Donald Ratliff Interview: Adaptive Route Optimization

I was privileged to have the opportunity to have an extended conversation with Donald Ratliff. Donald is a titan in the world of route optimization whose insights and ideas have helped to shape the whole industry. In this and future discussions with other industry leaders, we will explore what is available in the marketplace today. We will also take a hard look at route optimization methodologies and their application in real-world scenarios.

H. Donald Ratliff (don.ratliff@deliverydynamics.com) is the founder and CEO of Delivery Dynamics Inc. He has spent more than 40 years developing and implementing new last-mile delivery concepts for enterprises including USPS, UPS, RNDC, Pepsi, and Coca-Cola. His contributions to logistics optimization have been recognized by his election to membership in the National Academy of Engineering and as a fellow of IISE, INFORMS, and AAIA. He is a Georgia Tech Regent’s Professor Emeritus and was Founding Director of the Georgia Tech Supply Chain and Logistics Institute. He also serves as Co-Executive Director of the GT Panama Logistics Innovation Center.

Delivery Dynamics (www.deliverydynamics.com) provides technology for dynamic route optimization, adaptive route optimization, and hybrid route optimization as well as an integrated platform that includes machine learning for automated continuous improvement of data together with virtual route monitoring and analytics to facilitate continuous learning and improvement by people. We work closely with our clients to help them make their last-mile delivery routes people-friendly and maximize the value that they get from route optimization.

Bill Mathews: Don, thanks for sitting down with me today. We are exploring the various styles of route optimization available in the marketplace. Let’s begin by talking to me about what Delivery Dynamics offers.

Donald Ratliff: We provide the full gamut of last-mile delivery optimization from what most people call dynamic route optimization to what we call adaptive route optimization. The basic difference between the two is that dynamic route optimization starts from scratch and generates new routes for each new set of orders. On the other end of the spectrum, adaptive route optimization maintains a set of routes where each customer has a preferred route assignment and adapts these routes for each new set of orders. The idea is not to change these assigned routes very much, but when orders are different you must often make changes. Adaptive route optimization is pretty much an automated version of the actual process used by companies that try to use fixed routes. I have seldom seen companies that truly had routes that are always the same. They have routes that they would prefer to always be the same, but invariably things happen that require the routes to be changed. Adaptive route optimization uses the same philosophy but tries to make the processes more efficient by automating and optimizing what are typically manual activities with companies using fixed or static routes.

Bill Mathews: Because it's just a perfect day scenario, right?

Donald Ratliff: Usually, but in a lot of cases it is hard to say exactly what planners were thinking when they generated their fixed routes. They have a set of fixed routes that they start with each day. Typically, they will have a Monday set, a Tuesday set, and so on. For some days the fixed routes work fine but on other days they don’t. When the fixed routes don’t work so well, planners must adapt them to the day’s orders but try to keep them as close as possible to their fixed routes. This is generally done manually, mostly because it's quite hard for a computer to do this well using traditional route optimization concepts. What we've done is develop a new methodology to automate the adapting process. The process starts with what we call assigned routes where customers are assigned to their preferred routes. These assignments define the sets of stable routes that are adapted to each day’s orders. We have also built a very good dynamic route optimizer because there are some problems where dynamic route optimization is the best strategy. However, for most last-mile delivery problems, adaptive route optimization works better than dynamic route optimization. If you want to use adaptive route optimization, what you need is enough repetition in your demand so that you can generate good and reasonably stable assignments of customers to routes.

For a long time, I thought generating optimal dynamic routes was the Holy Grail for last-mile delivery. We know now that while we can usually generate good routes, nobody can generate optimum dynamic routes, it's not possible, it will never be possible because the number of possibilities is just too big. We will never have computers big enough, never have enough time, and never can get a solution that we can truly say is optimum. I mean, some claim to generate the ideal solution but that's just not true because it's not possible to get that.

The big negative with Dynamic route optimization is that the routes that you get are what I call unfriendly. New routes are generated from scratch every day. Sometimes you get repetition but a lot of times you get something that is just dramatically different from what you had generated before. This means that drivers must servicer unfamiliar customers. Dynamic optimization focuses on minimizing travel. It assumes that the service time is fixed and that it does not change. That is obviously not true because if you send a new driver to a customer and you compare the time it takes them versus the time it takes the regular driver, the difference is often dramatic. Sometimes twice as much or more. Even when we say we're optimizing, we're not optimizing the route, we're optimizing the travel in the route. However, making the travel smaller sometimes makes the service time bigger.

Bill Mathews: Right. And then of course you're gonna have

the complaints of the drivers to the dispatchers, and the dispatchers have to listen to the drivers and therefore change the routes anyway, going through the manual processes, to make people happy.

