A Developer's Guide to the Stride Hybrid Solver

[00:00:01]So I'm very pleased to introduce Daniel Savv. Daniel is a senior technical adviser at D-Wave. He's responsible for the uh AMIA region and he's based in Munich, Germany. Daniel has over 10 years of experience in enterprise software across verticals including manufacturing, finance, supply chain and logistics. and he works with our customers along their entire journey from identifying the right optimization use cases to formulating complex problems and achieving successful adoption of D-Wave's quantum platform.

[00:00:33]Daniel holds a masters of science in industrial engineering and economics and today he will take you hands-on through building real optimization applications with Stride hybrid solver. So welcome Daniel and thank you >> amazing. Thank you Susan for the introduction and welcome everyone to today's webinar. Thank you for the amazing turnout. Agenda wise, we have prepared a couple things today. We're going to start with a brief introduction to to D-Wave in terms of our technical portfolio of our current capabilities and also introduce the stride hybrid solver.

[00:01:09]Um afterwards we're going to be becoming more and more hands-on and more and more concrete in terms of example by introducing Ocean SDK which is our in-house software development kit. This is the way to start building the quantum quantum based applications and we're going to be presenting a problem very applicable to pretty much every industry um which is going to be the school bus routing problem as a modeling example coming from um you know the math behind that. How do we how do we formulate that um in terms of applied mathematics um operations research and how do we implement that using uh ocean SDK on the strides over and I'm going to show you some code. So it's going to be technical but not that technical that we're going to be writing the code it itself. There are a couple of snippets that we are going to implement in in here as well.

[00:02:01]Um, in terms of housekeeping, the total time scope of today's presentation is going to be around about 45 minutes and the remaining 10 to 15 minutes we're going to use as a Q&A session. So, please, as Susan already introduced, use um the the option below to formulate your questions and we're going to respond to them either during the presentation or afterwards in the Q&A session.

[00:02:24]Amazing. So, my main goal for today is going to be three things. I'd love you to start experimenting with stride. So take take one real world problem that you have from your industry and try modeling it um using the hybrid solver.

[00:02:40]Explore the open source examples that we have on GitHub are the fastest starting point to our solution and to the problems that we have depicted so far and tackled so far. And last but not least, just come talk to us if you believe you have a decent use case coming once again from your industry, from your field of expertise, so that we can figure out together what the best approach for for that is going to be.

[00:03:04]And with that, let's start with introduction and with some motivation.

[00:03:09]Of course, the most obvious question for today, why quantum computing and why now?

[00:03:15]We're getting that question asked pretty much in in every meeting and that that we are starting with right and a short definition of quantum computing I love using is is that quantum computing is an energyefficient way to compute especially for problems of high complexity and as you can imagine in real world combinatorial problems explode. The number of possible solutions grows faster than any classical machine any classical computer can enumerate.

[00:03:45]So on on this slide we have listed three failure modes or usual failure modes of classical approaches on these problems.

[00:03:51]Right? First and foremost the classical computer may may not be able to handle that complexity of the problem. It may take too long in terms of computing time. So what I usually observe um from customers examples that in some cases a solution to the problem may take up to 10 or 12 hours for a decision that you may need hourly. So you're eventually having old data in the system which is not that efficient.

[00:04:22]And last but not least it may consume or burn an unreasonable amount of computer energy.

[00:04:29]So either of these or a combination of those are signals that we believe that a quantum computer may outperform that and that doesn't mean that the quantum computer is a general replacement for the classical computer. In my eyes, it goes hands in hand, right? There are different classes of of problems that are perfectly suitable to be solved on a quantum computer, on a quantum analer to be specific. And the same same goes uh for the classical machine, right? Um and today we're going to be focusing exactly on that.

[00:05:04]So let's start with the hardware behind that. Dwave is driving a dual platform approach and these are the main paradigms that you would see and usually observe in a quantum world. Right? On the left hand side we have the kneeling machines and on the right hand side you have the gatebased system. The focus of today's presentation is going to be the kneeling machine um which uses the natural tendency of quantum systems to settle into their lowest state of energy which makes it perfectly suitable for combinatorial optimization and any kind of optimization problem you might be facing. Be it scheduling, be it rooting, be it any kind of assignment. Um those are problems that organization face pretty much in in in in their operations on a daily basis.

