Decision-Focused Learning without Decision-Making: Learning Locally Optimized Decision Losses