Approximate Newton Policy Gradient Algorithms