老手駕駛的數位孿生: 以資料驅動的最後一哩配送解決方案

Abstract

隨著全球疫情的肆虐以及網路的發達，無接觸採買以及上網購物的需求增加，如何進行有效率得配送變得愈發重要。在這樣的趨勢下，物流業雇用越來越多沒有太多經驗的人來作為職業司機將變得無可避免，因此，如何有效地培養新手司機將貨物有效率地配送到每一位顧客手上變成一項非常重要的課題。 本篇論文旨在藉由學習老手司機的送貨策略來解決上述問題，以減少新手司機的訓練時間以及金錢成本。我們採用深度神經網路(deep neural network, DNN)以及基於規則(Rule-based)的方法來判別老手司機送貨路線中的停車地點，並將這些停車地點與當日的配送地址結合，來得知老手司機如何在送貨的過程中，根據送貨地址決定停車位置。 結果顯示在判斷送貨停車點的準確度包含precision, accuracy及recall分別可以達到91%、92%以及94%以上，而送貨地址配對停車點的正確率亦可超過95%。在測試資料中，113個建議停車點中與對照組方法相比，我們可減少6筆送貨地址與建議停車點距離超過30公尺之案例。

The increasing importance of improving logistic efficiency has become critical in recent times due to the global impact of the epidemic and the growth of the internet. As a result, it has become necessary to hire more drivers, including those with little experience as professional drivers, in order to effectively deliver packages to customers. The training of these inexperienced drivers poses a significant challenge. This thesis aims to address this issue by learning from the delivery strategies of experienced drivers to reduce training time and related expenses. We adopt deep neural networks (DNN) and rule-based methods to identify the delivery parking spots of experienced drivers along the delivery route and pair these delivery parking spots with the delivery addresses of the packages. Our results show that the precision, accuracy, and recall of the detection of delivery parking spots are all above 91%, 92%, and 94%, respectively. The accuracy of pairing results of delivery packages and delivery parking spots can exceed 95%. Compared with the baseline method, our proposed method can reduce the number of cases where the distance between the delivery address and the recommended parking spot is more than 30 meters by 6.