CAREER: Optimization and Sampling in Stochastic Simulation

点击翻译按钮获取中文摘要

The objective of this Faculty Early Career Development (CAREER) Program project is to develop new methods for optimizing and predicting performance of complex systems that are described by stochastic simulation models. Such systems arise in various areas such as finance, engineering design, systems biology, and manufacturing, and are often characterized by complexities, nonlinearities, and uncertainties in their dynamics. The major challenges in the optimization and prediction of the system performance are the expensive evaluation of system models, lack of structure in the performance measure, huge search space, and the need to address the balance between efficiency and accuracy. This research aims to make strides towards these challenges by developing new theory and methodologies. The proposed methods will be applied to modeling of a class of biological systems from experiment data and studying modes of behaviors of these systems, helping to reveal functional mechanisms and design principles of biological systems. This project also supports the PI's educational objective to integrate research with course development and in-classroom teaching, engage more females and underrepresented minorities in engineering, and expose high-school students and middle-school girls to the field of industrial engineering and operations research.If successful, this research will provide a set of new algorithms that possess both superior practical performance and rigorous convergence guarantees for the following two problems: (i) simulation optimization; and (ii) characterization of the response space of a system model. For simulation optimization, an algorithmic framework will be developed by integrating the central idea of model-based methods from deterministic nonlinear optimization with classical gradient-based search in a seamless way. To efficiently explore the response space, a new approach is proposed to sample from the response space and the parameter space iteratively, which takes advantage of the simple structure of the parameter space to circumvent the nonlinearity of the model while using the information on the response space to expedite the search in the parameter space.