Download files
Abstract
Effort estimation is an important and challenging issue in software engineering. Software developers and/or project managers need to estimate the effort early in the software development life cycle, to evaluate the appropriate budget and manpower. Surprisingly, effort estimation for software customisation is not investigated extensively in the research literature. This thesis concentrates on understanding factors that influence the effort of customisation projects, and building an effort estimation model for software customisation projects.
A two-stage effort estimation model is proposed, that is analogous to COCOMO. The first stage of the model estimates effort based on general project characteristics. The second stage modifies this estimate by considering various productivity drivers. Two data sets are analysed to develop the models, both drawn from Release 12 of the ISBSG (International Software Benchmarking Standards Group) data repository.
For the first stage, a manual stepwise regression approach is used to form models from general project characteristics. The accuracy of the models is evaluated using mean and median absolute error, and mean and median magnitude of relative error.
To improve the accuracy of the stage 1 model, a set of productivity drivers is identified in the stage 2 model, that are related to variations in software productivity. Their effect is captured by multiplying the initial estimate by values that increase the estimate for things that imply worse than average productivity and reduce the estimate for things that imply better than average productivity. The resulting estimation accuracy is evaluated using the same criteria. Effort multipliers are determined by two approaches, one statistical and one using a Genetic Algorithm (GA). They are found to improve effort estimation accuracy significantly, with values found by GA having a better effect.
The specific coefficients in the models identified here may not apply in an organisation's particular context. However, we believe that the model structure we have proposed and validated, the factors that we found to be relevant for effort estimation, and the methodology we used are generally applicable.