Abstract. The irrelevant and redundant features may degrade the learner speed (due to the high dimensionality) and reduce both the accuracy and comprehensibility of induced model. To cope with these problems, many methods have been proposed to select a subset of pertinent leatures. In order to evaluate these subsets, two main approaches are generally distinguished: (1) filter approach: which considers only data i.e. algorithm-independent; (2) wrapper approach: which takes into account both data and a given learning algorithm i.e. algorithm-dependent. In this paper, we address the problem of subset selection using a-RST (a generalized rough sets theory). We propose an algorithm to find a set of (~-reducts which are non deterministic reducts. To select the best one among them, we also propose a Hybrid Approach by putting filter and wrapper together to overcome the disadvantages of each approach. Our study shows that generally the highest-accuracy-subset is not the best one as regards to the filter criteria. The highest accuracy subset is found by the new approach with minimum cost.