一种软件成本估算模型的评测和优化方法
文献类型：学位论文
| 作者 | 解浪 |
| 学位类别 | 硕士 |
| 答辩日期 | 2001 |
| 授予单位 | 中国科学院研究生院 |
| 授予地点 | 北京 |
| 导师 | 王永吉 |
| 关键词 | 软件成本估算 偏差 方差 预测区间 局部偏差 多组织数据 |
| 学位专业 | 计算机软件与理论 |
| 中文摘要 | 软件成本估算和管理是软件项目管理的核心任务之一，是项目计划、资源调度及人员分配的重要参考依据。在过去的几十年中，软件工程研究领域提出了大量的软件成本估算方法，但这些方法在实际应用中的效果并不十分令人满意，仍然有大量项目超支和失败。软件成本估算的误差过大是导致项目超支和失败的最重要的原因之一。评测并优化估算方法是提高估算效果的重要途径。如何计算估算误差是评测估算方法的基础，降低估算误差则是优化估算方法的核心研究内容。本文重点关注的参数化估算模型是现有的成本估算方法中常见的一种，该类模型大都需要历史项目数据来校准。一方面，由于软件开发过程中各个环节的种种不确定因素，历史项目数据会存在随机性，这种随机性将导致估算结果的不可信，给估算模型的评测和使用带来挑战。另一方面，由于软件项目数据的稀缺，校准时使用的项目数据可能来自于多个组织或者同一个组织时间跨度较大的项目，带来了历史数据的差异性。如何分析并解决历史数据的差异性对估算结果造成的影响也是一个挑战。针对以上的问题，本文提出了一种软件成本估算模型的评测和优化方法。其中，模型评测方法将估算误差的均值和方差相结合（Mean-Variance Combination, MVC），提出可以综合评测估算模型性能的指标。而优化方法部分首先针对COCOMO II的多组织数据特征，提出局部偏差的概念和度量方法。其次基于该度量方法提出了基于CII-B1和CII-B2两个新模型来处理这些局部偏差。最后，提出用推算出的工作量估算区间代替单一估算值。针对COCOMO系列模型提出了CBPI（COCOMO Based Predicted Interval）区间计算方法。综合以上的方法，本文设计了一个估算模型评测和优化实验系统。该系统验证了本文评测和优化方法的有效性：首先验证了MVC方法指标的稳定性超过传统指标（stdMRE和MMRE在交叉验证下的均值）；其次验证了CII-B1和CII-B2模型对局部偏差的处理效果；最后验证了CBPI方法获得的区间比直接使用估算误差获得的区间窄，同时保证工作量实际值在同样概率下落在区间内。该系统方便研究人员根据需求快速组装相应模块完成模型评测或者优化部分的实验 |
| 英文摘要 | Software cost/effort estimation and management is one of the core tasks of software project management. And it is an important basis for planning project, scheduling of resources and allocating human resource. In the past few decades, a number of software effort estimation method have arisen in the field of software engineering, however, the performance of these methods in application did not satisfy users. There exists large number of failed and overspent projects. One of most important reason that lead to the failing and overspending of project is that the estimation error is too big. Evaluating and optimizing estimation methods is an important approach to improve the performance of estimation. The foundation of evaluating estimation method is how to calculate the estimation error. The core work of optimizing estimation method is how to decrease the estimation error. This paper focuses on parametric estimation model, which is a common software effort estimation method. This type of model always needs to be tuned by history data. For one thing, because of the uncertainty factors in the development of software, the history data has randomness. It would lead to the incredible of estimation result and pose a challenge to the evaluating and applying of estimation model. For another thing, because of the history data’s scarcity, tuning approach always cross-company data. The cross-company data is the data from different companies or the data of different period at same company. The cross company data may has unconsistency. How to analyze and deal with such unconsistency to improve the performance of estimation model is another challenge.In order to solve the problem above, this paper proposes an evaluation and optimization method for software effort estimation model. One part of this method proposes a mean -variance combination method (MVC) for evaluating, which propose composited indicators to evaluate the performance of estimation result. Another part of this method is to do optimization. At first, in order to deal with the unconsistency of cross-company data, this paper proposes the definition of local bias and the metric of local bias. Base on the metric of local bias, this paper propose two new models, named CII-B1 and CII-B2, which are extended from COCOMOII model. After that this paper proposes a modified Ridge Regression method to solve new models. Finally, this paper propose to use predict interval in the application of estimation and propose a new interval method, named CBPI (COCOMO Based Predicted Interval) for COCOMO model. This paper designs an experimental system for the evaluation and optimization of estimation models. This system proves the efficiency of the three methods above. At first, it proves that the stability of MVC’s indicators is much better than the one of traditional evaluation method (the average value of MMRE and stdMRE). Second, it proves the efficiency of CII-B1 and CII-B2 for dealing with local bias. Finally, it proves that CBPI would get a narrower interval and retain the same probability of containing the value of real effort. The experiment system can help researcher to finish their experiments on model evaluation and optimization by assembling modules according to their requirement. |
| 学科主题 | 计算机软件 ; 软件工程 |
| 公开日期 | 2011-06-10 |
| 源URL | [http://124.16.136.157/handle/311060/10430]  |
| 专题 | 软件研究所_基础软件国家工程研究中心_学位论文 |
| 推荐引用方式 GB/T 7714 | 解浪. 一种软件成本估算模型的评测和优化方法[D]. 北京. 中国科学院研究生院. 2001. |
入库方式： OAI收割
来源：软件研究所
浏览0
下载0
收藏0
其他版本
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。