A Q-learning approach to attribute reduction

Attribute reduction is a paradigm covering both theories and techniques of selecting required attributes with constraints related to rough set. Currently, though various searching strategies have been developed for achieving such a purpose, few of them take the rewards of identifying attributes into account. In this study, inspired by the popular reinforcement learning mechanism, a Q-learning based procedure is designed to search qualified attributes and then construct the expected reduct. Specifically, state is regarded as the temporary result of selected attributes, action is regarded as variation of such a temporary result if random strategy is performed. Immediately, the reward can be obtained, which offers guidance on identifying attributes with the greatest reward. Moreover, considering the random factors emerge in our scheme, an ensemble device is also used to further improve classification performance of selected attributes in reduct. Finally, comprehensive experiments over a total of 15 UCI datasets clearly validates the superiorities of our study against 5 state-of-the-art approaches.