A game theoretic framework for feature selection

Feature subset selection plays a key role in both dimensionality and noise reduction. Moreover, it is often used to enhance accuracy in classification and clustering problems while decreasing their complexity. Inspired by Markov Decision Process, the presented paper considers feature subset selection as a one player game and uses Reinforcement Learning paradigm to select best features. In order to have an optimal traverse in the search space, we introduce a Monte Carlo graph search to overcome the complexity of the problem of concern. Finally, a low cost evaluation function is used to evaluate each state. The evaluation function leads search process into the most promising regions by rewarding each state. The results on the benchmarks prove superiority of our method over other well known methods in the literatures.