Natural Policy Gradient Methods with Parameter-based Exploration for Control Tasks