The traveling salesman problem poses a challenging question: What is the shortest route that visits every point exactly once? This complex issue is one that FM Logistic's warehouse operators confront daily in Poland.
Spanning the size of eight football fields and housing over 17,700 picking locations, the facility sees numerous operators on electric trucks navigating a maze of storage spots. Each unnecessary step taken by these operators can lead to increased time, fleet wear, and delayed order fulfillment.
FM Logistic, a prominent logistics provider in over 14 countries, had previously optimized their routing system. Their initial model utilized a fast, cost-focused allocation logic designed for real-time operations. While effective, it made decisions incrementally, limiting overall route coordination across the warehouse. With many operators sharing the same space, even minor improvements could lead to substantial gains.
Innovative Algorithm Development
To enhance their routing strategy, FM Logistic partnered with a new technology that autonomously generates and refines algorithms. Instead of relying on fixed rules, this approach acts as a coding partner, producing new code and iterating to discover improved solutions.
The team provided a foundational program, their existing algorithm, which made routing decisions based on immediate conditions. This established a baseline for the new system to build upon, allowing it to explore variations and improvements.
Performance Measurement
To gauge the effectiveness of each algorithm, FM Logistic created a custom evaluation function using data from 60 tours, representing over an hour of operational insights. The goal was to minimize average travel distance per pick while avoiding operational failures.
Specific penalties were implemented to discourage unworkable solutions, such as exceeding forklift capacity or violating order priorities.
Results Achieved
The newly developed routing logic produced immediate and measurable improvements:
- 10.4% increase in routing efficiency compared to the previous best solution.
- Reduction of over 15,000 kilometers in warehouse travel annually at full operational capacity.
This enhanced efficiency enables FM Logistic to manage larger order volumes without increasing staff or fleet size.
Rodolphe Bey, Group CIO at FM Logistic, remarked on the partnership's success, noting that the 10.4% improvement was achieved on an already optimized baseline, leading to quicker fulfillment and better working conditions.
Key Algorithm Improvements
The new algorithm outperformed its predecessor through a series of experiments that generated numerous candidate programs. The three main enhancements include:
- Density-based starting points: The new system identifies clusters of nearby items to establish efficient starting points for routes.
- Two-step filtering with distance simulation: A quick filter discards unsuitable orders, followed by a detailed simulation to find the most efficient path without hindering operations.
- Flexible route building: If a truck cannot be filled efficiently from a starting point, the algorithm returns those orders for better-suited routes later.
Future Directions
The pilot project in Poland has showcased the potential of advanced routing solutions in large-scale warehouse operations. FM Logistic is now considering applying these techniques to other high-volume e-commerce facilities and exploring further optimizations in road transport and product placement within warehouses.
This initiative involved collaboration between FM Logistic's team and the AI for Science team at Google Cloud, along with contributions from Google DeepMind.
