Approximate Newton policy gradient algorithms