同步最佳化預測和決策框架下的智造供貨管理

Abstract

採購與供貨成本控制一直是營運與供應鏈管理最重要的一環之一，許多領域廣泛地應用報僮模型與相關延伸來最小化成本，我們修改了經典報僮模型的基本假設，將原本為已知或是定值的商品價格設為隨機變數，利用相關變數來預測未來多期的價格，透過價格、需求與訂購數量之間的關係使成本最小化。大多數過去的研究會先利用統計或機器學習模型預測出所需要的參數，過程中並未考慮利用此參數推論決策後的損失；預測的最佳解未必是在做決策時的最佳解，因此，同步最佳化預測與決策的研究開始興起。在本論文，我們將報僮模型結合到此框架，並將其運用在深度學習的預測模型上，提供決策者一個較穩健並能最小化決策後成本的方法。

Newsvendor modeling is widely applied in many industry contexts to manage supply under uncertainty. We amend the assumption of classical newsvendor models to operate under uncertain feature-based prices over time. We minimize inventory cost through exploiting the relationship among price, demand and order quantity. Most existing studies are conducted in the manner of first predicting the parameters and then using them to make better decision. However, the best solution for pred6iction may not be the best solution for the sake of decision-making. Therefore, the study of simultaneously optimizing both prediction and decision is getting the more attention recently. We integrate the newsvendor model into the predict-then-optimize framework though the loss function of deep learning. Our modeling framework is operated in the robust lens by minimizing the regret of making the decision upon the predictive parameters.