ИСТИНА |
Войти в систему Регистрация |
|
ФНКЦ РР |
||
Multiple classification problems are usually hard to solve. With increasing number of classes, classification algorithms rapidly degrade, both by error rate and by computational cost. Multi-layer perceptron (MLP) type neural networks (NN) solving such problems are subject to the same effect: greater number of classes requires increasing the number of neurons in the hidden layer(s) (HL) to build a more complex separation surface, making the NN more prone to overtraining. An alternative way is to build a hierarchical classifier system, uniting the target classes into several groups and solving the recognition problem within each group recursively at the lower-lying levels of hierarchy. The authors of this study are developing an algorithm for adaptive construction of such a hierarchical NN classifier (HNNC). Each node of the constructed hierarchical tree is an MLP with a single HL consisting of only a few neurons. Such an MLP is knowingly unable to recognize all the required classes in a multiple classification problem. So after it has been trained for a specified number of epochs, it is applied to all the samples of the training set, and its output is analyzed. If the majority of samples from some class k ”vote” at the MLP output as belonging to another class m, the desired output for class k is changed to be the same as for class m. In this way, classes are united into groups, and as this modification is performed in the way favorable for the MLP, we obtain a system with positive feedback, rapidly converging to a trained NN with a high rate of recognition into several adaptively formed groups of classes. Afterwards, each group is subject to further recognition in an iterative way, thus providing adaptive construction of a HNNC. In this study, a new modification has been introduced into the algorithm. Now the target classes may not only unite, but under some conditions any target class may split into two new classes, possibly simplifying class separation borders, increasing efficiency and stability of the algorithm. The presentation displays the results of the algorithm without and with the new modification on several well-known benchmark problems.