Summary
Murray Thom, VP of Product Management at D-Wave, Sam Lorenzo, CEO at InspirationQ, and Sergio Gago, Managing Director – Quantum Computing at Moody’s Analytics, are interviewed by Yuval Boger. Murray, Sam, Sergio, and Yuval discuss quantum and non-quantum approaches to solving the Travelling Salesperson Problem, benchmarking for TSP, how companies should approach optimization problems, and much more.
Key takeaways
- Benchmarking the Travelling Salesperson Problem fairly means comparing quantum annealing against the best classical solvers available, not weak baselines built to make quantum look good.
- D-Wave's hybrid solver combines quantum annealing with classical heuristics, and the panel discusses where that combination actually helps versus where classical solvers alone still win.
- Moody's approaches optimization problems from a business risk perspective, asking whether a quantum or hybrid method delivers measurable value before adopting it, not just whether it runs.
- InspirationQ's role is to build neutral, reproducible benchmark methodology so companies can judge optimization claims on the same terms across vendors and problem sizes.
Read the full transcript on the Quantum Computing Report site
Transcript
Yuval Boger: Hello, Murray. Hello, Sam. Hello, Sergio. Thank you so much for joining me today.
Sergio Gago: Hey, Yuval. Hey. Good morning. Good afternoon. Thank you. Nice to be here.
Murray Thom: Nice to be here.
Yuval: So, Murray, who are you and what do you do?
Murray: My name is Murray Thom. I'm the Vice President of Product Management at D-Wave Systems.
D-Wave is the leading provider of quantum computing systems, software, and services. We're producing two models of quantum computing systems, and we're laser-focused on bringing people business value in their practical quantum hybrid applications.
Yuval: And Sam, who are you and what do you do?
Samuel Fernández Lorenzo: Yeah.
My name is Samuel Fernández Lorenzo. My background is in physics. I have a PhD in quantum technologies. Previously, I led the quantum computing research in finance for BBVA, and now I'm co-founder and CEO of Inspiration-Q. Inspiration-Q is a spinoff of the CSIC, the largest public research institution in Spain, and our mission is essentially to offer quantum advantages in the short term.
So we develop our own quantum-inspired and quantum algorithms, as well as specialized solutions, in the finance industry at the moment. But with a very pragmatic vision, because we prefer to work only with those backends that may provide short-term advantages.
Yuval: Excellent. And Sergio, who are you and what do you do?
Sergio: So, thanks, Yuval, for having me here today. My name's Sergio. I am the Managing Director of Quantum Computing at Moody's, and I'm also the editor of the Quantum Pirates newsletter, sharing with everyone the news of what happened in the week in quantum. And I am a lover and longtime partner of the traveling salesman problem, which I think is what brought us here today.
Yuval: Excellent.
I should have asked you who are you and what do you not do, because you do so many things. But let's talk about the TSP, the traveling salesperson problem, which is one famous kind of optimization problem. Murray, would you explain to us what TSP is about and why it is so popular and important?
Murray: Sure. Well, let me take a shot at it, and Sergio, being the fan that he is, can probably improve on the answer that I give. The traveling salesman problem is a very common problem that's well known. It's about the notion of someone who needs to go out—a salesperson—who is going to visit a number of customer sites, and they need to find the optimal route to each of those customer sites so they can reduce the distance traveled.
So if you imagine they have customer sites A, B, C, and D, should they do them in that order, or should they start with B, then go to A, and then go to C and D? Basically, if you have a small number of sites, you can set up all the combinations, look at them, and say, "Okay, which one is the shortest?"
But as you add new sites to that problem, the problem complexity grows exponentially. And so it's a fantastic problem from the perspective that it's an example of an NP-hard optimization problem, where you've got this huge solution space that you're trying to look through. There's a variety of solutions.
On average, solutions are not particularly good. With quite a bit of work, you can improve solutions and get good operating solutions. And then you have to do an enormous amount of work to try to find optimal solutions, and more work still if you want to prove that those solutions are optimal, because the solution spaces are so large.
So in my experience at D-Wave, where we've been helping people tackle their industrial optimization problems, scheduling problems, and logistics, the traveling salesman is a really helpful example to raise with people so they can make a connection to the problem types that are suitable for quantum computing technology, and an important part of high-value applications in industrial optimization.
