Study On Procedure Neural Networks' Model And Learning Algorithms

Procedure Neural Networks (PNN) is a novel artificial neuron Networks model. It is based on information processing pattern of biological neural system and application background of practical matters. Both Input and output of the networks are cither procedures or functions. Procedure-type's inputs to networks relax synchronization instantaneous limit on inputs in the traditional neural network models. So, it could be seen that the structures research, function approximation properties and learning algorithms of procedure neural network models is quite significant.In the theory and model section of this paper, the concepts of procedure neurons and procedure neural networks are presented. Two sorts of models of procedure neural networks are shown, they are expanded on base functions model and projection compounding model. The equivalence of two models is proved. Three sorts of feed forward procedure neural network structure is expatiated, they are radial base PNN, self-organizing PNN and parallel connection PNN. The feed back procedure neural network structure is researched. In the course of research of PNN algorithm, Applying the concept of vertical base function, integral operation is convert to additive operation and complicated aggregation in time field is avoided. Two sorts of general base function property is analyzed, they are trigonometric function and walsh function. Arming at projection compounding model, a learning algorithm base on weight function expanded on certain base function is proposed. The algorithm based on discrete walsh conversion is researched, and the validity of algorithm is proved. The elaborate algorithm of feedback PNN is presented. The simulation experience proved availability of the model on accelerating convergence. In the last section of paper, arming at practical matters in oil field exploitation, the application of PNN in the complex water flooded layer identification is shown.