Basis refinement strategies for linear value function approximation in MDPs