However, in our experience, immediately after starting there, you have to work quite a bit further to get into people's actual industrial challenges. The traveling salesman is an important component of industrial optimization challenges, but there are usually a lot of other constraints and factors that you're trying to take into account. So as we talk about traveling-salesman-like approaches to finding good solutions, the heuristics we might use, the choices we make, and the way we optimize them get complicated by the real-world situations that people are dealing with.
Yuval: And I think the traveling salesperson problem is not just about a salesperson, right?
They all do things on Zoom anyway these days. But it could be a UPS truck that's got multiple deliveries, and even the optimization criteria may vary from problem to problem. Maybe you want the minimal amount of time, maybe you want the most fuel-efficient route. Is that correct?
Murray: Yeah, that's an excellent point.
There are companies that have looked at routing vehicles for garbage collection. For instance, Groovenauts and Mitsubishi have done work like that with our systems in order to reduce CO2 emissions. So that's an application where people are trying to impact some large global challenges.
There's benefit to the business in terms of increased efficiency, and there's benefit to the world in terms of reduced emissions. And just like you said, that's the practical reality, where it's not just one vehicle visiting all sites. You also have choices to make in terms of how different resources—in this case, vehicles—get assigned to different areas of the routes that are going to be run.
That's also not unlike manufacturing floors, where people are delivering resources from where they arrive at the shipping location out to the manufacturing locations. You've got a fairly complicated vehicle routing problem as you try to imagine that material moving across the floor and arriving at those fabrication locations just in time.
Yuval: And Sergio, I think TSP is not necessarily a quantum problem.
It was formulated a long time before quantum came into vogue. Could you tell us a little bit about the various classical, quantum, or quantum-inspired approaches to solving these kinds of optimization problems?
Sergio: Right. So indeed, the TSP is not necessarily a quantum problem. It's a problem that potentially, hopefully, can be solved in a more efficient way with quantum computers or quantum-inspired algorithms.
My guess is that it's one of the most explored, analyzed, and studied problems in computational complexity in the world. There may be others, but certainly this is one of the top five, if I had to guess. The other point is that there is not just one TSP or vehicle routing problem.
There are actually, to my knowledge, at least three of them, right? One of them is a satisfiability problem: if I have this budget, or this amount of available kilometers or route to cover, is there a solution, yes or no? And this problem may be easier or more difficult than the others.
Then there is the actual optimization problem, which says, "Hey, for this graph, for this specific list of nodes or cities or parts that my robot has to traverse in a factory, what is the best route?" The most efficient, as you said, in terms of energy spent, or kilometers or centimeters traversed.
And then there is the search problem itself: give me, if it exists, the route with this specific amount of weight, or this matrix that I want to output. The optimization problem is the one that's NP-hard, I believe, whereas the search itself is NP-complete. And that's just one single problem; then of course we add how many hubs or depots we want to put, and how many different buses or delivery boys we're going to put on the map.
So there have been hundreds, possibly thousands, of researchers working on the TSP, and I think we humans have become really good at finding heuristic solutions, right? There's one thing that's common, and I think it's unbeatable, at least classically: it is an NP-family problem. So of course the complexity is going to grow exponentially or more with the size of the problem.
And of course, we have all those challenges, with a couple of universities in the States who are trying to beat each other all the time. A few decades ago there was the Christofides algorithm, which is one of the main ones that people are using today. Branch and bound, in terms of optimization, is quite famous—not just for the TSP, but for many optimization problems out there.
And then simulated annealing, and I guess this is a nice segue into D-Wave. Simulated annealing is one that fights all the time with two other families: genetic algorithms, and the more naive ones like nearest neighbor and stuff like that, which will probably depend a lot on what type of matrix you have.
So nearest neighbor may be the quickest and easiest solution for a specific graph, but a terrible one for another graph. So I think one good point to consider is: is there a better algorithm for a type of TSP or for a specific dataset, right? That would be a nice discussion to have.
