The High Precise Optimization Algorithm and rational construct study of multi-layered feed-forward neural network

In this paper, a High Precise Optimization Algorithm for manipulating multi-layered feed-forward neural network is studied. Its basic principle is: defining neural network average error as objective function, weights and thresholds as design variables, through design variables rationally sorted, objective function is dynamically formed. Compared the new method with BP, the optimum step-length can be not only acquired per time computing and objective function is gradually decreased, but also oscillation phenomenon can be overcome by the new algorithm. A high precision computing program of multi-layered feed-forward neural network is programmed. Rational construct of multi-layered feed-forward neural network is analyzed by optimization. Through computing neural network of typical engineering question, its validity and application prospect is showed.