Effort estimation is an important phase of software development projects. Estimation results are notoriously known for their inaccuracy. Numerous research studies tackled the problem, leading to a wide variety of effort estimating approaches that can be divided into several categories: estimation by analogy, parametric models, expert estimation, and artificial intelligence methods. Neural network-based models belong to the artificial intelligence (or machine learning) category, which can also employ fuzzy logic, regression trees, rule induction, Bayesian belief networks, evolutionary computation, grey relational models, and so on.
The authors present a brief review of publications (from 1993 to 2010) on the application of neural networks to cost estimation, and mention the strengths and weaknesses of the method. They emphasize that software effort estimation requires a sophisticated understanding of the relationships between the input factors (for example, project size and number of function points) and output estimation values. Neural networks have certain advantages over commonly used models such as the constructive cost model (COCOMO) and software life-cycle management (SLIM), which are based on single or fixed relationships between the inputs and outputs. The authors summarize prior research efforts on neural networks based on the databases used, evaluation criteria, and model input parameters. The conclusion is that the results of the publications reviewed provide evidence of neural networks being a suitable tool for software effort estimation, capable of outperforming other methods in some cases. Successful application of the neural networks method is limited by the insufficient size of the historical estimation databases, as this method is based on learning. The paper is not a systematic review, as it is based on a set of 21 papers, and the authors don’t describe the literature search process and how the papers for the review were selected. That leaves the question of the review completeness open. Actual estimation accuracies--the main driver for effort estimation research--are not described.
Academics and practitioners interested in software effort estimation are the intended audience of the paper. Readers who are interested in this topic can find additional information in other papers [1,2].