00:00:07 Introduction and Cédric Hervet’s background at Kardinal.
00:02:07 Kardinal’s approach to real-time route optimization with human input.
00:03:41 Real-time route optimization’s impact on supply chain and inventory management.
00:05:32 Development of route optimization algorithms and data importance.
00:06:22 Evolution of route optimization and the importance of accurate data.
00:08:00 Key companies and players in route optimization.
00:09:58 How Google’s innovations have inspired other companies.
00:10:51 Main sources of data for Kardinal’s route optimization.
00:12:55 Technical challenges of real-time online solutions.
00:15:38 Users gaining control over data and its impact on optimization.
00:18:00 Challenges in balancing data control and human expertise.
00:19:30 The impact of large companies like Amazon, Google, and Microsoft on data reliance.
00:21:00 The concentration of the map data market.
00:22:17 Exciting research and developments in AI and their potential applications.
In this interview, Kieran Chandler speaks with Joannes Vermorel and Cédric Hervet, founders of Lokad and Kardinal, respectively. They discuss the challenges of real-time route optimization and the importance of human input in conjunction with advanced technology. Hervet also shares his excitement about developments in AI, including reinforcement learning and the potential implications of quantum computing. The conversation touches on the idea of map data as a common good and the reliance on large tech companies for data, as well as the need to remain at the forefront of emerging technologies.
Kieran Chandler hosts a discussion with Joannes Vermorel, founder of Lokad, a supply chain optimization software company, and Cédric Hervet, Co-founder and Head of R&D at Kardinal, a route optimization software company. They discuss the recent advances in crowd-sourced data, quantum computing, real-time route optimization, and the role of humans in these processes.
Cédric Hervet explains that Kardinal specializes in real-time route optimization with a focus on context awareness. Traditional route planning is usually done manually, which is suboptimal. Humans, however, have the ability to handle emergencies and make decisions based on a global scope of priorities. Current route optimization software on the market provides static solutions, which become problematic when unexpected events occur, like traffic congestion or rescheduling.
Kardinal’s approach is to continuously optimize routes, allowing greater capacity to handle problems as they arise. They also emphasize not removing humans from the equation, as they possess knowledge and strategic vision that cannot be found or modeled in databases.
Joannes Vermorel agrees that utilizing human intelligence in conjunction with modern computing power is essential for supply chain optimization. Lokad’s strategy is to make the most of smart people who are very aware of the problems they are trying to address.
The conversation turns to the differences in time scale between Kardinal and Lokad’s approaches. Kardinal focuses on real-time route optimization, with decisions being reevaluated every minute or so. This is not the same as microsecond-level decisions required for tasks like piloting robots in a warehouse. On the other hand, Lokad’s decisions focus on the following day or up to one year ahead.
The conversation then moves on to the development of route optimization over the years. Hervet differentiates between the problems of finding the best route from one point to another, which Google Maps is designed for, and the more complex issue of determining the optimal order to visit multiple stops. This latter problem requires sophisticated algorithms and accurate traffic data to provide feasible and efficient routes. Kardinal focuses on transforming theoretical mathematics into practical solutions, ensuring that the optimized routes are realistic and manageable for drivers.
Vermorel highlights Google as a company that has driven significant innovation in online solutions, particularly in search engines, by providing more up-to-date information compared to competitors at the time. While not directly using Google’s algorithms, the approach serves as an inspiration for companies like Lokad and Kardinal, as they work on scalable, online solutions for complex problems.
Discussing the constraints and nonlinearities in route optimization, Vermorel points out that factors like employment regulations and driver-specific constraints add to the complexity of the problem. Hervet adds that there are two main sources of data for Kardinal. The first is from clients, who provide order information, constraints, driver availability, vehicle capacities, and other relevant details. The second source is from technological partners like HERE Technologies, which supplies distance data, traffic patterns, and real-time updates necessary for route optimization.
They discuss the challenges of real-time data processing, the importance of human input, and the reliance on large tech companies for data.