Yuval: And Sam, you said that you focus on near-term solutions to these kinds of problems. Could you describe the technology that you're using today? How would your company solve a TSP problem? And if I were a customer listening to this podcast today, excited by this problem, and I reached out to you—we're now in 2022—how large a TSP problem could I solve with your technology?
Samuel: Yeah. One aspect that I would like to highlight, which is often overlooked, is that the mathematical encoding of a use case is as important as the backend or the algorithm you use to solve it. Now, I like D-Wave machines, and surely it is one of the backends that we plan to work with soon.
However, D-Wave is at the moment restricted to the same encoding, the quadratic unconstrained binary optimization, AKA QUBO, problem. You can encode a TSP problem in this mathematical shape, but it's not efficient, as you get an encoding that grows with N squared, N being the number of places you want to visit.
So the potential speed advantage that you might get using D-Wave for this particular use case, you may very well lose with this encoding. Now, to overcome this, at Inspiration-Q we are developing specific quantum-inspired algorithms for the TSP problem. Indeed, we already have a first solution to it, and I think there is room for improvement to tackle problems with thousands of cities.
Yuval: What does—and Murray, I'll let you augment that in a second—but what does quantum-inspired mean?
Sam, that's for you, please.
Samuel: Yeah. Quantum-inspired algorithms are actually an emerging technology. It's a technology in itself. It's not just simulating how a quantum computer works. It's using all the mathematical tools and methods developed during these last decades by all the people working in the quantum information field.
There's been a lot of improvement in the numerical methods that they've employed to simulate, understand, and study quantum systems. And now the idea is to take all those methods and apply them to problems like the TSP problem or others—using ideas or encodings taken from the problems that we are studying with D-Wave and other technologies, and combining all these ideas to produce new algorithms and new techniques.
So yeah, this is a whole emerging technology, and in many cases we've seen that these techniques are able to improve on previous classical algorithms. And in some way, the collection of all these techniques is essentially what we call quantum-inspired technology.
And the benefit is that we can tackle problems in a way that is also compatible with future quantum backends. So I think it's a very pragmatic way to tackle this quantum future, and that's what we do.
Yuval: And Murray, you've published hundreds of different applications and use cases of optimization.
What's your current state of the art for TSP-type problems?
Murray: Well, that's an interesting question. Maybe the thing I should comment on, as an observation, is that when we're thinking about the approach that D-Wave is taking to the traveling salesman problem, we're really focused on practical applications.
And there it's very important to recognize the distinction between the types of algorithms you're going to use to solve the traveling salesman problem and what the quantum processor does. If we look at past industrial revolutions, the main industrial revolution was powered by the steam engine, which produced rotary motion, and the information age was powered by classical computing, which can do addition and multiplication blazingly fast.
Quantum computers are able to draw solutions from these QUBO problems that Sam was introducing with very high efficiency. Those problems are combinatorial in nature, which means that you're looking through a problem like, let's say, picking a set of players for a sports team.
I have a pool of players, and I'm going to pick the ones that are the best. But once those players are on the team, they don't have a particular order. They're on the court playing together. In a traveling salesman problem, the order of the cities that you visit makes a difference. So if you visit A, then B, then C, that's not the same solution as A, then C, then B.
And for exactly that reason, the problem formulation issue that Sam was just describing means that you don't want to try to brute-force it directly onto the quantum processor either. If you just take a very natural, straightforward description of a traveling salesman problem and put it onto one of our quantum processors, you might get tens of cities.
And if you compare that to, let's say, the best in the world, like Concorde, you can get up to tens of thousands of cities. What's more, you can prove optimality at hundreds or thousands of cities in just a few seconds or minutes. One of the reasons why classical algorithms like Concorde can do so well is that there are a lot of simple observations you can make—for example, when you're routing a single vehicle across a route, it's on a surface.
It's not moving in three dimensions or higher dimensions. And so by recognizing all of those constraints about mapping vehicles on the surface of the Earth, you can actually help classical computers do extremely well at large sizes. But, to the discussion we were having earlier, there are many problems in vehicle routing, or in planning the deliveries of fleets of vehicles, where a number of other constraints are at the heart of it.
So the way that we look at applying quantum annealing to this problem is by finding encodings where—and this is actually very much like the classical solution approach—you're starting with a feasible solution, and then you're looking to make exchanges. You're finding areas of high cost, and then you're looking to make swaps that keep you in a feasible space.
