預測採購與在線學習銷售之最佳化模型

Abstract

製造業是台灣的核心產業之一，一直以來，採購與銷售是控制製造業營運成本與獲利來源的探討重點，傳統建模上會從客戶的需求出發，依照客戶的需求決定產品組合，然而在造紙產業、橡膠產業等交貨時間長的傳統產業卻不是如此，需求對決策的擬定影響通常小於原物料價格波動帶來的影響，因此，本篇論文傾向開發有別於傳統從需求端建模的方式，從採購端出發，將主問題切成兩個子問題，即購入再售出，在購入階段，問題會以組合優化問題呈現，採購經理人會從眾多購入憑據中挑選出能使成本極小化的買進策略組合，我們透過機器學習模型進行購買憑據的分類，在「先優化再預測」的框架下對未來的買進數量、成本進行預測，然而，在銷售的環節，則是在接到訂單的同時，需立即作出相對應的決策，屬於在線學習優化問題，結合實務上的資料，致力於開發資料驅動的解方，提供決策者一套穩健的利潤極大化的銷售策略。

Manufacturing is one of the Taiwan's core industries. Procurement and sales have always been the focus of discussion on controlling manufacturing operating costs and profitability. Traditional modeling starts from the demand of consumers and determines the product set based on the demand of consumers. However, this is not the case in traditional industries with long lead time such as the paper industry and rubber industry. The impact of consumer demand on decision-making is usually less than the fluctuation of raw material prices. Thus, different from the traditional way of modeling from the demand side, starting from the purchasing side, the main problem is divided into two sub-problems, namely buying and then selling. In the purchasing stage, the problem will be presented as a classification optimization problem. The procurement manager will select a combination of features which can minimize the set of the raw material buy-in cost. We classify the features through the machine learning model, and predict the future purchase quantity and cost under the framework of smart predict then optimize (SPO). However, In the selling stage, when an order is received, the corresponding decision needs to be made immediately, which belongs to the online learning optimization problem. Combined with practical data, it is committed to developing data-driven solutions and providing decision-makers along with a set of profit-maximizing sales decisions.