Recognizing patterns of errors in scientific data

Investigates the application of machine learning techniques to the task of identifying and clustering errors in scientific data. A numerical model called WAM (WAve Model) is used by the Navy to predict significant wave height (SWH) based on wind speed. Comparisons of the WAM predictions and SWH calculated from altimetry measurements have indicated that the WAM predictions are inaccurate in some situations. We have conducted experiments using two well-known clustering packages, COBWEB and AutoClass, and have developed a new method for analyzing the output clusters from experiments that use real values for attributes. Our results have shown that this data does not contain a strong class structure, but that some groups of attributes are better predictors of WAM prediction errors than others