It's a loop that doesn't break. And you can use the type of problem-solving that quantum annealing provides to make those kinds of movements in the solution space in a hybrid algorithm. It's important for people to keep in mind that it's not about quantum or classical. We've probably done about 10 years of benchmarking, learning what the best capabilities of those two technologies are, in order to understand how they solve problems, and they have complementary skills, right?
Classical computers are very good at local updates, and quantum computers seem to be remarkably good at non-local updates, where they can move between large areas of the solution space. And that's an important part of problem-solving: when you get locked into a particular solution, just doing one swap or exchanging two cities is not going to get you out of it.
You need to reconfigure a portion of the problem. That's where we're applying quantum computing in these hybrid solutions, and taking that hybrid approach is what's allowed us to get to industrial optimization challenges in vehicle routing—for example, when we've looked at support vehicles delivering water, food supplies, or vaccines out to locations.
Sergio: If I may add something there—thanks, Murray, for mentioning Concorde. If I'm not mistaken, that's basically very low-level C optimizers on linear programming that try to optimize to the tiniest bit, both the in-memory handling of the combinations and the actual formulation of the problem. Also echoing what both of you are saying: the domain is super important, and so is how you do the mathematical formulation of it.
But I would like to dig a bit deeper on efficiency, whether it is tied to a specific problem or to the more general formulation—if we take, I don't know, the traditional Ising formulation for the TSP. We have Concorde, and we have OR-Tools, which are also really, really efficient and work pretty well for a lot of real-world problems with the sizes that we have today.
We also have companies like Gurobi, for example, doing really good classical optimization with very specific proprietary algorithms. So when we talk about efficiency, are we talking about orders of magnitude in solving the problem, or small steps, or maybe finding heuristic solutions that we could not find before?
Murray: Yeah, that's a great question. I think it's very natural for practitioners who are working in this space to think of quantum annealing as a kind of heuristic. It's part of a technology stack that they're putting in place in order to build solutions in their application domains.
Certainly, industrial optimization brings a lot of challenges with it: not just "Can I come up with a good plan?" but "Can I come up with a good plan that a workforce is actually going to use?" Because the driver on the road, in the traffic, is the one who can choose to deviate from a plan.
So you've got a lot of those hard factors and also the soft factors as you're looking to make those improvements. Certainly, as we have been looking at and understanding how quantum mechanics can be turned into useful work for us, we've realized that proving optimality, even in the sub-pieces that the quantum annealer is actually contributing, is not the quantum computer's role, right?
The quantum computer is basically helping to do these fast reconfigurations. So I would say that by the time you see its effect in the business-quality result, you're not looking for something that's going to suddenly change the business operations capability by a thousand times. I mean, it would be fantastic and terrible if that happened, in terms of the disruption to the business from that kind of rapid change.
It's more something that increases efficiency. Although, practically, what we find is that as soon as you increase the efficiency, the business will then grow, or they'll say, "Oh, there's some other factor that we would now like to include that we've always had to exclude from our problem-solving," and they'll bring that in, and the complexity will go up.
So really, for everyone who's running a business that is bound by problems this complex, you will always hit a ceiling. No quantum computer or classical computer will ever be large enough to reach your ambitions in terms of how big a scope you would like to run.
And so by combining quantum and classical computers together, you're effectively helping to push that ceiling as far up as you can, and you're also doing it in a way where you're setting yourself up to take advantage of CPUs, GPUs, and QPUs—basically all of them—as your compute resources in the future.
Because you certainly don't want to be in a situation like, let's say, in materials science. We've done some experiments in quantum materials simulation where, in two product generations, we went from not being able to solve problems to being 3 million times faster. If that happens in your business, you've now got to redesign all your application infrastructure, and that's the negative part of disruption.
But if you start building your applications with all these technologies incorporated together in a hybrid manner, you're effectively getting all the gains of the technology and of the algorithms as they build on top of one another, and that's where I think people are seeing real value today as they're using these technologies.