[00:05:52]Right? The gatebased side um is primarily anticipated to be used for solving differential equations for example modeling modeling drugs, modeling materials and so on. But it is believed right now and it's still a matter of research that is far away from production. Right? And this is the key differentiator between those two technologies that the gate based systems or the gate model systems are primarily used for research right now. And the kneeling machine is the the one we use on production right now. As of today, we have customers that plan their their production schedulings, do their operations on a kneeling machine using our solvers.

[00:06:37]And that's why we're going to be focusing on that today. So if your problem is pick the best combination out of an astronomically large set right analing is your tool today.

[00:06:52]So what does the kneeling stack look like from a product standpoint?

[00:06:58]Dwave has been researching on the hardware side for almost 25 years.

[00:07:05]This is how we got the journey started, right? And this is what you see as the first pillar forming that product portfolio, right? This is the advantage 2 machine, which is our sixth generation of kneeling quantum computer with almost 4,400 cubits. And this is the backbone of everything that comes on top. The second pillar is the leap cloud platform. This is the way to access those quantum computers and the available solvers without having to obtain the infrastructure behind that.

[00:07:35]Right? So we call it QCAS which is quantum computing as a service. And you can imagine that's comparable with any cloud provider on the market right now.

[00:07:45]So you're getting your API tokens and you can send your problems over the lead platform to D-Wave to our solvers and get a solution back without having to install a kneeling machine at home or in your organization.

[00:07:59]Right. [snorts] Um the next thing on the slide is ocean SDK which is a software development kit which we use to embed those problems or encode those problems um on the machine.

[00:08:13]it is very native to programmers. It's using Python functions and Python as a language as a general language um model to to to work with that. And this is actually one of the frequently asked questions I'm always getting, right? Do I have to have a quantum background so that I can start working with with D-Wave or start programming my my problems, my models on a quantum machine? And the short answer is no. You don't have to do that. This is taken care by the pre-built functions that we have in ocean and the next pillar which is stride which is the stri uh the hybrid solver that we're going to introduce later as well together with ocean solving them actually. So for you it means two names to remember for the remaining round about 40 minutes. Ocean is how you write your models and stride is what actually solves them. And last but not least, we have our professional services in house as well. And this is especially meaningful for organizations and for people out there that don't have experiments, be it in optimization, be it in quantum based optimization. We can leverage the the wisdom and best practices our experts have accumulated over time to help you write your programs using stride on ocean.

[00:09:31]So this is as a whole how we usually deploy applications for organizations and let me make deployed a bit more concrete with a few success stories from real world customers and use cases that we have built in the last couple of months and years.

[00:09:53]Let's take BSF as an example because it's a client for from my region um and it was a project I was um also involved with.

[00:10:03]BSF used our hybrid solvers to do their production planning. We're talking about multi-sight production planning for their filling clients. Thousands of products and constraints that they have incorporated into that model.

[00:10:18]And if we get back to the very beginning motivation slide that we have looked in T BSF have actually managed to improve both the time to solution from 10 hours to just a couple of seconds to 5 seconds. But not only that also to improve the solution quality in every part of the objective function that we have mutually agreed on.

[00:10:41]And this leads to the key key fact in all of this, right? Like we're not doing quantum just for for the sake of it like quantum for quantum sake. It's a high value problem formulated correctly run on the hybrid solver.

[00:10:57]And exactly that formulation is a skill that we're going to be giving a closer look into today.

[00:11:06]So let's meet the solver the stride hybrid solver that we we we have talked about to you and introduced just at the very beginning. What does hybrid stand for? First and foremost, it's a combination of classical and quantum resources working together to solve your problems.

[00:11:24]So under the hood, you submit one model and the solver itself orchestrates the rest. It breaks down your problem into subpros and sends them based on the problem you're you're submitting either to a classical resource or to a quantum resource.

[00:11:42]So once again there's no quantum background required in order to do that.

[00:11:46]All the fine-tuning of the different very specific parameters embedding as well also frequently asked question is taken care of um the strides over in in in the back end. It can support up to two million nodes we call them which means variables decision variables and and constraints in in your data and there are tons of features that are getting released which eats each cycle that we're going to be having a look um into in the next couple of slides.

[00:12:18]So how does one solver handle such different problems?

[00:12:23]By combining four different philosophies of or schools of optimization.

[00:12:29]Starting from the left once again quantum optimization.

[00:12:34]This is the heart of the solver. This is the heart of the wave the backbone of everything else. Once again it explores natively huge solution spaces via quantum effects. quantum tuna link for example so that you can find the lowest state of energy in your energy landscape after modeling your problem.

