An Empirical investigation into software effort estimation by analogy

Most practitioners recognise the important part accurate estimates of development effort play in the successful management of major software projects. However, it is widely recognised that current estimation techniques are often very inaccurate, while studies (Heemstra 1992; Lederer and Prasad 1993) have shown that effort estimation research is not being effectively transferred from the research domain into practical application. Traditionally, research has been almost exclusively focused on the advancement of algorithmic models (e.g. COCOMO (Boehm 1981) and SLIM (Putnam 1978)), where effort is commonly expressed as a function of system size. However, in recent years there has been a discernible movement away from algorithmic models with non-algorithmic systems (often encompassing machine learning facets) being actively researched. This is potentially a very exciting and important time in this field, with new approaches regularly being proposed. One such technique, estimation by analogy, is the focus of this thesis. The principle behind estimation by analogy is that past experience can often provide insights and solutions to present problems. Software projects are characterised in terms of collectable features (such as the number of screens or the size of the functional requirements) and stored in a historical case base as they are completed. Once a case base of sufficient size has been cultivated, new projects can be estimated by finding similar historical projects and re-using the recorded effort. To make estimation by analogy feasible it became necessary to construct a software tool, dubbed ANGEL, which allowed the collection of historical project data and the generation of estimates for new software projects. A substantial empirical validation of the approach was made encompassing approximately 250 real historical software projects across eight industrial data sets, using stepwise regression as a benchmark. Significance tests on the results accepted the hypothesis (at the 1% confidence level) that estimation by analogy is a superior prediction system to stepwise regression in terms of accuracy. A study was also made of the sensitivity of the analogy approach. By growing project data sets in a pseudo time-series fashion it was possible to answer pertinent questions about the approach, such as, what are the effects of outlying projects and what is the minimum data set size? The main conclusions of this work are that estimation by analogy is a viable estimation technique that would seem to offer some advantages over algorithmic approaches including, improved accuracy, easier use of categorical features and an ability to operate even where no statistical relationships can be found