Abstract: In assigning features to nodes of a tree structured classifier and determining terminal nodes in response to a training set of objects each being determined by a plurality of features, a selected characteristic, such as a cost function based on the minimum description length, of the plurality of features unused at prior nodes is determined along the path from the root to the present node, a feature is then assigned to the node having a preferred value for the selected characteristic relative to the other features. Child nodes of the node are created in response to the assigned feature with features assigned thereto in a similar way. The preferred values for the selected characteristics of the assigned features for the child nodes of the node are combined and compared with the preferred value of their father node. Classification of the father node as a terminal is based upon such a comparison.