[00:12:54]If we go next, tensor programming, very familiar for for the ones of you working with ML models or AI initiatives, it's an array native modeling. So if you think in numpy, for example, you're going to be feeling home with that, right? Vectors, matrices, arrays, and so on. Discrete and combinatorial modeling. That's a native way to formulate your problem with permutation variables. Be it lists, be it sets, any kind of discrete choices that you can um you can come up with.

[00:13:29]So, it's a really efficient way to increase the performance of uh of of your problem or of the solution you're going to be trying to to to achieve with that by just choosing the right variables. Right? For example, if we take a routing example, right? what we're going to be um um looking into into the next couple of slides and minutes. Um if you go with a list variable, it naturally represents the order of the stations I'm going to be visiting instead of having those binary variable gymnastics. So it saves you lots of resources by introducing that and by using list or combinatorial variables lets you encode constraints explicitly into that without actually stating them separately. And last but not least, mixed integer linear programming.

[00:14:18]That's the classical workhorse that is still in the mix, right? And the key point about that is linear pro pro um problems can also be handled by by the hybrid solver. But we are talking also about nonlinear relationship between variables that the stride solver is capable of solving.

[00:14:40]So concretely here's the vocabulary you get to model with. It's a rich growing numpy like symbol library. You don't have to memorize all of those names or you know like symbols that we have given here. You just have to know it exists.

[00:14:57]It's well documented in notion. So whenever you need one of those just remember um that that it is there and you can research on that um as a next step. The color coding is the release cadence over the years. So, Cubits is our annual in-house uh conference that we are hosting pretty much every year.

[00:15:14]Um, and on on this slide, you can see how the solver evolved over time. So, community feedback is also lending in those change requests over time. And we encourage everyone to submit feedback.

[00:15:29]We're actually looking into that under the hood all those symbols were together into a graph and that's the key mental model for Strat like to the question how is it different it's an asyclic directed graph that we are creating so if you've built something around um in in in PyTorch or Tensaflow it's pretty much the same idea right like you declare your structure and into that tree model, right? Like your constants, your variables that you you see theam on the example on the right hand side. And after declaring the structure, the solver takes care of the rest and handles the search of the optimal solution, exploring those options, right?

[00:16:18]And before we investigate the math behind that and the code behind that, where does a quantum powered application sit in a real enterprise landscape? This is al also one of the questions that we always get whenever we are talking enterprise right on the left hand side you see your existing world the only new piece that we are adding to that to all the ERP landscape and different ERP systems be it SAP or be it Oracle whatever you're using is the ocean SDK that you install on prem on the right hand side you see lib cloud service how the hybrid solvers actually network distributing those sub problems to either the classical resources or to the advantage 2 system and in between you have the integration pattern you extract the problem data from your ERP system co-le over the cloud API and write the solution back so from an architectures point of view is just another I would say restish API right and the most important point about up one is that no company sensitive data points leave the existing infrastructure. Right? The only thing we're actually submitting to LEP is the mathematical formulation of the problem and you're getting the optimal solution back from LEAP which is a bunch of zeros and ones or whatever like vectors that you can map back together to your sensitive data in the organization.

[00:17:53]Amazing. so much so much to the theory.

[00:17:58]Everything from here is going to be a bit more hands-on. So I'll show you the five moves you need. First of all, from a generic standpoint to create your applications, starting with the model creation, defining the different symbols, setting an objective, adding constraints, and submitting to the pro to the solver eventually.

[00:18:19]And afterwards, after we've done that, we're going to take a look into one concrete example.

[00:18:28]So how do we usually get started?

[00:18:31]Ocean once again as already introduced is a suit of Python packages for optimization on Dways. So the majority of the code is open source. It's on GitHub. Um you can go out there and research that and once again extensions and features and any kind of feedback is always welcome from the community.

[00:18:52]So you install it in just one line and then authenticate with your D-Wave credentials connect to leap and from then on you can start working on your problems.

[00:19:04]How does does that work concretely? So two imports actually get you started in in terms of instantiating a stride or formerly known NL object and sampler right first and foremost from Dwave optimization you're importing the model and that model is the empty place holder for your directed as cyclic graph that you're going to be using we're growing afterwards right the next thing is calling the sampler Right. Importing from lib hybrid n samper that's your gateway to stride in the lib cloud. It picks up your API as already mentioned um the token and connect to the right environment so that you can submit your problems and probably on that side one note on terminology you see NL often in that sampler that's the former name in the solver it means nonlinear once again because the solver is capable of solving nonlinear relationship relationships between variables it's just in the API names So how can we import some data? Your problem data enters as plain numpy arrays for example. So you import numpy snp. You can create whatever you need.

