Title: Sequential Kriging Optimization Methods for Stochastic and Multi-Fidelity Evaluations

Speaker: Deng (Dan) Huang, Ph.D., Post Doctorial Candidate, Department of Industrial & Systems Engineering, Ohio State University
 
Date/Time: Thursday, May 12, 2005, 10:00-11:00 am (MDT)

Location: Building 980, Room 24 (Sandia NM), Building 915, Room S145 (Sandia CA)

Brief Abstract: Sequential Kriging Optimization (SKO) is a method developed in recent years for solving non-linear expensive black-box problems in areas such as large-scale circuit board design and manufacturing process improvement.  The method is based on a stochastic process meta-model (kriging) that provides a global prediction and a measure of prediction uncertainty.  The subsequent evaluations are selected by maximizing the so-called expected improvement function.

We extended SKO to address stochastic systems by modifying the expected improvement function to achieve the desired balance between the need for global and local searches.  In studies using test functions, SKO compared favorably with alternative approaches in terms of consistency in finding global optima and efficiency as measured by number of evaluations.

We proposed another extension of the SKO method, named Multiple Fidelity Sequential Kriging Optimization (MFSKO), where surrogate experimental systems are exploited to reduce the total evaluation cost.  In this method, data on all experimental systems are integrated to build a kriging meta-model.  The location and fidelity level of the evaluations are selected by maximizing an expected improvement function that is related to the evaluation costs.  The proposed method was applied to 1) a metal-forming process design problem via Finite Element simulations, and 2) discrete event simulations of hospital emergency departments.

CSRI POC: Laura Painton Swiler, (505) 844-8093