In the wrapper approach to feature subset selection the learning algorithm itself is used in the selection process, thus making the process sensitive to the bias of the learning algorithm. The approach introduces a useful framework within which search algorithms and classifiers can be compared. To date, search algorithms used in conjunction with the wrapper method have focused on greedy search techniques, i.e. those which direct their search according to increased accuracy arising from the addition/removal of individual features. It is argued that since the error rate of the classifier is non-monotonic such searches may prune valid areas of the search space. In addition, the combinatorial nature of greedy searches makes their use infeasible for large (>20) feature sets. An alternative form of search, based upon the principles of genetic algorithms, is proposed and related work is discussed. Results indicate that improvements reported with the wrapper method are maintained whilst the robustness and applicability of the approach is increased. (8 pages)