[00:20:25]Right? In the example that we we're going to have a look in a minute. Um it's an array of distances. It could be times. It could be kilometers between the different stops. So in this case it's just an array. There's no special data format that Ocean is requiring. No file conversion. So if your data pipeline um ends in a numpy array or whatever pandas data frame, you're one line away from a model constant.

[00:20:54]How do we define those constants? By model add constant. So we are adding the distances between the different stops.

[00:21:02]And defining a variable is also pretty much um easier than expected, right? In this case, we're going to be defining that list variable we we've been talking about.

[00:21:13]And and once again, that's the native way for combinatorial variables to be smaller in size compared to all those uh binary variables and more natural to the models. So a list is one ordered row route over all items or over all stops you're going to be visiting. It could be ABCD. It could be a different order.

[00:21:34]This is the way to go. And what we're going to be introducing on top of that is a disjoint list. For example, it's a bit more complicated, but the same idea behind that is the partition of items across kordered items or lists. So it's used mainly for multi-vehicle routing problems, right? And in this case for a list variable, it could be just one vehicle that I'm going to be optimizing the route or the optimal allocation of stops I'm going to be I'm going to be going.

[00:22:05]How do we set an objective and add constraints?

[00:22:08]Um in most of the cases we are minimizing a certain value and that value could be any scalar symbol built from your graph, right? It could be the summed route time. It could be finances, costs that you're optimizing in a certain direction. Whatever the business needs, that's eventually what you're trying to achieve when we are talking value of the objective function, right?

[00:22:36]You can add multiple constraints in a very native way. Once again, you don't necessarily have to be thinking in cubos um and so on like natively programming and embedding on the quantum computer.

[00:22:47]It's just a function that you're calling in this case add constraints and you can define your um boolean um more or less comparison or whatever you need in that brackets, right? And the distinction over over there is pretty obvious, right? Like objective is what makes a solution good and constraints is what makes a solution legal.

[00:23:09]And the solver reports whether the return states are feasible.

[00:23:16]And in a real example, you're adding multiple numbers of those constraints, right? Like based on the problem you're facing.

[00:23:23]How do you submit the model to the sampler? Um pretty easy. You're calling sampler.sample um with a couple of um parameters in in the brackets. So let's start with those.

[00:23:34]Uh a time limit. Also very important. um that's your compute budget in seconds.

[00:23:40]So for hybrid solvers, you're usually trading time for solution quality. It's usually seconds. So there's a default number of seconds that is assumed. But if you have round about a starting value to start with, you can just you can just go for it. Label, this is how you would call the problem. Um so how it shows up in your leap dashboards eventually. So your future you will thank you whenever you give it a name. um in case you have to submit multiple problems or start debugging whatever you need. It's just a way to distinguish between the different problems.

[00:24:16]And the results, last but not least, they come back as states on the model.

[00:24:20]So you read the variable states to get your answer back returning from the sampler.

[00:24:28]Speaking of states, um what we can also do is to more or less lock the model. So it freezes the graph structure. So states can be can be attached to that.

[00:24:40]But also one of the best practices that the professional services team is always recommending is to start with an initial state. This is a way you can more or less warm up stride with a known solution. That's huge in practice, right? Like if you can imagine if you're optimizing any kind of schedules out there, be it the routing of a truck, be it your production scheduling or whatever, most of the companies have already a plan on how things are going, right? So that plan that is good enough to have worked with in the past weeks, you can start off um to to warm the model up.

[00:25:21]And that was the whole API, right? the generic example we we we were talking.

[00:25:25]So now now let's use all of that in a real world problem for today's webinar. I've chosen the school bus routing problem.

[00:25:35]This is going to be the heart of the session more or less. The school bus routing problem is a vehicle routing variant and it's a great template. So if you understand this example, you can adapt it to pretty much every vertical, right? It could be related to delivery fleets for example for field service technicians for stuffing right um shift assignment so anything with the science sequence allocate of resources and so on how does it look like we're talking about two coupled decisions right the very first one is the assignment which cap serves which didn't and the subsequent decision is the routing one. Which way am I am I going?

[00:26:24]In what order do I navigate my caps or my trucks or whatever.