Yuval: So I've learned from you that it's important how you express the problem, and that there are multiple ways to run it on the backend, whether classically or on a quantum computer. There are also hardware solutions—FPGAs that implement Ising models. So a question to all of you, maybe Sam first, then Murray, and then Sergio: how should a customer go about choosing the best provider, the best technology, to solve a TSP problem?
I'm a customer, I'm motivated. I understand that even a 5% improvement in my business has a huge financial impact. How should I go about making that choice?
Samuel: Yeah. I think the best depends on your needs, right? In any optimization problem, like the TSP problem, there are at least three dimensions we need to care about.
These dimensions are interdependent, so we cannot reduce the problem to any one of them. And we human beings often try to oversimplify things. The first is the number of variables—in this case, the number of cities you want to visit. Second, the quality of the solution.
And third, the amount of resources—say time, say the number of processors, et cetera. In summary, your budget, right? Depending on what you demand in this decision space, you may choose one option or another. Maybe you want a fast online calculation with few resources, even if you don't get the best solution, because what you really want is a fast solution.
Or maybe, as someone was saying, it's very important for you to have a very optimized solution, even if that takes more time and more resources. So there is probably no single algorithm or backend that is going to be the most suitable in every circumstance. You first have to ask yourself what your constraints are and then analyze the best solution, knowing very well what you really want.
Murray: Okay. The way I would answer that question—because it's a very difficult question—and the reason it's challenging is that there are so many people coming at this problem, or experiencing this problem, in so many different ways, right? They might be thinking, "Do I want to build my own solution for this?
Do I want to purchase a solution for this?" And there's also a lot of depth to it, right? I have gone through exercises as a product manager asking, "What is my customer's experience?" Well, if you start searching for industrial optimization,
Samuel: Phew.
Murray: the world of stuff that falls on you at that point is amazing, you know?
It's like, well, become an expert in algorithms, and then become an expert in tuning the parameters of those algorithms, and then become an expert in analyzing which of those algorithms is contributing on which instances on a given day. When I'm talking to math experts, that's the world that they're immersed in.
But a business leader who's trying to pick a solution is thinking about things like, "Well, even if I do run the best technology in the world, I have to make sure that I have high uptime, that I have high-quality solutions, that my software is not going to break, and that this is going to be updated and maintained over time."
So I think people are looking at composite classes of factors when they're trying to make this decision. It's sort of like: I want good people, I want good processes, I want good technology, right? So they're going to meet with a team and ask, "Does this team have experience?
Do they have the ability to understand my problem and design a solution that's suited to my experts and their needs in operations?" Then they're also going to look at what processes are going to be used. Are we going to use an agile development method? Can we do continuous deployment?
Am I working with a group of folks who are very conservative, so that we need to gradually introduce this new operating process? That's an important business decision at the time you're picking a solution provider. And then on the technology side, in the best-case scenario, I don't want to be trying to pick winners, right?
So that's a big part of it. It's our customers who drew us in the direction of building hybrid solutions. Samuel was mentioning quantum-inspired earlier. A lot of work was initially done on CPUs with simulated annealing, and that was a big test bed against quantum annealers to see what they could do.
And when quantum annealers started beating simulated annealing, people started to explore and say, "Oh, the quantum processor actually has access to this big memory space. Let's make a classical, quantum-inspired version of that using a GPU, and now we'll create slices of the memory space of the solution spaces we're searching," and now you're incorporating a graphics processing unit.
Both of those have contributions to make. It's not that one is better than the other. And now we have Nature papers showing that quantum annealing can reach high-quality solutions faster—fundamentally faster—than any classical approach can possibly do.
So on the technology side, it's: how do I get the best benefit from all of those things? And keep in mind that the technology is not just the processor; it's also the deployment infrastructure and your ability to scale up and scale down. You don't want to be using these resources when your problems are not actually running.
So people are thinking about it in that broader IT context and making sure the solution suits them there. I think people, processes, and technologies is how people judge those providers.
Sergio: So I'm going to try to give the perspective of the non-vendor, or better said, the user of those algorithms, both classical and quantum. I will start by saying that I think both the D-Wave and Inspiration-Q solutions, among others, are really good, really smart, and really efficient, having personally seen both of them.
