The Study On Learning Algorithm Based On Sensitivity Theory For Madaline Neural Networks

The learning algorithm based on the sensitivity-theory for MADALINE neural networks, which is one of multi-layer perceptions neural networks, is a supervised learning algorithm. Until now, there is no satisfactory training algorithm for the discrete multi-layer perception neural networks.The principle of minimal disturbance in the paper is looked on as a fundamental idea to designing algorithm. In the paper, a means to perform the principle successfully found and a algorithm based on the sensitivity theory of MADALINE for Madaline neural networks has been worked out. In addition, "local recycle" phenomenon, as a failure mode of the new algorithm, is analyzed in the paper. The way to break the situation is found and the algorithm called as "disturbance algorithm" is figured out. Finally, the new algorithm shows better performance in leaning and generalization through the simulation experiments in comparison with MRII and the "monk's problem".Other good performance of the algorithm is also analyzed and revealed in the paper,such as leaning capability having strong sensitivity to the number of hidden-layer neurons and generalization capability having stability to networks' structure.To large extent, the new algorithm has achieved the generaliziton of the training rules of perceptions in the discrete multi-layer perception neural networks sucessfully. It has solved the difficulties in the neural networks fields well.