在现实世界的优化应用中,机器学习与优化方法的结合已经成为决策分析的重要方法论。本文就是要讨论数据驱动下,带有不确定参数的优化问题。这种问题通常通过“Predict, then Optimize”的范式来解决。在涉及预测再优化问题中,机器学习方法的目的是最小化预测误差,并不关注如何将预测结果用于下游的优化问题中。
16
Elmachtoub and Grigns: Smart -Prodiet, then Optimize”
since for all . Clearly, replacing the constraint with in (il) results in an upper bound. Since this is true for all values of , then
In fact, one can show that inequality (治) is actually an equality using duality theory, and moreover, the optimal value of tends to . Intuitively, one can see that as gets large, then the term in the inner maximization objective becomes negligible and the solution tends to . Thus, as tends to , the inner maximization over can be replaced with maximization over , which recovers (5i). We formalize this equivalence in Proposition below.
Proposition 2 (Dual Representation of SPO Loss). For any cost vector prediction and realized cost vector , the function is monotone decreasing on , and the true SPO loss function may be expressed as
Using Proposition we shall now revist the SPO ERM problem (四) which can be written as