The Structural Optimization Research For Network Based On IAGP Algorithm

With the development of society and industry, the neural network has beenwidely used in many areas of life. However, the application of neural networks cannottake full advantage of the performance of the network, for example, the networkcannot get a good generalization, the precision is very low, convergence speed is veryslowly, training time is relatively long, training error is not small. In recent years,people found that the adjustment of adaptive network structure influences theperformance and efficiency of the network, so the optimization of neural networkstructure is an urgent problem to be solved. And then scientists have proposed somenew algorithms and the improved new algorithm to optimize the neural network.In view of the above problems, in this paper, The IAGP algorithm which is basedon ZhaMiNa’ s AGP algorithm, a new pruning algorithm and a new growingalgorithm. Network pruning is based on the sigmoidal activation value of the nodeand all the weights of its outgoing connections, Network growing is based on thecorrelation of significance measure’s variance, we direct copy those nodes. And IAGPalgorithm firstly takes pruning algorithm to make the network’s structure simplified, ifit do not achieve network performance requirements, then adopts growing algorithmto add nodes, after that, if it still cannot meet the requirements, then the networktraining will be iterative, until it meets the network performance requirements. Inorder to test the effectiveness of the proposed algorithm, the paper made a simulationexperiment about comparing every algorithm with other algorithm, and IAGPalgorithm is applied to the prediction of transport capacity and the total output valueof domestic tourism, the simulation experiments show the new algorithm is proposedin this paper can get the expected effect in the practical application, the training time,the convergence rate, the ability of generalization, training error, these show itssuperiority.The main research work of this paper are as follows:Firstly, this paper systematically describes the pruning algorithm, growingalgorithm and hybrid algorithm’s research status and some disadvantages, and itintroduces the application and classification of each algorithm.Secondly, introduce the basic knowledge of neural network (including theintroduction, development history, structure, characteristics and application), analyze the BP algorithm, and expound BP’s core idea.Thirdly, analysis of the factors affecting network performance, including thetraining time, convergence rate, generalization and errors etc.Fourthly, based on the above shortcomings, propose network pruning algorithmwhich is based on the sigmoidal activation value of the node and all the weights of itsoutgoing connections, it improves the efficiency of pruning, but also makes sure theperformance of the network. This paper presents a new growing algorithm, it is basedon the contribution of variance correlation, directly copy those related to high degreenodes, makes the network convergence’s speed quick, at the same time, it alsoprevents over fitting phenomenon, and does some comparative experiments, it wasachieved good results.Fifthly, put forward IAGP algorithm, according to certain rules, take networkpruning algorithm and growing algorithm to optimize the neural network structure,and take the algorithm in the approximation of the nonlinear function, theexperimental results was quite satisfactory.Sixthly, IAGP algorithm is used to predict the transportation capacity and thetotal output value of domestic tourism, do comparative experiments, analyze andcompare the advantages and AGP algorithm in solving the problem.