Identification of Unbalanced Warship Designs Using Multivariate Outlier Detection Procedures

摘要

The paper aims to identify unsuccessful warship designs using advanced outlier detection procedures. Therefore, we generate a data set covering information on general characteristics, offensive/ defensive power, and sensor capabilities of more than 800 different classes of (surface) warships from the cold war era to the current era (including planned future classes). On this high-dimensional data set we apply two approaches, one based on self-organizing maps (i.e., a special kind of artificial neural network, Liebscher et al. 2012), the other based on minimum spanning trees (Kirschstein et al. 2012). Both methods can identify outliers (i.e., "unbalanced" designs) and reveal cluster structures in the data. In a second step we compare the outliers' characteristics to a reference group (which is also determined in the first step) to decide whether they are "good" or "bad" ones.