Comparison of different Neural Network architectures for software cost estimation

Abstract

This research will observe the use of Artificial Neural Networks (ANN) for estimating software cost. Constructive Cost Model (COCOMO) is the most famous estimating model for software cost, which will be used in this research. The model estimates software cost by calculating several variables which are created by expert with some equations. Furthermore, ANN helps to estimate COCOMO effort accurately. This research offers multilayer feed-forward neural network to adjust COCOMO effort estimation parameters. Also, an algorithm such as Back-propagation is applied to improve the architecture by comparing actual effort with estimated effort and updating the network. However, there are several types of neural network architecture. This research tries to compare several types of architecture by testing each architecture model to dataset. This paper concerns with two different architectures. The difference of this two architecture is basic architecture only uses effort multipliers as input layer while modified architecture divides input layer into two categories such as effort multipliers and scale factors. The result is the proposed model increases the accuracy and each model has different result.