CAJ | 학술논문

精确地估算软件成本是软件项目成功开发的一个重要保证,直接影响着软件的风险控制和质量保证。为了更好地解决单一估算模型的不足,提出了集成多案例推理（CBR）模型的软件成本组合估算模型。首先,采用六种距离计算公式刻画新旧项目相似度,构建了六种CBR模型,并运用粒子群算法（PSO）来优化CBR模型族中的属性权重。其次,在CBR模型族的基础上,运用支持向量回归机（SVR）模型将不同CBR模型的估算结果进行集成,提高软件成本估算结果的精度。利用Desharnais数据库对模型有效性进行检验,实证结果表明,在六种CBR模型中Euc-CBR、Min-CBR、Gau-CBR和Mah-CBR模型估算结果没有明显差异,Gre-CBR和Man-CBR模型略优;提出的SVR组合估算模型估算精度明显优于单CBR模型和其他线性组合估算模型,能有效提高软件成本的估算精度。
정학지고산연건성본시연건항목성공개발적일개중요보증,직접영향착연건적풍험공제화질량보증。위료경호지해결단일고산모형적불족,제출료집성다안례추리（CBR）모형적연건성본조합고산모형。수선,채용륙충거리계산공식각화신구항목상사도,구건료륙충CBR모형,병운용입자군산법（PSO）래우화CBR모형족중적속성권중。기차,재CBR모형족적기출상,운용지지향량회귀궤（SVR）모형장불동CBR모형적고산결과진행집성,제고연건성본고산결과적정도。이용Desharnais수거고대모형유효성진행검험,실증결과표명,재륙충CBR모형중Euc-CBR、Min-CBR、Gau-CBR화Mah-CBR모형고산결과몰유명현차이,Gre-CBR화Man-CBR모형략우;제출적SVR조합고산모형고산정도명현우우단CBR모형화기타선성조합고산모형,능유효제고연건성본적고산정도。
As a significant part of software process, software effort estimation plays a central role in controlling software cost, reducing software risk and guaranteeing software quality. The software development process is knowledge-intensive and under a dynamic development environment, which will increase the difficulty of solcware effort estimation. Therefore, software effort estimation is one of the most challenging activities in software development process. According to the idea of model integration and combination forecasting, the study puts forward the support vector regression based combination of multiple case-based reasoning to estimate software effort, in order to overcome the shortage of single estimation model and improve the estimation accuracy. The basic case-based reasoning （CBR） method for software effort estimation, including Similarity measure, Weight optimization, Number of most similar projects and project adaptation, Evaluation criterion, is introduced. Furthermore, six independent CBR methods, derived from Euclidean distance （Euc-CBR）, Manhattan distance （Man-CBR）, Minkowski distance （Min-CBR）, Grey Relational Coefficient （Gre-CBR）, Gaussian distance （Gau-CBR） and Mahalanobis distance （Mah-CBR） are constructed for software effort estimation based on the literature review. Moreover, particle swarm optimization （PSO） is adopted to determine suitable weights for each attribute due to its strong search capability. The CBR methods proposed in the study identify the estimation relationship between effort and other attribute from different aspects, and get a different estimation result, which will generates a gap between the estimation results. For this reason, the study applies the support vector regression （SVR） to combine the multiple CBR estimation results in order to integrate the estimated knowledge and improve estimation accuracy. Based on statistical learning theory, SVR have good generalization ability considering the structural risk minimization principle. The structure diagram of SVR based Combination estimation model is described detailedly. The experiment is carried out using Desharnais dataset. The study compares the results of different CBR methods from the MMRE, MdMRE and Pred（0.25） evaluation criterions at both training and testing stages. Experimental results indicate that there is not obvious difference between Euc-CBR, Min-CBR, Gan-CBR and Mah-CBR. However, the Gre-CBR and Man-CBR outperform the other independent methods slightly. The results from the comparison between combination estimation and single CBR estimation show that the SVR combination method outperforms the independent CBR methods. Moreover, the results also show that the SVR combination method can get more accurate estimation than other linear combination methods such as Mean Combination estimation model, Median Combination estimation model, Best Combination estimation model, WMC Combination estimation model.