of different performance measures on a regular basis. The methodologies that come out from the research will provide sensitivity analysis tools to measure the impact and risk of inaccurate model assumptions in these calculations. The results of this research will also be used to develop new courses in undergraduate and graduate levels, and to establish mentorship of students from under-represented groups. The implementation will be widely disseminated through open software for both academic and industrial use.
