Abstract:The reason is analyzed why the convergence of standard BP algorithm is slow.Detail research is carried out on Levenberg-Marquardt BP(LMBP),which is the fastest improved version of BP.It is found that the calculation of inverse matrix of [JTJ+μkI] involved causes primarily the problem.LU decomposition is employed to avoid the time-consuming calculation of the inverse matrix,so as to decrease largely the calculation amount of LMBP.In addition,the steps for the calculation of the augmented Marquardt sensitivity matrix are simplified,which also decreases the calculation amount to a certain extent.The improved LMBP algorithm is implemented by authors with Microsoft Visual C++ 6.It is found that above two improvements caused LMBP to converge rapidly.Many trials are conducted on Traingdx from Matlab,which is based on the steepest descent algorithm,Trainlm from Matlab,which is improved on LMBP in Matlab,LMBP,and ILMBP,which is improved on LMBP by authors.The results show that the mean convergence speed of ILMBP is 23 times quicker than that of LMBP,and 9 times than that of Trainlm.

Comparative Study on Fast Learning Algorithms of BP Networks

The Performance Comparison of Several Kinds of Modified BP Algorithm

Comparative of Pattern Classification of BP Neural Networks Improved by Numerical Optimization Approach

An Improved BP Algorithm Based on Parallel

New and Rapidly Converging BP Neural Network Algorithm

Study on Fast Intelligent Learning Algorithm of Neural Network(i)

An Improved Back-propagation NN Algorithm And Its Application

On Improved Algorithm of LMBP Neural Networks

Study of the Optimized Algorithms of BP Neural Network

An Algorithm for Fast Convergence of Back Propagation by Enlarging Error

Research on Improvement of Convergence Speed of LMBP Algorithm in Neural Network

Comparison of BP Algorithms in Matlab NN Toolbox

Implementation and Comparison of Improving BP Neural Network Based on MATLAB

BP Neural Network Algorithm Improvements and Convergence Analysis

A Fast Learning Algorithm of Neural Networks by Changing Error Functions

The Research on the Relation of Self-Learning Ratio and the Convergence Speed in BP Networks

QLMBP: A Quick BP Neural Network Algorithm

Fast Learning Algorithms for Feedforward Neural Networks

The Study of Approximate Optimal Step Size of B-P Neural Network

Application of Improved BP Network Algorithm in Image Recognition

Comparative study on function approximation performances of GRNN and BPNN