Towards Zero Shot Learning in Restless Multi-armed Bandits