So the first thing is kudos to both of you, because you're definitely trying to do something that I've tried to do for most of my professional life, which is solving the TSP, right? So thank you for that. Hopefully we get there together soon. I think one of the problems we have out there is that in some cases we are mixing layers of conversation.
We talk business, then we talk engineering, and then we talk algorithmics, right? Or computational complexity. If we keep the conversation on each layer, then we make sure we're talking at the same level, but sometimes those conversations start to mix, and then it's a bit chaotic.
And I think you touched, Murray, on a lot of those points: what the integration is going to be, what it will look like, how it will improve over time, how my problem will evolve over time. You touched on a really good point that we've seen over and over with AI, cloud computing, and so on: okay, now we have better machines or better technology to do things, so now let's tackle more things.
We are like gases, right? We always use all the available space that we have at our disposal. If we start from the ground level, computational complexity, I think we need to come to an agreement—and I'm not sure if today is the place for it or not—but there's a lot of conversation about what BQP is, what NP is, and whether we can achieve quantum advantage for general NP problems, right?
And if we enter that conversation, it's opening a can of worms that would probably need a three-hour podcast, or a 30-hour podcast, and more people to come in and comment. Maybe we can agree on something that I think both of you said: sometimes it's about the local efficiency for your type of problem and improving what you have today.
You don't necessarily have to go into claiming quantum advantage or claiming P equals NP—otherwise some of us would have a lot of money in that regard. So let's say that there is a business advantage in how we do things. From the user perspective, there is one thing I would really like to see, which is more business-side benchmarking. On the TSP specifically, and many other problems as well, you can see plenty of classical benchmarks—you mentioned simulated annealing, which is really good.
Genetic algorithms tend to be very good for certain models, certain types of problems, or certain datasets. The same can be applied to many other complex and not-so-complex problems, at the end of the day. Perhaps it's because the classical algorithmic problem has been around for much longer and is more mature, in a way, and of course quantum technology is still in its early stages.
But to your question, Yuval, about what would be nice to have, or how a client would choose: that benchmark could be extremely useful in helping each other improve on what we have. And I think that's the beauty of quantum and quantum-inspired: we keep pushing each other to improve, right?
Now there are post-quantum algorithms. Suddenly we have much better classical encryption algorithms that have nothing to do with quantum, but the ecosystem itself helped push the boundaries of what we thought was possible. So for me, it's pure benchmarking: saying, "Hey, this is how far we can truly get in time, accuracy, or cost," which are, at the end of the day, the three things that matter to companies, or a combination of the three.
And sure, some cases will require speed-ups in calculation time at the expense of some accuracy because, as you say, Murray, you need to be more reactive to soft variables. In some cases, you just want to get something as cheap as possible and put it out there in the market, and in other cases, you want the bloody best solution you can get.
So that is something that I think can help the industry in general a lot, and also the client base. It would remove part of the fog surrounding the ground level I was mentioning. Is BQP under NP? Well, we think we know that's not true, right? Does that mean we cannot get some business advantage by using quantum-inspired algorithms, or quantum annealers, or a combination of hybrid algorithms that take the best of both worlds?
I think the answer is yes, but we need to navigate through the fog.
Yuval: So, my next-to-last question—and thank you, Sergio, for that—touches on the point of benchmarks. If I were benchmarking a CPU, I could use PassMark, and if I were benchmarking a machine learning algorithm, I could use MLPerf. Is there, or should there be, a standard benchmark for TSP?
Maybe Sam quickly, and then Murray.
Samuel: Yeah, I think the problem of benchmarking is a critical one, I agree, because for many problems in quantum computing, at least, we know that there should theoretically be an advantage. But in the case of optimization problems—and of course TSP is no exception—we don't have that theoretical advantage, in principle, as far as I know.
And we won't have it. We know that even if we have a quantum computer, the complexity of the problem is going to grow exponentially. Now, that doesn't mean we are not going to find an advantage, because we are using heuristics, and we can use one processor or one algorithm and find an advantage using those means.
So yeah, there is no way but to benchmark different solutions for real use cases; it's just a matter of finding real business value. I think that's what we all have to do. And in that sense, I think all of us have to make an effort to invite the whole quantum community to start doing these benchmarks and to compare solutions to find the truth, right?
