Neural networks in specification level software size estimation

Presents a neural network approach to software size estimation. A multilayer feedforward network is trained using the backpropagation algorithm. The training and testing data consist of randomly generated structured analysis descriptions as input data and corresponding algorithm based size metric values as output data. The size metrics used in the experiments are Albrecht´s (1979) function points, Symon´s (1988) Mark II function points, and DeMarco´s (1982) function bang metric. The experiments indicate that neural networks can learn to calculate software size estimates. In each of the experiments it was found that the results depend on the features of the input data, the metric, and the convergence criteria used. The results also encourage the development of a general input set to represent size-related features of graph-based system descriptions