Implementation of AI Award - Analytics Awards 2020
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Analytics Awards 2020 Implementation of AI Award HEINEKEN Ireland Keg Census Background HEINEKEN produces draught beer and cider for the Irish market at its brewery in Cork City using locally sourced ingredients and services. Products produced in Cork include Heineken, Coors Light, Fosters, Orchard Thieves, Murphy’s and Beamish along with a range of smaller brands. Draught Beer and Cider are supplied in three sizes (20ltr, 30ltr and 50ltr) of returnable aluminium kegs. The kegs are filled in the brewery, distributed to pubs across Ireland and the empty kegs are then returned to the brewery to be re-used. While kegs can have a lifespan of many years, about 10% of kegs are lost every year for various reasons. The HEINEKEN Logistics function forecasts keg requirements and purchases new kegs every year to ensure there are enough kegs available to meet demand. Deciding how many kegs to purchase each year is an important business decision as too few kegs could result in lost sales and reputational damage, while too many kegs results in excess capital tied up and increased storage costs. The keg distribution journey does not always follow the diagram above and the attrition rates can fluctuate, so Logistics conduct a stock count (keg census) every year to determine how many kegs are in circulation. They count all the kegs in the brewery, storage, distribution depots and in outlets such as pubs and hotels. A number of alternatives to a physical count have previously been considered by Logistics but ruled out as not being practical or accurate enough to meet their needs.
Business Understanding – Describe the challenge Conducting a keg census in outlets presents two main challenges. Firstly, the census is conducted by delivery drivers but in many instances the kegs are delivered early morning before outlets open, with the drivers leaving the kegs outside so they don’t get access to the premises. The Logistics planners apply an average keg quantity to the outlets without a physical count to estimate the total keg population. Secondly, as there are about 7,500 outlets in Ireland, conducting a census in even half of them is a time consuming (and therefore expensive) process resulting in overtime costs and delayed deliveries. The objective of the project was to address either one or both of these challenges. The whole area of analytics was new to Logistics so it was clear that success would be dependent on Senior Logistics Management having enough confidence in the Machine Learning approach to deploy the solution. Data Understanding The standard approach adopted by the Analytics Team is to always start with the business problem or decision. Only then is the availability of the data which could help with the analysis considered and if it is not available then the challenge is how to obtain it or substitute for it. Following the integration of data from a number of sources, exploratory data analysis was carried out which was useful for developing the team understanding of the logistics data and as a first look at correlations amongst the variables. Not all the previously assumed correlations were shown to exist which led to some interesting discussions amongst the stakeholders. Data Preparation Normal data preparation tasks such as data cleansing, creating training and test data sets and feature selection were carried out. Data Modelling A number of different algorithms were evaluated for their suitability and it was decided to adopt a stacking approach to build the final model. Three tuned models utilising Linear Regression, Random Forest and Gradient Boosted Tree were selected to be trained and then combined by a Linear Regression meta-learner to return the final prediction model.
Evaluation The objective of the project was to predict how many kegs of each size are located in outlets. Therefore in order to evaluate the model performance, its prediction on previously unseen holdout data was compared to the actual keg quantities in the hold out data. However, there was a benchmark to compare the new model against as Logistics previously used an averaging method to estimate the keg quantities in outlets where it was unknown. If the Machine Learning model was not at least as accurate as applying an average then it would not be deployed. The table below outlines the performance of both the previous average method and the Machine Learning model developed during this project. The Machine Learning model delivered significant improvements in accuracy for both 20ltr and 30ltr kegs. While there was a slight reduction in accuracy for the 50ltr kegs, at only 0.1% it is quite marginal. Deployment This was a cross functional project which came about as a result of a request from the Logistics function for assistance with a problem they were having. The Machine Learning process developed out of this project has been deployed to replace the previous labour intensive process of physically counting kegs on customer premises. The advantages to Logistics include: • Improvement in the accuracy of keg census. • Reduction in cost of physical count. • Elimination of disruption to normal keg delivery operations. • Greater flexibility. Logistics are now free to conduct a keg census at any time without significant advance planning.
Deployment contd. The most recent application of the Machine Learning Keg Census was during the Covid-19 crisis. All Pubs in Ireland closed on 16th March 2020 resulting in significant business and logistical challenges for HEINEKEN. Outlets had on the premises at the time of closure over 100,000 kegs of various sizes made up of Empty kegs, Opened kegs and Un-opened kegs. In order to empower key decision makers across the business to make informed decisions impacting on credit exposure, product quality, logistics (uplifting kegs to be returned), supply of kegs to ensure availability of fresh stock for pub reopening, etc., there was an urgent requirement for a keg census. With the Pubs closed, no deliveries were taking place and with the Field Sales and Technical Service Teams unable to visit Pubs due to lockdown, there was no way to physically count the number of kegs in circulation. This was not the problem it could have been, as the Machine Learning Keg Census was able to determine how many kegs were in circulation and where they were, thereby allowing Logistics to proceed with their crisis planning.
You can also read