Application of BP Networks Improved by LM Algorithm in Pattern Classification

Shuo Ding,Xiaoheng Chang,Qinghui Wu,Youlin Yang
2014-01-01
Abstract:Standard BP algorithm has disadvantages such as slow convergence speed and local minimum in pattern classification.To overcome these problems,LM (Levenberg-Marquart)algorithm is used to improve standard BP algorithm.LM-BPNN and standard BPNN are established to classify two-dimensional vectors on a plane.And their generalization abilities are tested.The simulation results show that in the condition of the same number of training samples and small or medium-scaled networks,LM-BPNN has a much faster convergence speed than standard BPNN,as well as higher training accuracy and higher classification precision.And LM-BPNN also effectively overcomes the disadvantages of standard BPNN,including slow convergence speed and local minimum.
What problem does this paper attempt to address?