Feature-Based Q-Learning for Two-Player Stochastic Games