The Research On Sotfware Reliability Measure Model And ELM Algorithm

Inthepastseveraldecades,softwarehasbecometheindispens-ablefactorinthesocialmodernization. Itisappliedtoanincreasinglynum-ber of key areas of social development. Along with the improvement of thesocial intelligent, computer software plays an increasingly important role.The reliability of software is the lifeline to these industries that have highrisks and require high security.However, Software reliability technology,compared with the technology of hardware reliability, is still in its infancy.And the research about software reliability is not mature neither in the the-ory nor in the applications, which draws more attention of researchers inthe software engineering. So far, mathematical modeling based on residualerrors of software is still very popular.In this thesis, based on comprehensive analysis of software failure da-ta, thethesismainlyinvestigatessoftwarereliabilitymodelingwiththepur-poses of proposing new models and new methods. The thesis achieved thefollowing innovative research results on these aspects.Firstly,softwarereliabilitymodelingwithremovederrorsandcompou-nded-decreased-rate is proposed. Based on the three well-known softwaremodels based on the rate of removing errors, a new model is proposed,which has more applications and is more consistent with the practice. Themodel parameters are estimated by the first-order differential combinedwith the least square method, finally the capacity of the proposed modelis verified by the public databases. The experimental results show that themodel is in more accord with the actual software testing and more valuablefor practical applications.Secondly, a new software reliability estimation model based on piece-wise quartic polynomial is proposed. Given the long-term and complex software testing process and the feature of software failure data, the bend-ing function and curve based on the piecewise quartic polynomial are es-tablished, then software reliability estimation model based on piecewisequarticpolynomialisdesigned. Becauseofthegoodlocalaccurateinterpo-lation property of this curve,the outputs of the model in the control pointsare superposition with control points, the experimental results on the in-ternational databases of software errors also show the effectiveness of themodel.Thirdly, new machine learning technology is used. Extreme LearningMachine algorithm, a new learning algorithm attracting much attention re-cently, has been applied in many areas. However, there is sparse researchin the field of software reliability and few relevant literatures. So in thisthesis, the method randomly choosing input weights and biases in the o-riginal ELM algorithm is modified, instead, a brand new algorithm is pro-posed. The improved algorithm is applied to the approximation of knownfunction, and then to the simulation of public databases. The results showthatthe new machine learning method can be appliedto softwarereliabilityestimation, but the performances need to improve by better theory support.