Ordering-based Conditions for Global Convergence of Policy Gradient Methods