Robust Policy Learning over Multiple Uncertainty Sets