Active Learning of Soft Rules for System Modelling

Using rule learning algorithms to model systems has gained considerable interest in the past. The underlying idea of active learning is to let the learning algorithm influence the selection of training examples. The presented method estimates the utility of new experiments based on the knowledge represented by the existing rulebase. An extended rule format allows to deal with uncertainty. Experiments with different artificial system functions show that the presented method improves the model quality respectively decreases the number of experiments needed to reach a specific level of performance.