By editor 
A 15-minute delay from the point of shipping to when it arrives at the intended location is now seen not only as an inconvenience to your customers but could ultimately threaten the overall profitability of your business. Data from various industries show that the ‘last mile’ of shipping can account for more than half of all freight costs incurred by a company. Therefore, as customers increasingly ask for same-day delivery, this has created multiple challenges for traditional models of freight logistics.
The main focus of the guide will be to address how conventional systems typically utilize only static routing daily to create their schedules and therefore suffer from a lack of flexibility when faced with dynamic variables such as increased traffic or vehicle malfunction. Today’s successful companies that want to stay competitive in their markets will invest in scalable AI solutions to provide for the ability to continuously optimise routes and fulfil customer requests as quickly as possible.
The guide will provide all the necessary information for the design of AI-based systems for optimising route planning and implementing in the cloud for a streamlined & real-time operation of logistics within a business.
1. The Shift from Static VRP to Dynamic Optimization
The Vehicle Routing Problem (VRP) forms the basis of logistics thinking—the mathematical problem of determining the best route for a fleet of vehicles. VRP models are static and assume all variables remain constant.
Going forward, the Dynamic Vehicle Routing Problem (DVRP) will be the more realistic representation of how logistics organizations operate in modern society. Unlike static VRP models, DVRP represents an ongoing calculation. New factors emerge in real time as logistics operations are affected by:
- The addition and cancellation of customer orders;
- Traffic congestion and weather changes;
- Driver availability and vehicle condition.
With an increasing number of vehicles on the road, the volume of possible combinations of routes increases exponentially. For example, when fleets exceed 500 vehicles, the volume of route combinations becomes virtually limitless. Therefore, in order to scale, AI-based routing engines must transition from a process of “batch” processing to continuous optimization.
Read: The Weights on E-commerce Benefit Margins
2. Architectural Blueprint: Building for Global Scale
A high-performance AI driver is not an algorithm; it is a multi-tenant cloud-based architecture. When processing data from thousands of vehicles, an event-based architecture will be necessary.
High-Speed Data Ingestion
Your engine needs to accept many thousands of GPS telemetry pings every second. You may want to utilize technologies such as Apache Kafka or AWS Kinesis to process incoming data asynchronously and avoid bottlenecks during high traffic.
Microservices vs. Monolithic Solvers
Unified platforms are usually large. An individual, elaborate calculation can hold up the total system if the program is large. To deploy a flexible and scalable program with a high degree of reliability, it will typically require experience in logistics software development services using cloud platforms.
Using a microservices architecture allows a company to separate the computational processes (the “solver”) from the user interface (the UI) and provides for scaling of the computing resources separately, offering the opportunity to scale based on the requirements in the purchase area.
Elements of Event-Driven Architecture
A more traditional means of orienting engines has been to have a set period in which an entire set of routes or deliveries needed to be re-calculated. A more scalable solution is to implement a set of triggering “events” for re-optimizing all of the routes, where the triggers can include:
- A newly created, high-priority request.
- An instance where the dispatch vehicle has migrated away from its assigned route by more than ten percent.
- A traffic event warning predicting fifteen or more minutes of delay.
3. Algorithmic Intelligence: Heuristics vs. Machine Learning
Your engine’s brain: Pick the Slowest vs. Fastest way to make it work!!
Meta-heuristics (The Standard Way)
The Best-Case Framework is one of the most reliable standards in the industry today is that Google has developed in recent years is the [external link to OR-Tools]. Google’s OR-Tools is using meta-heuristics such as Genetic Algorithm (GA) or Tabu Search (TS), and they can provide “good enough” solutions quickly (not much time needed), making it the de facto standard for time-sensitive dispatching.
Deep Reinforcement Learning (New Technology)
From Amazon’s perspective about recent developments in Machine Learning are from the Reinforcement Learning (RL) point of view and are rapidly evolving towards this approach. Reinforcement Learning uses multiple delivery scenarios to predict complicated developments such as where/when difficult to park in the neighborhood, and how many left turns made during school hours on average people take when driving during school hours, and so on.
A Hybrid of Both Approaches: The way most world-class providers develop supporting engines today is based on a hybrid of both Meta-Heuristic and Deep Learning (DL) solutions where the DL engine provides the first “warm start” (initial solution or what the provider believes is the most likely to occur) to develop a preliminary sequence, and then the meta-heuristic solution fine-tunes and refines it before submitting it as the final solution.
4. Solving for Scalability with Geospatial Sharding
One of the biggest barriers to scalability in algorithms is known as “The Scalability Wall.” This occurs when an algorithm that has worked well to optimize a task for 50 vans will stop working completely when trying to optimize that same task for 5,000 vans.
Engineers are now able to overcome this challenge by utilizing new tools—with geospatial indexing among them—and using tools such as Uber’s H3 (hexagonal grids) and Google’s S2. By breaking a metropolitan area into a grid system, the AI can run localized optimizations on each grid in parallel and dramatically decrease computation time.
In addition, by deploying these types of engines using Kubernetes (K8s) container technology, the system is able to be “elastic.” This means that during peak dispatch hours (generally 8:00 AM) additional “solver pods” can be automatically started and added to your cloud resources, while at night during most hours (tomorrow’s low-demand hours), those additional pods can be turned off in order to reduce cloud billing costs.
5. Integrating Real-World Constraints
Distance is NOT an effective predictor of logistics performance. Time is technically the only measure of the accuracy of the supply chain! All scalable and shared supply chain engines need to address “Chaos Factors” to enhance accuracy:
- Predictive Latency: Using historical data to identify the locations where outbound loading docks have historically been delayed during the Friday afternoon rush.
- SLA Compliance: When managing your routing needs it is extremely important to identify the hard delivery windows associated with your customer commitments. Example: Must be delivered by 2:00 PM.
- Vehicle Constraints: Taking into consideration the physical restrictions of vehicle heights under low bridge structures or placement of “Green Zones” in urban areas for solely electric vehicle traffic.
- Driver Retention: Planning the routes to maximize the amount of time the driver has remaining when they reach their destination.
6. Key Performance Indicators (KPIs) for Success
When the ROI of an AI solution has been validated, there are four key performance indicators (KPIs) on which to track the changed cost structure and effect of the AI through use of the AI solution:
- On-Time Performance (OTP) — To ensure success with an AI solution, ideally OTP will reach 98% or above within the time frame that was previously promised to customers.
- Cost Per Delivery (CPD) — Generally speaking, the total cost per delivery will be reduced through AI optimization for the customer through a minimum of 10% to a maximum of 15% in saved time and fuel costs.
- Driver Utilization (DU) — The number of empty miles (known as “deadhead” miles) that drivers have to travel while going from pick up to drop off should be minimized as much as possible. The best way to track DU is to keep the average deadhead mileage at less than 50%.
- Engine Latency (EL) — For the AI solution to be effective for real-time updates provided to drivers regarding routes that were recently provided by the AI solution, the AI solution needs to return a newly optimized route in less than 500 ms.
Conclusion
To succeed in today’s world of commerce, you need to build a scalable AI engine that will help you improve the speed and reliability of your delivery system.
The shift towards an event-driven, cloud-native architecture is the first step toward creating an intelligent route optimizer that can help your organization capitalize on the unpredictability of last-mile deliveries.
Ultimately, as companies begin deploying autonomous drones and robots for deliveries, these AI engines will become the nerve center for this new era of global delivery service.
Author’s Bio:
Akshay Tyagi is a content writer at NetClubbed, an expert logistics software development company. He focuses on translating complex technical concepts into clear, actionable strategies that help brands refine their voice and achieve measurable results.