Improved LMBP Neural Network Algorithm And Its Application

As the society develops gradually, the large-scale accumulation of knowledge and information have already made people overwhelmed. Especially the advent of Internet tidal wave since the21st century induces an explosive growth of information people take in. The study of intelligently analysis, processing and prediction of massive data even interdisciplinary field are paid attention to and valued by people more and more, among which an ascendant study topic is neural network. It can be used to solove practical problems in engineering and science effectively through building the smart model for widely application fields and is a topic of highly research value and application prospect. In recent decades, neural network has been applied to medical diagnosis more and more and has contributed to medical development larger and larger. People apply the neural network to classification diagnosis of heart disease, analysis of cancer cells, times optimization of organ transplantation, analysis of respiratory tract infection rates, analysis of electrocardiography and electroencephalogram medical cases and so on.BP(Back Propagation) neural network is a common used and popular network model. It is proposed based on the Widrow-Hoff learning algorithm and ADALINE(ADAptive LInear NEuron) network which have the limit that they can only solve linearly separable problems. BP network can be used to train multi-layer network struct and can adjust the network parameters by self-learning, so it’s capable of dealing with the non-linear data analysis and prediction and the performance optimation problems. Due to its powerful non-linear modelling ability, BP network spread as soon as it was proposed. However, BP network algorithm has2main deficiencies:(1) the low convergence rate of network causes network training spending long time for BP network uses the first order derivative of the target function, which leads to a linear convergence rate of network, and when the practical problems are large-scale it may take hours even days to work out;(2) network may converge locally and cann’t figure out the optimal solution for BP network uses the local optimization algorithm and besides the surface of target function of multi-layer network struct is much complicated in general. People have conducted kinds of optimation and improvements on it according to the different actual requirements and many refinement algorithms make sense, but this problem is still the sticking points of research by far.Directing against the deficiencies of problems above, this dissertation adopts a derived algorithm, LMBP(Levenberg-Marquardt Back Propagation) algorithm, of BP neural network and we improve the convergence rate and the accuracy of network parameters calculation furtherly on the basis of LMBP algorithm. A recursive Cholesky decomposition algorithm based on a partitioned matrix is proposed according to the theoretical analysis of Hessian matrix in LMBP algorithm and the idea of divide and conquer. This algorithm speeds up the convergence process of network in2aspects:one is that use Cholesky matrix decomposition to avoid calculating the inverse matrix by analyzing the positively definite property of Hessian matrix, and then decompose Hessian matrix into a lower triangular matrix and an upper triangular matrix with converting it to a linear equation group solving question; another is that don’t do it directly when decomposing the matrix with Cholesky algorithm but uses a divide and conquer method, namely, first partitioning the matrix, decomposing the block matrix recursively, then calculating the block matrixs to get the final triangular matrix we want. This dissertation proceeds a numerical calculation of algorithm improvement theoretically. A conclusion is drawn that the improved algorithm is faster compared with the traditional standard and existing algorithms through the theoretical analysis and calculation of algorithms time complexity. A simulation experiment by Matlab is given, which confirms that the data results of experiment dovetails with the theoretical calculation approximately. At last, we use LMBP algorithm to model and train the meridian values which affect human sub-health status, and the prediction accuracy of final model to human sub-health status is approximate to83%. Confined to the factors such as lack of theoretical standard on medicine, complexity of human physiology meridian, coverage of district and age of participants and measurement errors of meridian value, the prediction result has a certain reference value.