Improved Algorithms, Based On The Parallel Structure Of Bp

The research about artificial neural network has gotten important development in recent years.It inspires many scientists' enthusiasm and interest in the field of computer science,brain neural science and artificial intelligence.Artificial neural network is simulation to the information-processing mechanism of the brain.It is expected that realize the brain function by simulating the structure and thinking of brain.Its theories have been applied in many fields.The multilayer perception,trained by the back propagation algorithm, is currently the most widely used neural network.It solves the question of hidden layer learning rule that utilize the method of error back propagation.The essence of back propagation networks is that make the change of weights become little by gradient descent method and finally attain the minimal error.Since it adopts the steepest descent method in nonlinear programming,it has the drawbacks that easy converge to a local minimum point of the error function,and it may converge very slowly.The systems of parallel computer now can be classified as SIMD and MIMD.At the beginning of the development of parallel computer,the type of SIMD systems played an important role.However,with the development of micro-processing chip technology,Single-processor performance has been greatly improved since 90's,and results in the development of the MIMD systems.SIMD is no longer popular.But both of the two have there advantage and default,so we can design new parallel structure by combining the advantage of the two,so as to solve some special problems.A brief overview about neural networks and a review of approximate algorithms for combinatorial optimizations are given;we also introduce the learning rules of neural networks very clearly.Then the paper introduces the basic principles of BP neural network and how to derive the algorithm,its strengths and drawback.A variety of effective improvement BP algorithms are given.Parallel theories are being widespread used in many fields.This paper use Parallel structure to optimize BP network structure.A new algorithm named PBBD(Parallel Based BP algorithm) is given,it uses many networks which are different in learning speed to train by parallel,aim to speed up learning process and get out of the flat area.The experiments show that the improved algorithm is feasible.