A Comparison of Classification by Two Versions of the Method of Generalized Portraits

The modified method of generalized portraits is compared with the original method. Particular attention is paid to the existence of solutions for the two methods and the shape of the resultin decision surfaces in the training space. One finds that solutions always exist in the original method, regarless of the placement of the training vectors. The indecision region is a hyperplane, which may, in extreme cases, cover the entire test space. In the modified method a solution exists only if the members of one class can be encapsulated in a hyperellipse to the exclusion of the other class. The effect of stepcoding is investigated. The results are displayed by means of two-dimensional examples.