The Paradox of AI in Last Mile Solutions

Ron Dombrowski October 8, 2019

As we move further into the 21st century, we see more and more AI around us. My car tells me when I am leaving my lane and will brake if the car in front of me stops. So, why does finding a cost effective solution for routing pickups and deliveries take so long?

To start, we need to look at what AI does.

Artificial intelligence has a machine learning (ML) component. That is, there is a process where the AI tool can be taught by observing and then can retain the knowledge and use it later. Show an AI learning tool enough pictures of cats, and it will be able to recognize a cat. Show it enough pictures of lane markings and it will be able to tell you when you are leaving your lane.

What is being learned is "what does a cat look like" or "what does a lane look like". But as tricky as this might seem to be, at least cats look like cats and lanes look like lanes.

When attempting to find an efficient set of routes for a given day’s work, it’s harder than that.

Ask a route planner what "good" routes look like and you will probably to told that good routes are loops that start and end at some location and never cross. A lot like petals on a flower.

But you will also be told that this easy to observe pattern is not easy to find in real life. You can’t actually drive these attractive loops that can be shown on a map. You must drive on the roads that have you have been given, not those that you would like to have. Some roads are faster than others and, in many cases, there is no direct road from one place to another. There will also be cases where multiple routes use the same roads, and cases that there will be some route overlap as well. Once you are on the fast shared road, it will really make no difference whether you get to your next stop by taking the exit to left or the one to the right. The pretty loops disappear quite quickly.

Could a good machine learning algorithm figure out this pattern? Probably.

But let’s consider that each day’s problem is different. There will be days when two stops across the street from each other have very different delivery time windows. If one requires an early morning delivery and the other requires one for late in the afternoon, the vehicle is not going to park and wait. Instead it will have to come back later, or another vehicle will have to be scheduled for that later delivery. And let’s add the complexity that you have a limited number of vehicles with, say, lift gates, and some small number of customers require delivery from one of them.

Since the solution is not going to be just straightforward pattern matching, some cost minimizing algorithms use an AI approach of starting with a random solution and then use machine learning to find better and better solutions over a period of time. And since it takes time to learn, it will take longer to find a set of efficient routes than it will to recognize a cat or a lane departure.

Also, note that before any of this assignment of work to routes can be done, there is another preliminary step that needs to be performed which also can take a significant amount of time.

The assignment algorithm needs to know how far apart the stops are. The time and distance from any one point to another is critical. Both time and distance have costs, and the most efficient solution is the one that minimizes these costs.

The algorithm for finding these times and distances for many points is way more complicated than finding the time and distance from single origin and destination points that you feed into the routing app on your phone or computer. It takes a significant amount of time to get accurate numbers here.

Are there shortcuts to doing this? Yes, there are. But there are also cases where the algorithms using these shortcuts do not give accurate times and distances, especially for stops that are only a short distance apart.

So, the paradox here is that we have great artificial intelligence to match patterns and help make decisions, but when given very different last mile logistics problems each day, the way to get the best set of routes becomes more than just pattern matching and there is just a lot of computation that needs to be done.

Effective route assignment does take some time, but it takes a lot less than if you try to do it by hand, and by not doing it by hand you will the benefits of getting very efficient routes using AI.

Visit us at Strategic Movements and see how and see how we use AI to produce efficient routes for a wide range of last mile scenarios.

Like the blog? Sign up and be the first to hear what we have to say.