软件成本估算的粒子群算法类比模型及自助法推断

Analogy Estimation Based on Particle Swarm Optimization and Bootstrap Inference for Software Effort

软件成本估算是控制软件进度、降低软件风险和保证软件质量的有效措施,已引起产业界和学术界的广泛关注。为能更准确地估算软件成本,采用粒子群算法优化加权类比估算模型(PSO类比模型)中各个特征属性的权重,避免权重选择的盲目性;同时采用非参数自助法对原始数据抽样,在自助法子样本的基础上,分析PSO类比模型中的类比项目个数和成本计算方式对模型精确度的影响,并进一步计算新项目估算成本值的可信度以及在一定置信水平下的区间值;采用Desharnais数据库验证PSO类比模型与自助法推断方法的有效性,根据MMRE和Pred(0.25)两个标准将PSO类比模型与一般类比模型、支持回归机、人工神经网络、径向基神经网络和分类回归树进行估算精度比较。研究结果表明,采用粒子群算法优化权重的PSO类比估算模型能得到较高的估算精度,且自助法能有效校正PSO类比模型中的相关变量并检验结果的可信度,可以为管理者在软件项目风险分析和项目规划方面提供有益参考。

Wide attention has been attracted on software effort estimation by both software industry and academic community owing to its high ability in controlling software schedule,reducing software risk and guaranteeing software quality.This paper investigates the improvement effects of estimation accuracy in analogy-based model when particle swarm optimization method is adopted to optimize the feature weights.Meanwhile,nonparametric bootstrap was employed to generate samples for calibrating the number of the most similar projects and the project adaptation in PSO analogy model.Reliability of estimated value and confidence interval were calculated by nonparametric bootstrap too.Experiments were carried out using software projects from Desharnais dataset in order to verify the effectiveness of PSO analogy model and bootstrap inference method.Estimation accuracy of the PSO analogy model was compared with ordinary Analogy,SVR,ANN,RBF and CART in terms of the error measure which is MMRE and Pred（0.25）.The empirical results show that applying particle swarm optimization method to optimize the feature weights is a feasible approach to improve the accuracy of software effort estimation.Moreover,nonparametric bootstrap can calibrate the variables in PSO analogy model and calculate estimation accuracy and confidence interval effectively.Results of the proposed model are beneficial to assess estimation accuracy,analyze risk and plan project.