Amazon, the global e-commerce giant, has always been at the forefront of innovation, especially regarding its fulfillment logistics. In this case study, we cover three main issues Amazon faced on the way to fulfillment logistics optimization:
Amazon chose to use AnyLogic because it is compatible with various data sources and provides AI integration, allowing for innovative solutions for comprehensive fulfillment analytics.
For each issue, a unique approach and model for problem-solving were employed. Below, we will delve into these, highlighting the problematic aspects, the applied solutions, and the outcomes.
Issue 1: Facility location
Problem
Amazon needed to determine the best places to set up grocery stores to meet customer demand throughout the year. The simulation had to help find new potential locations and evaluate the effectiveness of current stores.
The main goal was to optimize fulfillment logistics, keep travel time as short as possible, and ensure that groceries could be delivered in less than one hour.
Solution
The team combined deep reinforcement learning with the simulation model they built in AnyLogic. The simulation defined an optimal number of stores and specified the locations they should serve. The DRL-based AI agent enabled Amazon to see more effective fulfillment logistics options.
Results
The AI agent outperformed the previous heuristic algorithm by 38%, significantly reducing the average travel time between the store and the customer from around 17 minutes to approximately 10 minutes.
Issue 2: Network fulfillment logistics and topology
Problem
The Amazon team wanted to plan a strategy for placing fast-selling products closer to customers. At the stage of building a model of fulfillment logistics organization, they observed that 23% of goods from far away fulfillment centers (FCs) ate up the overall network efficiency and speed.
Solution
The chosen geo-cluster for the experiment on fulfillment logistics
The Amazon team set three directions for the work to succeed in the task:
- Divide the network into regional clusters.
- Classify products and FCs based on local demand.
- Target network topology improvements.
The team divided the entire US into multiple geo-clusters, making them self-sustaining for fulfillment logistics. Also, they categorized the FCs into high- and low-velocity centers that manage high- and low-velocity products, respectively. To test the approach, they built a simulation model focused on one region – South California.
Results
The proposed network reduced the miles traveled per package to about 48%. Consequently, the total travel distance was reduced, and the demand was met. Following the results, Amazon decided to test the geo-cluster approach on other parts of the fulfillment logistics network in the future.
The precise modeling of truck movements on real roads significantly enhanced the simulation's detail. For an in-depth understanding of incorporating historical traffic data in AnyLogic, visit the Road Traffic Simulation page.
Issue 3: Last-mile route planning
Problem
The team faced the challenge of creating a delivery plan to minimize travel distances and shorten delivery times without adding extra resources.
Solution
For the last-mile route planning issue, Amazon built an optimization model to test the alternatives for fulfillment logistics. The simulation aimed to improve route sequences from delivery stations to customer doors. The model processed three types of information:
- The number of vans Amazon needed for a day.
- Package distribution between vans.
- The best combinations of packages to be delivered to the neighboring areas.
The team developed an optimization model and integrated it into the delivery simulation. This allowed the simulation to pause and wait for the route sequence of each van to be optimized. Once the model provided the optimized results, the simulation would continue. The optimization process for each route takes approximately 20 seconds.
Results
Using the provided data, Amazon received proof that optimized route planning can decrease overall mileage by 9%, improving last-mile delivery for thousands of packages each day.
Siva Veluchamy of Amazon presented the case study at the AnyLogic Conference 2023. If you are interested in other Amazon cases, check out the Simulation for Transportation Network Optimization via Truck Yard Revision case study. It is about Amazon's solution for reducing delivery delays and transportation network optimization developed with the AnyLogic simulation model.
The slides this case are available as a PDF.