Vermorel explains that working with real-time data presents multiple challenges. For one, the speed of light is finite, which means that even though data can be transmitted quickly, it can still take seconds to process when there are multiple data centers and thousands of round trips involved. Additionally, there are many factors that can slow down computer systems, such as software updates or other background processes. Ensuring that real-time systems operate efficiently on a global scale requires significant expertise.
Another challenge is the dependency on partners, which can affect the availability and reliability of enterprise software services. The more dependencies there are, the more potential problems and downtime can arise. This means that the service’s uptime will only be as good as its dependencies, often resulting in lower availability and reliability.
Hervet highlights the importance of maintaining human input in route optimization systems. He shares a story about how their initial algorithm-generated routes were mathematically optimal, but the drivers were able to identify issues that the algorithm could not see. For example, a driver might know that parking would be impossible during a specific time due to parents picking up their children from school. Hervet emphasizes the need for a balance between algorithm-generated routes and human expertise to make the best possible decisions.
He also believes that having control over data is crucial. Users need to be able to understand and interact with data to make informed decisions. Kardinal aims to augment human decision-making with computational insights, allowing for a combination of human expertise and data-driven optimization.
When discussing the issue of relying on large companies like Amazon, Google, and Microsoft for data, Hervet agrees that there might be too much reliance on them. However, he also acknowledges that technology is fast-moving and there are not many providers of map data worldwide.
The conversation begins with a question about whether map data should be a common good. Vermorel acknowledges the importance of open maps and suggests that technological advancements could push mapping closer to being a common good. However, he also points out that the mapping industry is highly concentrated, with only a few key players. While competition exists, it remains limited.
Hervet then discusses his excitement about developments in operations research and artificial intelligence. At Kardinal, they have PhD students working on solving online optimization problems, expanding the scope of the mathematical field. They also consider other advancements in AI, such as reinforcement learning, which teaches algorithms to make decisions without explicitly defining the best choice. Hervet notes that this approach is philosophically different from Kardinal’s current methods, which involve defining a solution space and ranking solutions within that space.
Although reinforcement learning has shown promise in various applications, Hervet admits it has limitations when dealing with the range of constraints they face at Kardinal. However, they continue to monitor its progress, as it could become more suitable for real-time decision-making in the future.
The conversation then turns to the potential implications of quantum computing. Hervet mentions Google’s recent claim of achieving quantum supremacy, meaning that quantum computers can solve problems in significantly less time than classical computers. Quantum algorithms could be used to solve complex problems like the Traveling Salesman problem, which is central to Kardinal’s work.
While quantum computing is still a long-term prospect, Hervet acknowledges its potential to democratize problem-solving by making difficult problems easier to solve. If this happens, companies like Kardinal would need to stay at the forefront of the technology to help their clients perform better.
Kieran Chandler: This week on Lokad TV, we’re delighted to be joined by Cédric Hervet, who’s going to discuss with us how the increases in quantum computing and the ability to route optimize in real-time has led to delivery companies changing the way they actually operate. So, Cedric, many thanks for joining us today.
Cédric Hervet: Thank you for welcoming me. Kardinal is a company that specializes in route optimization in real-time with a great deal of context awareness. Usually, routes are optimized by humans manually, so they plan schedules for their drivers or technicians. This is clearly suboptimal, but on the other hand, humans have a great capacity to handle emergencies, unforeseen events, and problems. They can envision the global scope of their priorities and make decisions. However, algorithms are not fully equipped for that, especially optimization algorithms. There are software programs on the market that provide route optimization, but they do it in a very static way. They get data, mix it somehow, and provide an optimized solution for schedules. But this is problematic because the first event encountered will destroy the quality of the routes. Once trucks are on the road, there are issues like traffic congestion, late or absent customers, and rescheduling appointments. All these events can damage the performance. At Kardinal, we believe the right way to optimize routes is to never stop optimizing them. This way, you have a greater capacity to handle problems as they arise. Another key aspect of what we do is that we have a strong focus on not removing humans from the equation because they know their job and they know things that can’t be found or modeled in any database. It’s important not to remove them from the process because they have a strategic vision of their overall activity.
