Categorization of programs using neural networks

This paper describes some experiments based on the use of neural networks for assistence an the quality assessment of programs, especially in connection with the reengineering of legacy systems. W e use Koho-nen networks, or self-organizing maps, for the cat-egorization of programs: Programs with similar fea-tures are grouped together in atwo-dimensional neigh-bourhood, whereas dissimilar programs are located far apart. Backpropagation networks are used for general-ization purposes: Based on a set of example programs whose relevant aspects have already been assessed, we would like to obtain an extrapolation of these assess-ments t o new programs. The basis f o r these investiga-tion is a n intermediate representation of programs in the f o r m of various dependency graphs, capturing the essentials of the programs. Previously, a set of met-r i c ~ has been developed to perform an assessment of programs on the basis of this intermediate representa-tion. I t lis not always clear, however, which param-eters of the intermediate representation are relevant for a particular metric. The categorization and gener-alization capabilities of neural networks are employed to improve OT verify the selection of parameters, and might even initiate the development of additional met-rics.