The training strategy for creating decision tree

This paper describes a decision tree based on OCR, consisting of three parts: the root level, the Kohonen level and the MLP level. The proposed training strategy aims at creating an object-oriented decision tree classifier. The entire classifier is composed of a few separate sub-trees, each of which is a sub-classifier for some special pattern category and effectuated with a Kohonen Self-Organizing Feature Map (SOFM) and the Multiple Layer Perceptron (MLP). The growing and pruning training algorithms of a neural tree are proposed to train four pattern categories samples, corresponding to digits, uppercase letters, lowercase letters, and the mixture of digits and letters, respectively. After building the tree, there exists a number of leaf nodes that contain more than one character class, they must be further broken down into individual clusters since a definite recognition is demanded for a given input pattern. For this purpose, the MLP level training strategy is incorporated. The experimental result shows that the training algorithm strategy is feasible in the decision tree.