Kieran Chandler: This idea of using a human brain and making the most of that, particularly in those emergency scenarios, is a really interesting one. I think it’s something you’d probably agree with as well, Joannes, making the best use of the human brain in terms of being an addition to optimization?
Joannes Vermorel: Absolutely, the idea of making the most of smart people who are very aware of the problems they’re trying to address with their supply chain, and using the best of what modern computing power has to offer is, at a very high level, also the strategy of Lokad.
Kieran Chandler: Great. Today, we’re talking a bit about real-time route optimization. Why is that of interest to you from a supply chain perspective?
Joannes Vermorel: Obviously, Lokad, when we think about supply chain optimization, we don’t exactly think at the same time scale. If I compare what Kardinaland Lokad, who are both enterprise software vendors, are doing: Kardinalis doing route optimization, so decisions that can be re-challenged every minute or so. It’s not exactly like microseconds, as you’re not piloting real-time robots doing picking in a warehouse. It needs to be swift, but not at the microsecond level. On the contrary, Lokad focuses on decisions for tomorrow or up to one year ahead, so that’s the time range. The fact that Lokad can optimize supply chains typically involves making decisions for longer time periods.
Kieran Chandler: We are discussing inventory rebalancing between locations, such as stores or warehouses. This heavily depends on the agility of tools like route optimization, as provided by companies like Kardinal. The more agility you have with your routes, the easier it is to rebalance stock between stores, which lowers the cost for optimization. How has route optimization developed over the last few years, especially given our increasing reliance on smartphones and GPS systems?
Cédric Hervet: Route optimization has two main problems. The first is going from one point to another and finding the right road, which is what tools like Google Maps are built for. The second, more difficult problem is having n stops to visit and determining the order in which you will visit all these stops, taking into account factors like traffic. Our focus is on this second problem. To solve it efficiently, you need smarter algorithms than just enumerating all possible combinations, which is where the math comes in.
The ability to model this problem and develop algorithms has been around since the 1960s. However, the actual implementation of these algorithms relies heavily on the availability of accurate data, such as traffic data. If the data given to the algorithm is wrong or inaccurate, it will produce infeasible routes. At Kardinal, our focus is on providing routes that are feasible and practical for drivers to follow.
Kieran Chandler: Speaking of data availability, who are the key players that have driven the growth and expertise in route optimization?
Joannes Vermorel: There has been intense development in online solutions for various problems. Historically, one company that made a significant impact in this area was Google. Before Google, search engines like Yahoo and AltaVista updated their indexes once per quarter, resulting in outdated search results. Google was innovative in many ways, including its ability to provide more up-to-date search results.
Kieran Chandler: So, Joannes, can you tell us about the transition that you made from being a search engine optimization company to a supply chain optimization company?
Joannes Vermorel: We initially started with online solutions to provide the best results for queries. However, the reality is that new pages are constantly added to the index, and initially, we were only doing a weekly refresh. But, it was already 20 times faster than most of our competition. So, there was a transition towards a problem where we wanted always up-to-date results under changing conditions. Low Cad and Kardinaldon’t use Google algorithms specifically designed for search engines, but it was a source of inspiration to us, to see what you can do at scale with the proof that it can actually work.
Cédric Hervet: And, many other players started to do similar things on different types of problems. I think there was a new wave of people who were thinking about how to have the online version of a problem that is much smarter and also very different from what Kardinal is doing right now compared to what people were doing in the 50s. All those constraints and nonlinearities make optimization hard to represent. You have nonlinear constraints, such as the fact that maybe your driver cannot drive more than X hours because there is an employment regulation that says so.
Kieran Chandler: Cédric, can you tell us about the data that’s actually of interest to Kardinaland where you get it from?
