Discriminatory software metric selection via a grid of interconnected multilayer perceptrons

Software metrics quantify source code characteristics for the purpose of software quality analysis. An initial approach to the difficulty in mapping source code to quality rankings is to multiply the number of features collected. However, as the number of metrics used for analysis increases, rules of thumb for robust classification are violated, ultimately reducing confidence in the quality assessment. Thus, a metric selection method is necessary. This paper examines the ability of a grid of interconnected multilayer perceptrons to select an appropriate subset of software metrics. Local interconnections between the multilayer perceptrons, in the form of feature evolution heuristics, allow publication of discriminatory features. The combination of competitive publication of discriminatory features with a limited number of inputs leads to classifiers that conform to robust classifier design rules. This paper examines the determination of discriminatory feature subsets by a grid of multilayer perceptrons in relation to a gold standard provided by a software architect.