Manipulation of Training Sets for Improving Data Mining Coverage-Driven Verification

The constant pressure for making functional verification more agile has led to the conception of coverage driven verification (CDV) techniques. CDV has been implemented in verification testbenches using supervised learning techniques to model the relationship between coverage events and stimuli generation, providing a feedback between them. One commonly used technique is the classification- or decision-tree data mining, which has shown to be appropriate due to the easy modeling. Learning techniques are applied in two steps: training and application. Training is made on one or more sets of examples, which relate datasets to pre-determined classes. Precision of results by applying the predictive learning concept has shown to be sensitive to the size of the training set and the amount of imbalance of associated classes, this last meaning the number of datasets associated to each class is very different from each other. This work presents experiments on the manipulation of data mining training sets, by changing the size and reducing the imbalances, in order to check their influence on the CDV efficiency. To do that, a circuit example with a large input space and strong class imbalance was selected from the application domain of multimedia systems and another one, with a small input space that affects the coverage occurrences, was selected from the communication area.