Cédric Hervet: Sure. There are two main sources of data. The first one is obviously from our clients, who provide us with the orders we have to optimize. They give us the most accurate description of their activity, such as legal constraints for working hours of the drivers, drivers’ availability, where they start, where they take the service, what kind of vehicle they are driving, what kind of capacity is needed, and whether they can transport dangerous goods or conduct specific technical interventions that require a particular skill set. All this data defines constraints over their activity, and we need to understand it. The second source is the data coming from the client, describing the orders themselves, such as packages to deliver or interventions like repairing IT equipment. We rely on technological partners like HERE Technology, which is our partner for getting the distance data we need to understand how long it takes to go from one stop to another, and how traffic changes over time. We also need to get real-time traffic updates to adapt as needed. HERE provides us with this data, and we use our algorithms to provide updated solutions.
Kieran Chandler: Cédric, you mentioned the rising growth of online solutions. From a technical perspective, what challenges does that introduce in terms of being able to work in real-time?
Cédric Hervet: Real-time introduces a lot of complications. Firstly, there is no such thing as real-time because the speed of light is finite. Even though it’s incredibly fast, it still takes time. The problem arises when you have distributed computer systems, and you need to go back and forth from multiple data centers. If you do thousands of round trips, it takes seconds to get results.
Kieran Chandler: Achieving real-time systems can be quite challenging when you operate globally. What are some of the difficulties that you face?
Joannes Vermorel: Well, there are many things that can prevent you from having a good real-time system. For instance, our computers may seem super fast on average, but there are times when they get stuck because of updates or other reasons. So, the reality is that computers can actually be pretty slow in the worst case. Also, the speed of your system will typically be whatever is the slowest that you have. This means that if you have many machines, the slowest one can be very slow. Real-time is in itself a set of challenges that are very complicated. Another complication is that introducing dependency to partners means that you have to make your service very available and reliable, even if your partners are not. The more dependency you have, the more potential problems there are for downtimes. Your service is only as good as your dependencies, which means that every time you move down the chain, you get something with lower availability, lower uptime, and lower everything. So, real-time is quite a challenge.
Cédric Hervet: Yes, I agree. And we’re now entering an era where we can actually have control of that data. For example, with Waze, you can now say if there’s a police speed camera somewhere. Do you think that’s a positive impact? The fact that we can now control these pieces of data is obviously very important to have that capacity, especially in the context that I was describing before. At Kardinal, we are paying very much attention to keeping humans on board the system because once they lose control over it, and everything is too automated, they can’t really check that the algorithm is doing something. They lose so much understanding of what is happening that they can’t really provide their expertise. And they always have an expertise. I have this short story of when we started to do what we’re doing, we were trying to challenge the routes of drivers, and we were proposing our own optimized tours. They always had an example of something that the algorithm could not see. A funny example is that we had this very beautifully optimized route, which was obviously the perfect way of visiting all those stops. But when the driver saw this, he was not focusing on the general aspect of the route, which was kind of better than what he would have done anyway. He was really focusing on some specific stops. He was telling us, “Okay, so you’re telling me that I will deliver this person here at 4:45 p.m., and here there is a school, and I know that every parent will be parking in that street, and I won’t be able to park myself to just make that delivery.” This is obviously mathematically optimal, but I know that this delivery at this specific time means that it’s just 15 minutes, but it’s impossible to deliver someone in this 15 minutes. And this is really something. For those working in data, the cost of knowing that in advance to avoid that space, fix up at 50 minutes in advance anticipating that fact is very costly for us. It’s pointless because we have someone in the truck knowing that. The key interaction we are trying to implement with them is okay. You know stuff that we will never know, and we will never really try to know that because it comes out to too great cost for us, so just give us that input. Okay, you can challenge the algorithms even when you’re on the road. And what we’re trying to do is that many events can happen. It can be problems from the…
Kieran Chandler: Could you explain to our listeners how you combine user input and data to make informed decisions?
