基于DBN的软件可靠性预测模型的研究  

Software reliability prediction model based on deep belief networks

作　　者：王国涛[1] 石振国[2] 吴小景 Wang Guotao Shi Zhenguo Wu Xiaojing(School of Electronics ＆ Information School of Computer Science ＆ Technology, Nantong University, Nantong Jiangsu 226019, China)

机构地区：[1]南通大学电子信息学院,江苏南通226019 [2]南通大学计算机科学与技术学院,江苏南通226019

出　　处：《计算机应用研究》2016年第12期3739-3742,3773,共5页Application Research of Computers

摘　　要：安全攸关系统广泛应用于交通、工控、航空等与国计民生相关的安全攸关领域,对可靠性有着非常高的要求。而控制软件往往是安全攸关系统的核心,因此对它的可靠性预测精度必须达到很高的要求。将深度置信网络(DBN)应用于软件可靠性增长预测模型(SRPM)的研究。针对DBN中核心模块RBM的无监督学习,采用了动态模式跳转算法(DMH)。该算法通过动态地维护一个模式集,然后借助模式集中模式的跳转来完成RBM中状态的跳转,使RBM的无监督学习具有很高的学习效率。通过与参数动态调整的动态模糊神经网络(SADFNN)、BP神经网络(BPN)以及基于萤火虫算法的BP神经网络(FABP)建立的SRPM进行预测能力的比较,实验结果表明基于DBN建立的SRGM的预测结果精度最高且最稳定。Safety-critical system is widely used in transportation, industrial control, aviation and other areas related to the national security and people＇ s livelihood, it needs extremely high reliability. Control software is usually the core of security-critical system, so its reliability prediction accuracy must be very high. This paper applied the deep belief networks （DBN） to the prediction accuracy of software reliability prediction model （SRPM）. It used the dynamic mode-hopping MCMC （DMH） for the unsupervised learning of RBM which is the kernel module in DBN. The algorithm was based on dynamic maintenance a model set, with the help of the model-hopping to complete the state-hopping of RBM, making the unsupervised learning of the RBM have a high learning efficiency. The SGPM＇ s predictive ability established by DBN compared with SGPM＇ s predictive ability established by dynamic fuzzy neural network with parameters dynamic adjustment （SA-DFNN） BP neural network （BPN） and the firefly algorithm of BP neural network （FABP）. Simulation results confirm that the prediction of the SRPM based on DBN is highest and most stable.

关 键 词：深度置信网络 软件可靠性预测模型 动态模式跳转 限制波尔兹曼机 无监督学习 

分 类 号：TP311.5[自动化与计算机技术—计算机软件与理论]