[00:26:30]So it's a free word relevance. Once again we can be talking school districts but the same structure appears in last month deliver patient transport waste collection pretty much every right.

[00:26:42]What are the primary objectives we're trying to to achieve with that? That's our main function. We're looking into a couple of things, right? We want to minimize the number of buses used. Fewer buses is better, right? We got to have fuller buses in terms of seat utilization. We don't want to be having empty seats on different buses. And last but not least, we want to have shorter routes for everyone involved, which results in time wasted or anything that is associated with that pretty much operational costs that you're releasing.

[00:27:21]And three objectives means we'll need to prioritize them somehow or weight them somehow. So this is coming in a second as well.

[00:27:31]What are the main constraints that you have um in in practice? Each route should start and end at its caps depot station.

[00:27:41]That's very important. Pickup time windows must be respected. Every student is on exactly one cap. And that's actually very important because this constraints or exactly this constraint doesn't have to be added additionally because it's part of the disjoint list. Right?

[00:27:58]Disjointed list give us this one for free more or less by construction implicitly.

[00:28:04]We have a certain capacity limit per per cap that we we got to be respecting.

[00:28:10]And of course in the real world example further constraints may be incorporated be her constraints or beat soft constraints.

[00:28:21]So from a mathematical standpoint how does it look like?

[00:28:26]We are always starting with a mathematical model right and afterwards we can always translate that using the Python functions into code. But what we are trying to achieve in here, let's define our variable in this case or disjoint list.

[00:28:41]So we are talking about n students partitioned across k caps and air k is that ordered list of students on bus k right?

[00:28:55]Rki is the student at the stop. So we have the stops incorporated in there and P is the partition between those. So the single variable choice simultaneously encodes who is on which bus and in what order they are being picked up assignment and routing in one symbol in this case.

[00:29:19]How do the constants look like? This is our our data and that data is usually being imported right the constant symbols usually in such cases for routing examples or whatever are huge arrays or matrices. So the first one the very first one t may represent the student to student travel time.

[00:29:41]It's a symmetric array with zeros on the diagonal because that's the the distance from from me to myself more or less.

[00:29:48]That's why it's zero.

[00:29:50]We have the capacity vector which is C the number of seats that each bus is having and we have the school vector s which is which school is which student assigned to.

[00:30:06]So constants once again are where your data engineering meets model and everything the the solver needs to know about the real word this array so that it can process it and come up with the best solution.

[00:30:19]How does the objective function look like?

[00:30:22]We've been talking about three terms.

[00:30:25]It looks scary at the very beginning, but if we break down into the different terms, it it will eventually all make sense, right? So, the very first firm uh term that we just introduced was the number of buses that we are going to be using.

[00:30:41]The second one, the difference between the capacity and the taken seats counted on buses actually used is your empty seats. Right? We want to minimize those unused seats seats on used buses.

[00:30:58]And the third one, that's your travel time between students on the way.

[00:31:07]So that's the slicing of that travel time that is depicted over there.

[00:31:13]Route shifted against itself gives consecutive pairs index into B. So that's pure numpy thinking if we want to formulate it that way. And last but not least, we already introduced that as well. Multiple objectives means in the most cases we got to implement some priority. And we are implementing that priority by by choosing weights which is the three symbols be behind that. In this case I want to optimize first the number of buses uses so in exactly that order then the unused seats and then the travel time. So it's beta greater than alpha greater than gamma.

[00:31:55]And let's translate right now all of these into code.

[00:32:02]It starts with the imports. Once again, we don't have to go through every line.

[00:32:06]You're getting all the slides afterwards and there are multiple examples online, but just to give you a flavor of how the application may look like from a structural standpoint.

[00:32:16]Over here, it's a bit of housekeeping, right? Like we have one container that keeps the model constants organized.

[00:32:23]Standard imports. We've talen about that. the model um calling the solver exactly the ones from part two. We have the model constants data style containers capacities school to student school of student the four travel times matrices and sizes. So it's nothing glamorous on that slide. It's just a discipline of how do we want to model that right pays off whenever your model grows exponentially and you have all your constants in one place. It's testable. It's swappable whenever you need to add something or um amend something in your constants.

[00:33:02]How do we add the variable? We we've been talking about the disjoint list variable. That's the math from the slide before. It's just in Python one to one with the formulation and students kaps.

[00:33:13]This is the whole decision space of the problem in one call.

