A Fast Compositive Training Algorithm of Forward Neural Network

The thesis presents a fast compositive training algorithm of forward neural network, which integrates the advantages of traditional BP algorithm and single parameter dynamic searching algorithm (SPDS algorithm). It is well known that the BP algorithm, mostly used in many fields, has the disadvantages of slow convergent speed and the possibility of network paralysis. But SPDS algorithm overcomes these drawbacks of BP algorithm, and its training speed is much faster than BP algorithm and has better forecasting precision for the same samples. By numerical experimentations, it comes to the conclusion that the compositive training algorithm is good for training neural networks