Guided Policy Search via Approximate Mirror Descent