报告题目：Algorithm Improvements for IPOPT

报告人：Wei Wan
Ph.D. candidate, Chemical Engineering Dept. Carnegie Mellon University, USA

报告时间：2015年5月8日15:00
报告地点: 教九429

报告摘要：
Nonlinear programming (NLP) has become an essential tool for process systems engineering. Interior point NLP solver IPOPT is widely used because of its high performance. However, it still has several drawbacks. Dependent constraints are common in chemical process models but violate the regularity conditions, which may cause the interior point method to fail. In my project, we aim to improve IPOPT's performance for dependent and ill-conditioned problems through algorithmic modifications and data-based options recommendation. Three structured regularization methods have been developed to handle the dependent equality constraints and results show an average reduction of more than 50% of the iterations. In addition, we are developing a new restoration design to better handle ill-conditioned problems. An IPOPT online submission system is developed for the data-based options tuning.

报告人简介：
Wei Wan is a PhD student working with Larry Biegler at CMU. She completed her B.S degree at CSE department in Zhejiang University. Her research focuses on nonlinear programming algorithms and KKT matrix regularization method. Wan is the owner of Chu Kochen Honors Scholarship and Top 10 College Students in Zhejiang University.

报告题目：Trust region methods for reduced model-based optimization

报告人：John P. Eason
Research Assistant, Chemical Engineering Dept. Carnegie Mellon University, USA

报告时间：2015年5月8日16:00
报告地点: 教九429

报告摘要：
A common approach to incorporate complex simulations and black box functions in optimization models is to fit an algebraic surrogate model to these complex functions. This allows a fully algebraic system description, suitable for modern nonlinear optimization solvers. However, small errors in this reduced model can result in a large deviation from the true optimum. In this work we investigate trust region algorithms capable of adaptively controlling this error during the optimization process to provably converge to the true optimum. In particular, we focus on the case where the black box function appears in the constraints. A filter-based method is developed to handle feasibility and is shown to have superior performance to a previous penalty function method.

报告人简介：
John Eason is a PhD student working with Larry Biegler at CMU. Before arriving at CMU, he completed his undergraduate degree in Chemical Engineering at the University of Tulsa, conducting research there in the field of PSE with Selen Cremaschi. His research focuses on process optimization with a focus on derivative free/simulation-based optimization methodology.