Determining a good feature set is critical to the performance of learning algorithms such as classifiers. Recently, researchers have proposed evolutionary-based feature extraction methods that aim to find a good feature set by combining the original features with new features generated by mathematical transformations of the original features. In this paper, we propose dynamically collecting past performance information on promising features and operators to use in our mutation method. We consider how to make our evolutionary algorithm more efficient and reliable by reducing overfitting. Preliminary results using UCI data show that our dynamic mutation method only slightly enhances the classification accuracy but it produces more reliable results.