Integration Of Neural Networks Based On The Bee Colony Algorithm

The neural network ensemble based on artificial neural network to learn the sameproblem at the same time, by participating in the integrated neural network integratedcombination of the output of the same input example input conditions, that is output by theformintegrated neural network output in this example is to co-decision. Generalization error isequal to the average generalization error of the integration of individual network and thedifference of the mean difference, so if you want to enhance the generalization ability ofneural network ensemble purpose should first maximize the generalization ability of theindividual network Secondly, we should try to improve integration in the degree of differencebetween the various networks.In this paper the neural network ensembles are used for regression problems and we dosome researches on how to produce individual neural network in ensemble and design a goodneural network ensemble algorithm to try and obtain better generalization performance.The main work and innovation in this paper are as follows:(1)Aiming at multiple repeat samples in the subset of training samples which isgenerated by the Bagging method(not help to improve its generalization ability). This paperproposes a techniques that guarantee a certain size of training samples, randomly selected asmall portion of the samples to obtain a training sample subset. This training sample setwould have neither the duplicate samples.(2)Aiming at the difficult to determine the nodes number of the hidden layer of BPneural network in practical applications. This paper presents a way to determine the BP neuralnetwork hidden layer nodes. Using of artificial bee colony algorithm to optimize the structureand connection weights of BP neural network through a special encoding.(3) The paper presents a selective average integration based on the correlation of thesmallest individual network combination. Firstly, all individuals classified. Secondly, eachgiven a network number of individual combinations of representatives of the ABC algorithmcombined search of all the individual networks, minimum correlation coefficient of the simpleaverage of the individual group cooperation as such. Finally, all delegates once again theaverage integrated. While another is based on the network output sequence variance measureindividual differences in the degree of selective integration, namely the use of the ABCalgorithm to combinatorial optimization, select a maximum variance combination to integratethe predictive value of all network.(4) Using neural network selective ensemble based lowest correlation coefficient andneural network selective ensemble based maximum variance combination to apply in typhoon intensity prediction and regional precipitation forecasts, and the test results show that itspredictive ability is better than general neural network ensemble method, can be used for theactual prediction...