This is what matters in the end: to find value and to make progress in society using these technologies.
Murray: Yeah, I think that's an excellent point. One of the excellent things about having these new emerging technologies is that we're pushing each other to do better, just like the point Sergio was making, which I thought was an excellent point.
With problems like the traveling salesman problem, where the complexity is explosive, it's not like other problems where you can say, "Well, I can show you that this is the best way to solve the problem. The smallest number of additions you can turn this into is this many, it's N, and this is how many we're doing per nanosecond, and as a result, this is the fastest way to do it."
It's much more like: well, 100 people have tried this problem, this is how well they did, and I was able to do a little bit better than them. It's more of a relative comparison. There's a researcher named Carleton Coffrin at Los Alamos National Laboratory, and he remarked that what's fantastic about quantum annealing is that it's going to advance the entire optimization field, because people now have something to compare themselves to that's not classical computing at all.
Its fundamental methods and its scaling curves are going to be different, and that is causing people to open their eyes and say, well, maybe it is possible for a classical algorithm to do better, and that's causing them to explore that space better. We're constantly looking at online benchmarks in a variety of spaces.
Certainly, if you're going to do just native surface-of-the-Earth TSP problems, we're not going to do as well, and that's because we're not optimizing to do extremely well in that space, because we're talking with companies about the industrial problems they're working on. It might not be obvious to everyone, but there's this no-free-lunch theorem in these algorithm spaces: you design an algorithm for a particular class of instances of a problem, and if it's the best algorithm at that, it will absolutely be worse at another group of problem instances, right?
You have to make choices. For instance, if the solution space of your optimization problem is a bowl, the fastest algorithm is to run downhill. And the moment you start to get some undulations in the surface of the bowl, that's no longer the fastest algorithm. But both of those algorithms are important when you don't know the structure of the problem you're trying to solve.
And herein lies the point Sergio was making about business benchmarks: how do people orient themselves when they're trying to create industrial solutions in these incredibly complex spaces? It's very challenging at a detailed level with the technology stack people are looking at, because if I work with two groups of people in, say, vacation scheduling—nurses, and let's say people who work in airports, or something like that—
the context and the details of how they run their businesses actually change the quality of the solutions and the approach—the formulation, as Samuel was saying—that you would take in those two cases. However, businesses do compare themselves, right? And certainly, when we're talking to people about Save-On-Foods, and we say we took an important grocery optimization problem from twenty-five hours down to two minutes, there are other people running business operations who say, "Okay, well, I can understand some of the complexities of grocery optimization."
That's a reference point. And I think as we do better, and as we get more of these use cases built, there will be more of the reference points Sergio was talking about, which will be important for businesses as they try to make comparisons and orient themselves.
Yuval: Excellent.
So Sam, Murray, Sergio, how can people get in touch with you to learn more about your work?
Samuel: Well, visit our website, inspiration-q.com, or just get in touch with me directly through LinkedIn, Twitter, or whatever. Yeah, we'll be waiting. You will be welcome.
Murray: Yeah, you can find me on LinkedIn, I guess, if you're looking for me specifically.
For D-Wave Systems, if you want to learn about our quantum computing systems, come and check out our website. You can also get free access to our quantum computers through our Leap cloud service. If you search for D-Wave and Leap—as in jump—you can sign up and get access to the systems and our open-source code examples.
For developers, maybe the more natural way would be to go through our GitHub, our SDK, and our open-source examples there, and start looking through our Python and C toolkits. And for businesses, we often work with them to help them through a phased approach, from discovery to proof of concept, called our Launch program.
So contacting us through sales is a great way to get connected and learn how to identify business problems that are suitable for quantum computing and quantum hybrid approaches, and then start to build them out.
Sergio: So I won't be able to give anyone cloud access to any algorithm, because I don't have that.
But feel free to go to quantumpirates.substack.com, or look me up on LinkedIn, as well as what we do at Moody's in quantum. I'm especially interested if you have a good exponential solution for the TSP or any other algorithm. I would really love to hear from you and take on the challenge of improving on each other's results, standing on the shoulders of giants, I guess.
Yuval: Excellent. Gentlemen, thank you so much for joining me today. Thank you. Thank you very much as well.
