New multivariate calibration methods, including model maintenance, often involve selection of multiple tuning parameters to determine the final calibration model. By using multiple tuning parameters, the respective penalties (constraints) provide additional regularizations in developing models. The difficult part in using multiple tuning parameters is selecting a good combination of tuning parameter values from the large number of possible combinations. Selection of single tuning parameter value for simple calibration models is often accomplished through some form of cross-validation and the resultant root mean square error of cross-validation (RMSECV) values are evaluated. However, even determining a “good” tuning parameter combination with this one measure model quality is sometimes difficult. Including additional measures of model quality allows selection of tuning parameter values to form better balanced models. However, selection of values for multiple tuning parameters based on many different model quality measures is challenging due to local anomalies. The best tuning parameter combination can be indistinct. Several fusion processes are presented to rank calibration models formed using multiple model quality measures. The fusion methods allow for simultaneous evaluation of any number of calibration model quality measures to select an appropriate combination of tuning parameter values. A calibration model updating situation is presented for characterizing the fusion processes. This situation consists of a near infrared pharmaceutical data set, including primary lab samples and secondary full batch samples, in conjunction with a Tikhonov regularization calibration maintenance approach. Another example presented consists of model development for a smooth regression vector using several different types of spectroscopic data sets.

Ranking multivariate calibration models formed from multiple tuning parameters: Model penalties

251st National Meeting (2016)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.