Donald Ratliff: Yes. Another problem we have is that when we generate a set of routes that are not compact, that is if they're overlapping and tangled up, they could be very efficient in terms of our mathematical model, but they're putting the planner in a bad position. The planner is having to say, this is a good set of routes but if these routes are tangled up the planner thinks they're probably not the best set of routes, which is often right. The planner must put his or her job on the line by approving that these routes are good. If the routes don’t look good, planners start changing things manually. If we knew that the solution generated by the optimizer was the best possible, we could tell the planner that the optimizer is generating the minimum cost set of routes. However, we will never be able to say that. So not only are the planners suspicious but, they're sort of justifiably suspicious. Right? Routes being compact doesn't mean they're good but it's harder for somebody to find something obvious that's wrong with compact routes. So, the routes themselves are not, I don't know a better word than friendly. They're not friendly for the driver, they're certainly not friendly for the route planner and they also make it more difficult for managers. Managers want to have the right number of trucks and drivers available each day but if there is a radically different set of routes this Monday than last Monday it is hard to know what the right number is. Also evaluating routes that are not reasonably compact and repeatable, creates a hard-to-resolve mismatch between the math vision of good routes and the people's vision of good routes.

Bill Mathews: And that is the core of my business is to try to get the math and the people to live together.

Donald Ratliff: Sometimes it is not possible to have routes that are both efficient and compact. For example, if you have a lot of tight time windows that must be satisfied, it is often just too hard for planners to route manually, so they must rely on the computer and the computer can't do it without having jumbled up routes.

Bill Mathews: Time Windows cause the routes to look like a plate of spaghetti on the map.

Donald Ratliff: That's right. And they cause you to give up compactness and stability. Maybe, at some point, we'll figure out how to achieve stability when routes are tangled but right now, it's hard for the route planner to see what to do. If the routes are compact, and there's something the planners don't like about it, like they're not balanced, the planner will work around the edges to improve balance. The routes may not be the best set of routes, but if you can keep it compact, they won't be terrible either. The world treats route optimization like it's a math problem and it is obviously more than a math problem. You must get the math and the people involved to be complementary and on the same page. The planner is going to try to make decisions based on what they think is good for the company, but they're usually not going to recommend routes that they know will make them look bad. If they feel like the optimized routes make them look bad, they will not accept the result.

Bill Mathews: So how can we help companies automate this process more?

Donald Ratliff: At one time we thought we could automate everything. I guess it may be possible but probably not in our lifetimes. The most difficult case to automate is when there are not enough resources to execute all deliveries. This almost always requires a human planner to decide things like whether a supervisor can take a route or whether some deliveries will be postponed. Creating and maintaining rules for the computer to make these decisions is extremely difficult. So, it seems unlikely to me that we can ever replace planners altogether. I think what we can do is automate the easy decisions and make it a lot easier for planners to do a good job. This will take much less planner time for daily route planning and allow them more time for continuous improvement.

Another question is what is the purpose of the automation? I see a lot of the marketing talk about big cost savings numbers like 30% from reducing trucks and drivers using dynamic route optimization but I don't know how this is calculated. When workload is not balanced over days, minimizing the number of routes will result in fewer trucks needed on some days but what can you do with the trucks you don't use? If you need 50 trucks on Monday and 45 on Tuesday, you only get the savings if you have something you can do with the extra 5 on Tuesday. The problem often is how to balance work over a fixed set of trucks as opposed to minimizing the number of trucks. If you have 50 trucks and drivers available today, you often prefer to use 50 trucks and drivers with roughly the same workload rather than packing 45 of the trucks and leaving 5 trucks and drivers idle.

Bill Mathews: Yeah, it always comes down to that. Especially when you consider the driver shortage that exists, right? On one hand. It'd be nice to take those 50 and make them 45 regularly. If you saw that you can do that, that'd be great. But often. What happens is exactly what you describe, right? You can get by with 45 many days but the peaks require 50.

Donald Ratliff: Yes. And if you're going to get rid of five, you need to see a lot of evidence over quite a long time before you start cutting back on trucks and drivers because you must predict what will be your heaviest day in the future. In fact, one of the things I like about what you called “geographic routes” is that if the routes have this stability and compactness, it makes it easier to predict truck and driver demand and to do any sort of improvement. If you have compactness and you see that you can't satisfy time windows, then you have some good ideas about where you can go and negotiate the windows. If you're doing dynamic route optimization, the routes are going to automatically change to fit the time windows. This makes it very difficult to determine where negotiating different time windows will significantly improve efficiency. If routes are more stable, then planners have better insights regarding changing the time windows. Often there is more fuzziness in the time windows than the optimization thinks.

Bill Mathews: That's one thing. The planner always has the capability to do that. An algorithm is never going to be able to see one stop causing all sorts of problems because it's got a short time window at a bad time of day. They can pick up the phone, call that customer, and say I need to make it this instead. And can you live with that? And, you know, the answer 90% of the time is sure, I just need to know when you're gonna be here.

Donald Ratliff: But if your routes are different every time you can't tell that's your problem, right?

Bill Mathews: Right. So, who would you consider to be the best target customer for adaptive route optimization?

