Improved Evaluation of Accuracy of Software Development Effort Estimation Algorithms

Abstract

ABSTRACT newlineSoftware development effort estimation (SDEE) research has grown many folds during the last few decades and is still a challenge. There are many algorithms in the literature and no one is able to advice on the best method. Instead of the conventional train and test method of modeling, updating of the model after completion of each project is proposed. Two candidate methods selected are the effort estimation by recursive filtering using algorithmic Kalman filter (Kf) and the machine learning Online Sequential Extreme Learning Machine (OS-ELM). newlineOne can process data as and when the projects are completed which involves model updating after each project. However, the knowledge required for estimating staff may be higher compared to the conventional approach. Predict software development effort at the beginning of the project using previously developed model and update the model after the completion of the project using actual effort data. The Kf which can process data recursively using the state prediction and state estimation using measurements is successfully applied in many branches of engineering. Specifically, it was initially applied to space missions for state estimation with great success. Hence, the Kf is a right candidate for SDEE. The Kf can also take into account model and measurement uncertainties. This method demands both the dynamics and measurement models to be linear. The COCOMO model equation is transformed into a linear equation by logarithmic transformation and used for SDEE. The effort prediction is done for each project using the derived coeffic ients. The Kf newlinealso provides model coefficient uncertainty and estimated effort uncertainty. One can newlineuse this uncertainty to identify outliers in the data. Both simulated and real world data newlineare used for evaluation. The Kf works well for simulated data. In spite of the newlineadvantages, the Kf requires many inputs and may be difficult to get accurate results in newlinepractice as software development is a complex process where it is difficult to specify newlineaccur