3-1 運輸箱設計的機器學習方法.pdf

《3-1 運輸箱設計的機器學習方法.pdf》由會員分享，可在線閱讀，更多相關《3-1 運輸箱設計的機器學習方法.pdf（15頁珍藏版）》請在三個皮匠報告上搜索。

1、A Machine Learning Approach to Shipping Box DesignGuang Yang,Cun(Matthew)MuJ To minimize the overall shipping cost by strategically selecting the best combination of k box sizes from thousands of feasible ones to be responsible for hundreds of thousands of orders daily placed on millions of inventor

2、y products.Overall lowest shipping costPack use k=18 box sizesselect from thousands of feasible box sizesXXXOverview Key:view each box size as a point in space and formulate the box design problem as a generalized version of weighted k-medoids clustering problem to recommend k=18 box sizes.Q:How to

3、define and measure the weight Q:How to define and measure the weight w wj jand distance and distance c cij ijfor each box size?for each box size?Box:10”x8”x7”Box:9”x8”x8”Box:22”x5”x3”Note that it is possible that c cij ij c cji ji,and often this is the case.Workflow Generate all candidate box sizes.

4、Pack customer order into full box set,often thousands of box types.To determine the box economic value,e.g.,weight wj.Evaluate how easily box Bican replace another box Bj,when an order is best packed(highest utilization rate)by Bj.Define box-box economic substitution cost cij.Solve the generalized w

5、eighted k-medoids clustering problem given wjand cij.Parameter tuning and result evaluation.1.Generate all candidate box sizes Requirements:This yields 3,391 different types of boxesBox IDLength(in)Depth(in)Height(in)Volume(in3)Weight Limit(oz)B0001853120800B0002953135800800B339136211410584800volume

6、=10800(due to handling capability)box size volume(in3)shipping rates($)2.Pack customer orders into full box set.Pack 3-months historical orders using all 3391 boxes.We developed a super efficient packing algorithm called gbp that can solve this 4D packing within 1%suboptimality,but 100 x or even mor

7、e faster than solving mixed integer programming problem using Gorubi.OrderOrder IDIDTicketTicket IDIDBoxBox ID ID SKUSKUx xy yz zDateDateFCFCX001T001B2342A04062996420002017/01/01FC1X001T002B1578A49517610453302017/01/01FC1X001T002B1578A49517610450302017/01/01FC12.Pack customer orders into full box se

8、t.This yields the box js weight wjwhich is estimated descriptively by its(discounted)effective volume(EV)contribution.Box ID#SKUsTotal Volume Packed-EV(in3)#TicketsTotal Box Volume(in3)WjB00019799309982426751204060.54%60.54%B00027374271228251233912079.98%79.98%B33911771014080574749744882.03%82.03%No

9、te:assume =1.3.Define box-box substitution cost cij A box type Biis competitive if products packed by other box types Bjs can be easily repacked using Biwithout too much sacrifice.222421322222-1-0.10-0421-0.20-1-0.053220.67-0.30-13.Define box-box substitution cost cijSpecifically,4.Select k=18 from

10、n=3391 box sizes Solve Imagine each box is a point on the space.The box economic value wjrepresents its contribution in packing.The box economic substitution score cijrepresent its capability in replacing other boxes in packing.Solve this generalized weighted k-medoids clustering problem.We develope

11、d a generalized k-medoids clustering problem solver skm using expectation maximization(EM),or greedy approach.XXXNote that it is possible that|i|j|and c cij ij c cji ji,and often this is the case.5.Parameter tuning and result evaluationAchievement Consistent performance by increasing K.Utilization R

12、ate(%)CurrentCurrentProposedProposedAchievement Consistent improvement in utilization rate and reduction in order split rate across date and fulfillment center(FC)/retail stores.FC1 FC2 FC3 FC4Utilization Rate(%)DateFulfillment CenterCurrentCurrentProposedProposedKey Developments gbp Solve 1D-4D bin

13、 packing problem using best-fit-first strategy in a recursive manner,take care of the weight constraint and handle the bin split.CRAN:https:/cran.r-project.org/package=gbp Demo:https:/gyang.shinyapps.io/gbp_app/skm Solve generalized k-medoids problem using expectation maximization(EM)based solver,or greedy approach based solver.CRAN:https:/cran.r-project.org/package=skm Demo:https:/gyang.shinyapps.io/skm_owl/Thank you!