Research On Combined Forecasting Method For Software Reliability Based On Multiple Neural Networks

With the popularity of computer and software widely used, software reliability predictionhas aroused increasing attention. As an important indicator, software reliability has directimpact on the quality of software, especially in some special areas such as the militaryindustry and financial field where the malfunction of software could lead to huge loss.Therefore it is urgency for us to work out a way to effectively measure and predict the qualityof software effectively in order to improve software reliability。The study of software reliability can be traced back to1960s and so far many softwarereliability models and methods have been proposed, such as the traditional MLR, Bayes,neural network, SVM and so on. Especially the neural network has higher predictionperformance in the field of forecasting for its self-adaptive and self-learning functions.However, it couldn’t upgrade the accuracy of software reliability prediction fundamentally byoptimizing individual neural network because of defects of the neural network such asover-fitting and local minimum. In contrast, the combination of those individual neuralnetworks can improve the generalization capability of the neural network and produceaccurate prediction despite the unsatisfactory prediction of a single neural network model.This article proposes software reliability prediction method based on neural networkensemble. After we use Bagging to generate an amount of individual neural networks, we useK-means clustering to select certain individual neural networks. In this process we use thedistance cost function to get the optimal k in order to solve the problem that value of k can’tbe selected automatically in the K-means clustering algorithm. Then we use dynamic weightto solve the output problem of neural network ensemble, and propose a dynamic weightedmodel base on fuzzy neural network, and solve its weight in different time by predicting theerror of single model to get the final result of neural network ensemble. And in this article, werespectively use two groups of experiments to verify the effectiveness of the proposed k valueoptimization algorithm which applied in neural network ensemble and the superiority ofDynamic weighted model compare to traditional fixed weight model.Finally, we combine the two improved methods in software reliability prediction base on neural network ensemble, and compare to the single neural network model and the traditionalneural network ensemble model. Experimental results show the application effect ofsatisfactory and the prediction result has been improved greatly compared to traditionalmethod.