Intelligence collected on known terrorist organizations allows a monitoring agency to construct a database of observations that includes 1) an observed level of chemical, biological, radiological or nuclear (CBRN) weapons development, 2) a set of features that describes the terrorist organization, and 3) a label identifying both the terrorist organization and the time in which the data was collected. An exploration of the multivariate data is helpful for determining which subset of features can be used to best allow discrimination between levels of CBRN weapons development. In feature weighting such as this, the goal is to remove irrelevant, noisy, or redundant features and assign a non-negative importance value to each remaining feature.

In this analysis, we discuss a random search algorithm that seeks to minimize classification risk on our CBRN training set using a nonparametric kernel-based method by modifying the weights of a Minkowski distance metric. The results are then compared to similar analyses performed with two competing approaches to see if the additional computational cost associated with the proposed method is worthwhile. Where possible, cross-validation is conducted on the data to ensure that the resulting classifier is not overly dependent on the training set.