A Program which Constructs a Multistage Classification

In recognition problems in which learning takes place for a large number of classes it is frequently not possible to separate them in terms of stable features in one stage. The author proposes an algorithm which divides according to stable features in such cases (without recourse to probabilistic methods). The algorithm divides in several stages. Stability of the features is required only on the subset of classes which remains undivided at the preceding stage of the classification tree and which is to be divided in the given stage. Certain examples are given and experiments on comparing multistage and one-stage classifications are evaluated.