Joannes Vermorel: Clients are not the only source of data. The user’s input is also an event for us, and if you think something is really better, you can choose to do otherwise. What we suggest is just a recommendation, and you can make an informed decision. You’ll really be the master of your domain because you probably know stuff that we don’t. To really answer your question, it’s essential to have control over the data because data, in itself, is pointless if you don’t have something telling you what this data actually means. But once you can provide insight into what the data means and what the decisions you’re providing are, humans are not removed, but they’re augmented. They can take better decisions because they have information about the impact of our computations, and with the other things they have in mind, they can make the best possible decision. This is what we are trying to achieve - the right combination between the two.
Cédric Hervet: Another example of this is the wildfires in the US. When people were putting in their routes to try and escape these wildfires, obviously the roads where there were fires were showing them as clear, and they were actually directing them to go that way. Having a way of adjusting that and taking into account factors out of context is essential when you’re doing what we’re doing.
Kieran Chandler: Would you say from a data perspective that we’re too reliant on some of the large companies like Amazon, Google, and Microsoft?
Joannes Vermorel: I would say probably yes, but also the fact that technology is really fast-moving. If you look at map data, there are not that many providers worldwide. The question is whether map data should be a common good. There are some people who are trying to do that with open maps and whatnot. The reality is that when you have technology that is super fast-moving, it’s hard for many companies to compete. Usually, when people say “winner takes all,” what they forget in technology is that frequently things rotate swiftly. So yes, there is not that many players in the market for maps right now, but I see plenty of changes where things that were considered as very hard to access are now getting closer to be more like common goods with open maps and whatnot. It will take a long time, but what I suspect is that most of the time, those things become commodities, and the problem will have moved to something else entirely. So, bottom line is probably right now, as long as there is some alpha competition, I believe that there is some competition. For maps, there are probably like four or five players, and that qualifies, but indeed, it’s still a fairly concentrated market.
Kieran Chandler: And if it’s not sort of wrapping things up today, you’re very involved in research and development. What real things from a research and development perspective excite you over the next couple of years?
Cédric Hervet: Well, first, there are what we’re doing at Kardinal.
Kieran Chandler: So, Joannes, what recent advances in optimization excite you the most?
Joannes Vermorel: We are PhD students working on solving the online version of optimization, expanding the scope of operations research as a mathematical field to handle problems in the proper way. We’ve also seen other things happening in the AI community as a whole. Reinforcement learning is a different approach from what we are doing with operations research. It’s teaching algorithms to know the best decision without explicitly telling them what possible decision is right, which is very different philosophically from what we are doing at Kardinal. We tell the algorithm the global scope of possible solutions to a problem and which is a better solution than another, so we can focus on finding the solution inside a closed envelope of possible solutions. Reinforcement learning provides another way of doing it and is probably very suited for real-time decision making.
Cédric Hervet: The limitation of this approach today is that it cannot handle the variety of constraints we end up with our techniques. But who knows, we’ve been very much surprised with what reinforcement learning could do in Pingo or even video games. Now, they can beat very strong players. This is something we are following, and it’s really prospective. But we’ve seen Google announcing that there’s quantum supremacy, which means having quantum computers solving problems in short times that are unavailable for normal computers. They had to enumerate all the solutions, and we know that there are quantum algorithms suited for solving the Traveling Salesman problem, for example, which is one of our core problems. They can solve it in seconds, whereas it takes thousands of years for singular computers just to enumerate. This is something we need to follow. Obviously, this is very long-term, and with our algorithms being smart in their conception, we can already match the speed of quantum conundrum algorithms. But quantum computing yields a kind of democratization of making all these problems that are hard by nature easy to solve, which is quite something interesting for us. If our problems become easy to solve tomorrow, we will have to be on the front of handling this technology to help our clients perform much better than today.
Kieran Chandler: Brilliant. Well, thanks for your time today anyway. It’s really interesting. Thank you very much. So, something for this week. Thanks very much for tuning in, and we’ll see you again next time. Bye for now.