Statistical Engineering Division
Seminar Series

Tony Kearsley
Mathematical and Computational Sciences Division, NIST
Statistical Engineering Division, NIST
NIST North Room 152
Teusday, October 18, 2005, 10:30-11:30 AM

Abstract

Many scientific instruments produce numerical values that are contaminated with noise. These values often depend on machine parameter settings. Machine parameter settings are usually selected to optimize some aspect of performance, but tuning instrument parameters to achieve peak performance is complicated by the fact that machine output usually depends non-linearly on the machine parameters and is further complicated by output noise. Even so, in some applications, parameter settings that optimize instrument output are sought. These optimal parameter settings can greatly improve performance, but are often difficult and costly to compute. In this short talk, I will briefly survey some frequently employed numerical techniques (derivative-free) appropriate for these calibration problems. I will then introduce our current early-stage approach to compute noise-specific approximations to instrument parameter derivatives. Given statistical assumptions about instrument noise, derivatives can be estimated and calibration parameters optimized. A work-in-progress, a numerical example involving spectral data will be presented. Currently, the two major customers of this work are the Polymers Division at NIST and the Department of Justice.