A new convergent variant of Q-learning with linear function approximation