Picking Optimization

Genetics & Machine learning at work in ASDA store operations

Team:

Supply Chain Technology

Published Date:

01.16.18

Locations:

Sunnyvale, CA

Bangalore, India

Picking Optimization

Technologies

  • Scala
  • Java 8
  • Spring Boot
  • Apache Ignite
  • Apache Cassandra
  • Elastic Search
  • IBM MQ
  • Medusa
  • Stati Platform
  • Spark
  • Grafana

Overview

ASDA, a British supermarket retailer giant, became a subsidiary of Walmart after a takeover in July 1999 and currently stands at 3rd rank by market share in the UK. ASDA operates online grocery delivery and pick up services at more than 300+ stores delivering over a billion items across thousands of trucks trips each year. A majority of the cost of fulfilment annually is attributed to the ins-store order picking process as most of the e-commerce grocery orders are picked and shipped from ASDA stores. Improving in-store order fulfilment velocity would ensure that more items are processed per hour/picker thereby improving the capacity of online orders fulfilled in a day (~directly impacting revenue), increasing on-time deliveries and driving operational costs down.

Customers who wish to order grocery from ASDA go on the ASDA grocery website, build their cart and book an appointment to receive their order at home. A location routing engine would create optimized routes based on the incoming orders with fulfilment processes working towards ensuring that these trucks are loaded and sent on their way, on-time. The picking optimization engine specifically deals with optimizing the manner in which these grocery orders are picked by store associates minimizing the time required to pick them.

This is a tactical problem that tackles 2 intertwining decision points: distributing orders among batches and determining a batch-picking route for each batch in order to minimize the total pick time to pick all items of all orders. Efficient picking process is however also a part of a multitude of larger strategic design decisions including Product storage assignment, Layout design, Order release strategy, Routing strategy and Level of automation. Some of the main challenges with the picking optimization problem arises due to the fact that picking was executed in retail-stores and may of the design decisions couldn’t be impressed upon.

Challenges

Picking groceries efficiently poses many challenges specially in a retail store environment as many proven warehouse-picking efficiencies cannot be adopted due to lack of automation, merchandize driven layouts and the nature of the orders. The picking process also tends to be manual not only increasing fulfilment costs but also limiting revenue opportunities making the picking process a bottleneck to the number of orders that could be promised to customers on the website.

Large order sizes: Unlike electronics or apparel orders, Grocery orders tend to be very large in size with items ranging from Milk and Meat to Potatoes and Pizzas

Cold chain compliance: Most frozen or chilled items cannot be kept out of the cold chain for long periods of time adding complexity and cost to the picking process. This requires shorter pick paths reducing the number of items that could be batched together

Complexity of batching orders: Hundreds of orders with thousands of items to be picked across thousands of square feet of shop floor gives rise to millions of combinations of how these items could be picked

Locating items on the shop floor: Tracking thousands of item locations on the shop floor with accurate co-ordinates and obstacle markings (staircase, lift, walls etc)

Limitations of the picking process: The space and weight constraints of the picking device along with other order related constraints of food safety regulations, maintaining quality and operational alignment limits the extend of optimization opportunities

Store environment: The picking process has to be tailored to store picking environment where customers would walk in the same aisles as the pickers. Another important aspect of the picking process is the accuracy of inventory and the quality of the products (expired dairy products or rotten bananas)

Complex Cost function: The optimization cost function tends to be very complex in terms of incorporating all probable costs and also in terms of measuring and separating them from each other. Some of the picking costs included are walk time, item seek time, item retrieval time, scan times, trolley loading, angle of entry and exit into aisles, tote consolidation etc.

Low tote fills and Low pick rates: Our pick rates are lower than some of the other market leaders. This is a combined result of the current order batching process, routing strategy and the lack of item-volumetric

Solution

Given these challenges ASDA decided to tackle this problem with a complex yet practical suite of machine learning, genetics and metaheuristic optimizations. The picking optimization has been successfully running for the last few months in ASDA stores.

Technology
The technology framework that supports the picking optimization landscape is built with multiple failover mechanisms that runs on multiple clusters of servers ensuring that complex computations are executed fast and service level agreements are met. A series of alerting and monitoring systems support the technology framework at crucial junctions. Multiple micro services supporting intricate store layouts, route visualizations, volumetric, shortest-paths, store profiler among others are tightly coupled with the optimization engine providing essential data for the optimization function.

The problem of picking optimization tackles 2 intertwining decision points: distributing orders among batches and determining a batch-picking route for each batch in order to minimize the total fulfilment time for all orders. The optimization engine takes a layered approach to solving the complex order batching and routing problem. The engine uses a combination of evolutionary computations and principles of genetics - starting from an initial solution and then moving on to generate better solutions using cross over and mutations. This is overlayed with machine learning approaches and meta heuristics. The layered algorithms iterate over a period of time to finally converge to a final solution - batches of orders with each batch having a planned pick route. Each batch adheres to operational constraints and other store/operational limitations.

Results

  • Increased pick speed
  • Reduced walk time
  • Higher utilization of trolleys and totes
  • Better utilization of trolleys


Any reference in this case study to any specific commercial product, process, or service, or the use of any trade, firm or corporation name is for information and convenience purposes only, and does not constitute an endorsement or recommendation by Wal-Mart Stores, Inc.

Recently viewed jobs