Size and effort-based computational models for software cost prediction

Abstract

Reliable and accurate software cost estimations have always been a challenge especially for people involved in project resource management. The challenge is amplified due to the high level of complexity and uniqueness of the software process. The majority of estimation methods proposed fail to produce successful cost forecasting and neither resolve to explicit, measurable and concise set of factors affecting productivity. Throughout the software cost estimation literature software size is usually proposed as one of the most important attributes affecting effort and is used to build cost models. This paper aspires to provide size and effort-based estimations for the required software effort of new projects based on data obtained from past completed projects. The modelling approach utilises Artificial Neural Networks (ANN) with a random sliding window input and output method using holdout samples and moreover, a Genetic Algorithm (GA) undertakes to evolve the inputs and internal hidden architectures and to reduce the Mean Relative Error (MRE). The obtained optimal ANN topologies and input and output methods for each dataset are presented, discussed and compared with a classic MLR model.