Validating and understanding software cost estimation models based on neural networks

This paper presents the cost estimation models based on artificial neural networks. one of the most important limitations of neural networks is the difficulty of understanding a neural network that makes a particular decision. For application to the cost estimation field, the neural network is used to predict the software development effort is the Radial Basis Function network. The COCOMO´81 dataset is used to train and test the RBFN. The accuracy of the RBFN depends essentially on the parameters of the middle layer, especially the number of hidden neurons and the values of the widths. After evaluating the accuracy of the RBFN, the Jang and Sun method is applied to extract the if-then fuzzy rules from the artificial neural networks. These fuzzy rules express the information encoded in the architecture of the network.