Donald Ratliff: I would say anybody who assigns regular delivery days to their customers. If you assign delivery days to your customers, then you force a repetition pattern. Right? The customers delivered each Monday will be a subset of those who have Monday as one of their assigned delivery days. In addition to improving daily route efficiency, routes that are more stable and compact provide better insights regarding how assigned delivery days impact delivery costs. This is particularly important when customer visit cycles are less than once a week. Making day assignments that balance the requirements for delivery resources over days and group customers that are near each other is an extremely difficult problem so planner insights regarding how they can be improved can make a big difference.

Bill Mathews: Right now, we're talking about a whole different kind of optimization at this point, the optimized visit schedule.

Donald Ratliff: Yes, and I think making day assignments is the most difficult optimization problem related to last-mile delivery. For example, if I have a remote location where have multiple customers, I'd like to serve all of them on the same days, but I also have limits on route times and truck capacities so making good assignments becomes terribly hard, particularly as the number of possible delivery patterns increases.

Bill Mathews: Right and trying to balance that. So you get the correct number of trucks, you know, make that as even as you can across the days.

Donald Ratliff: Yes.

Bill Mathews: That's a fun problem.

Donald Ratliff: Yes, but anytime you introduce more decisions, like additional day assignment patterns, optimization becomes more complex. The additional decisions add flexibility but usually make it much more difficult to take advantage of this flexibility. A good example is adding additional truck types. This can be a good thing if you know or can determine over time where specific truck types provide an advantage. However, the number of possible routes is dramatically increased so using dynamic route optimization to determine where to use each truck type can result in very poor routes. Adaptive route optimization is useful if you know where to use each truck type.

Bill Mathews: If you have a downtown truck, for example, I mean that makes sense.

Donald Ratliff: Right. This is a case where planners can often do better than technology in determining where to use each truck type.

Bill Mathews: Yeah.

Donald Ratliff: All right, let's consider where to apply the truck types. I go back to my term “people friendly”.

Bill Mathews: It's a great term.

Donald Ratliff: It's about building routes that are stable, compact, and balanced. Everything including where you apply the trucks is easier. If your routes are similar every Monday, then the planner can see that Route 1 is full every Monday and Route 2 is not full every Monday, so he or she can put a big truck on Route 1 and put a little truck on Route 2. These are the kinds of decisions where continuous improvement can have a big impact. If I were going to just be in consulting, continuous improvement for last-mile delivery would be my theme. The only way that route optimization works is to continuously improve planner knowledge and planning data. With last-mile delivery, the problems are very difficult and don't stay the same. Dropping an optimization system into a delivery operation and then just hoping it will get better is just a very bad idea.

Bill Mathews: That's exactly what happens 90% of the time though, as you know.

Donald Ratliff: Yes, yes. I blame those of us that sell route optimization technology for a lot of this. We sometimes get caught up in the idea that optimization is going to be easy and forget that the data must be good, and the routes produced must help the drivers, planners, and managers do their jobs better. All of that takes work. I think that we need more discussions, where people can understand that the right kind of optimization is a good thing but if you don't work at it, you can’t expect anything good to happen. You're going to get bitten in the behind.

Bill Mathews: You'll end up with a glorified whiteboard. You might as well use strings and a wall map. I think we can agree then that there is a spectrum with fully dynamic on one end and completely static on the other. Nearly every business problem falls somewhere in the middle of that spectrum, toward one side or the other. I think one of the problems I’ve seen is that the imperative to sell optimization solutions can sometimes push businesses off their ideal spot.

Donald Ratliff: Yes. It's also the case that until recently, automation was almost entirely focused on dynamic route optimization. Adaptive route optimization provides automation but uses the concept of assigned routes to assure better route stability and compactness. However, keeping the optimized routes in sync with the business needs requires that these assigned routes be maintained. If you let your assigned routes get out of sync, you're going to adapt to bad routes, which is obviously not going to be a great strategy. We spend a lot of time developing machine learning to automatically do the updating of assigned routes. We automatically keep up with what is done each day to adapt the routes. And if you regularly make the same changes, the computer can determine what changes should be made to the assigned routes to make them more stable.

Bill Mathews: That's something we didn't have 10 years ago for sure.

Donald Ratliff: Couldn't do that, right?

Bill Mathews: Right.

Donald Ratliff: The other thing is that if you're going to use machine learning, you've got to have some consistency. It doesn't work well if you adapt differently every time. Machine learning needs a tremendous amount of data when there is a lot of variability. Most of the time adaptive route optimization is successful and reasonably consistent in adapting the assigned routes to the daily orders. However, when planners adapt the routes, they must also be reasonably consistent in how they adapt for the machine learning to work well.

Bill Mathews: Don, it’s always a pleasure to talk to you. I really appreciate your experience and perspective in the route optimization world, and I believe the readers will be well served by learning from it.

Donald Ratliff: Thank you, Bill. I’m glad we’ve had the opportunity to dig into a subject that is a passion for both of us.