Constructing an accurate model for predicting effort of a software system (total hours spent by all software developers involved) is a challenge in Software Engineering. This is mainly because software development effort varies depending on a large number of factors/variables, which often depend on the particular type of software system and the particular development environment. In addition, the influences of those factors/variables on effort also vary depending on other factors/variables, which are specific to the company/organization where the development is undertaken.
Hence, a useful approach to overcome this difficulty is to construct a specific model, which consists of all the influential factors/variables specific to the target system and environment, and calibrate the model using local data, which were collected by the company/organization from the previously developed similar systems, in a similar environment. Among a number of available modelling approaches, multiple regression in particular has been successful for constructing an accurate effort prediction model for a number of software systems in this way.
However, as the number of the similar software systems in the local calibration data decreases, predictive accuracy of the existing models also deteriorate. This consequently leads to very poor prediction if only a very small number of systems exist in the data. This talk introduces new Bayesian statistical software effort prediction models to address this issue and evaluates these models' predictive accuracy in two separate case studies, in comparison with the multiple regression counterparts.
Last modified: Thursday, 28-Jul-2005 17:23:30 NZST
This page is maintained by the seminar list administrator.