A Hybrid Algorithm For The BP Network Training

The article has mainly discussed some algorithms about training BP Neutral Net-work.In the first paragraph the article gives the biological model of the human cranialnerve system.Expounds the theory of the neurons, and show the main character of thenenrons.People created the Artificial Neural Network according to the emulation of thehuman cranial nerve system. Artificial Neural Network is an information processingsystem which imitates the process of obtaining the information according to the trans-mission and reflection in human cranial nerve system.It's a theorizated mathematicalmodelofhumanbrainwithitsactivity.Itsfunctionunitisneurons.Theinputinformationfrom other neurons is weighted fitted and output the information by using the activityfunction in the neurons, then deliver the output information to the next.The ArtificialNeural Network has succeed the excellent performance which human cranial nerve sys-tem has.It has the characters of independence studing, currency, robustness and nonlin-earity. Some normal used Artificial Neural Network Model are Feed-Forward NeuralNetwork, Back-Forward Neural Network, Mixed Neural Network. The research on theFeed-Forward Neural Network never interrupt because it is widely used.BP Networkplays a important part in Feed-Forward Neural Network, so the improvement in BPNetwork algorithm has rapidly develop the Artificial Neural Network.nIn the second paragraph we show the theory of BP network and its algorithm.nBP Network is also called Back Propagation Network.It is one of most widely usedArtificial Neural Network Model.It reflects the relation between the data of input andoutput by studying and training.n The primary algorithm of training BP Neural Net-work is the traditional BP algorithm based on gradient descent algorithm.It adjust theweights by propagating the error back forward.nThen make the error function reach theminimum.Because the traditional BP algorithm has some weakness just as long timeand slow speed training, and the situation which the train can easily entrap in the partof minimal value, the article introduces some algorithm to improved the BP Network training.nIn the third paragraph we introduce some algorithm to train BP network. Variablelearning rate method change the learning rate to accelerate the convergence to the min-imal value.nAdd momentum method is add a part of the last weight correction to everycorrection on weight in the training network, thus accelerate the rate of convergencein studying, and avoid the network to entrap in the part of minimal value efficient-ly.Resilient BP algorithm correct the weight of the network using the partial deriva-tive of the error function.nReduce the computation efficiently, accelerate the rate ofconvergence of the network.FR conjugate gradient method and PRP conjugate gradientmethodhastheglobalconvergenceundertheexactlylinearsearch,avoidingthenetworkavoidthenetworktoentrapinthepartofminimalvalue.nLevenberg-Marquardtmethodis the fitting gradient descent method and Newton method.Compared with traditionalBP algorithm and other algorithm Levenberg-Marquardt method has less steps in itera-tion, faster convergence and higher precision in solving the nonlinearity problem.nThegenetic algorithm is a random search recursive computational procedure based on thesimulation of principles of evolution of living organism in nature.n It select the bestindividual according to the encoding, selection, trips variation to the original problem,It has a good currency and comprehensive optimization ability.Combined the abilityof comprehensive optimization of the genetic algorithm with the character of fast ofLevenberg-Marquardt method convergence, the article shows a mixed method.nAt last paragraph we according to the numerical experiment we compare thesemethods for BP network training.Artical improves the mixed method is a efficient wayto solve some nonlinearity problem.