基于组合神经网络的软件可靠性预测研究

A STUDY ON SOFTWARE RELIABILITY PREDICTION BASED ON ENSEMBLE NEURAL NETWORK

为了进一步提高神经网络的预测能力,提出了一种前馈神经网络混合学习算法,并将其应用于组合神经网络.该算法由一种模式提取算法(Alopex)和伪逆算法组成.在该混合学习算法中,网络的学习任务被分解为2个部分:隐藏层的权值先随机给定,然后使用Alopex算法不断地对其进行扰动;输出层的权值使用伪逆算法确定.所使用的组合神经网络由多个结构相同的前馈神经网络组成,每个前馈神经网络都使用混合学习算法(采用不同的初值)进行训练.实验结果表明,这种组合神经网络能够显著提高软件可靠性的预测精度.

To further improve neural networks prediction capabilities, a hybrid learning algorithm is presented and applied to an ensemble neural network. The hybrid algorithm combines the algorithm for pattern extraction （Alopex） and the pseudoinverse algorithm. By using the hybrid learning algorithm, learning tasks are divided into two parts：weights in the hidden layers are given randomly and perturbed continuously in some directions by using Alopex; weights in the output layer are gained by using the pseudoinverse algorithm. Then, several feedforward neural networks are assembled into an ensemble neural network. Each of these feedforward neural networks is same in structure and trained by using the hybrid learning algorithm （initial value is different）. The results of experiments indicate： the ensemble neural network can improve greatly the accuracy of software reliability prediction.