[00:33:24]And eventually last but not least the objective function that's the backbone of the whole optimization that we have even in on on here right the three different sub parts of that objective function objective one minimizing the number of caps used objective two minimizing the leftover seats and objective three minimizing the total time traveled.

[00:33:52]So, three penalties, three different weights, and one minimize at the very end. And from here on, it's just the two lines that you saw before. You can lock the model sampler sample with a time limit and read back the cup states, cup root states, and you have your optimal solution in here.

[00:34:18]Amazing. And with that I'm I'm going to end the code example and we can talk a bit in the remaining round about five minutes 5 to 10 minutes is how do you get started with this? How do you get your hands on this? Right?

[00:34:34]Those are actually on on this slide links you can be clicking on afterwards as soon as we distribute those. So you can either come talk to us based on the region. We have a global organization pretty much everywhere. So you can talk with a quantum expert. We have our quantum launchpad program which is free trial access to to leap for a certain time frame on problems that are meaningful. So there is an actual application process behind that and I encourage everyone with a decent pro problem from their industry to apply on that describe their problem um what they're thinking about that what their benchmarking data looks like and so on and we are going through each and every application afterwards.

[00:35:18]So you can run the concepts from today on yourself provided that you have that use case and uh get the access granted.

[00:35:30]The next one is Dwave documentation that's oceans SDK um backbone more or less. So all the different solvers, all the different initiatives, all the different use cases are in here in in in that documentation. You can learn about the functions. You can learn about the whole motivation from an optimization standpoint behind that research initiatives. Um emerging technologies like ML and AI and the translation into quantum our surrogate model. So lots of interesting stuff in there that I encourage everyone um to go online and explore.

[00:36:10]The next thing uh dwave examples I I briefly mentioned that throughout the presentation that's the fastest route to your own app by cloning an example that you have in there. For example, on on that GitHub repository, we have an MVP problem solved. You can just clone that, amend your parameters or adapt it to your problem your your organization is facing and try it out. So, a bunch of examples, lots of interesting stuff in there. I encourage you once again to go and uh check it out.

[00:36:47]Last but not least, we have training initiatives. We have D-Wave learn as a team within the organization which is represented by experts in different fields that I have created training courses.

[00:37:04]So there are a couple of courses that are already available and a couple in the making but I can walk you through uh the ones we we are already having um and we are we are also going through. So if today's formulation slides felt a bit fast probably the one in the middle foundations for quantum computing um is your way to um get up to speed um that's the basics the med symbols the variables the quadratic models behind that is just to um get you started. The one on the right hand side quantum programming core is the more much more complex one and it's instructor supported. So we we have someone walking you through the different models. We are ho hosting um trainings hours answering all your questions. So I encourage you if you're keen on exploring that to go out there and register. And last but not least on the left hand side we have the introduction to quantum computing which answers most of the questions that we we also addressed at the very beginning in terms of what is a quantum computer what is a cubit? What is the physics behind that? How does analer actually work?

[00:38:09]What is a superp position? What are different quantum effects? And what's the difference between a kner and a gatebased system? So very interesting non that's not the technical course writing the code and so on but just to understand the quantum background behind that and with with that we're going to be reaching the Q&A session just to recap the idea of the of the whole thing.

[00:38:39]I'm getting a survey in here.

[00:38:40]>> Yep. All right. Thank you Daniel. Um really well done. We answered a lot of questions during the talk.

[00:38:46]>> Amazing.

[00:38:47]>> Uh but we would appreciate it if everyone would just fill out this webinar survey and then I will read to you some questions that um we thought would be of interest to everybody. So let's see how we're doing.

[00:39:08]We'll just give it one more minute here.

[00:39:18]All right, I'm going to Well, pe people are still still answering. So, >> I also see lots of movement in the chat.

[00:39:27]Amazing.

[00:39:28]>> Yes, exactly. All right.

[00:39:30]>> So, I'm going to end the poll now and thank you very much. We can share the results here.

[00:39:37]So, looks like overall we did very well.

[00:39:40]Daniel did very well with his talk. Uh it was relevant to your organization's needs and there are some people who would be of interest in following up which we will do.

[00:39:52]Okay. So here are some questions um that came in.

[00:39:57]Um how does a stride solver compare against CQM in your example? Is it always from tests presumably from the minimum objective that one would decide when to use stride versus CQM? If one would use strides, how would that reflect in computational cost?

[00:40:15]>> Great question. Um, so the CQM solver is actually the former generation of a hybrid solver. So what we always encourage people and recommend is to use the most recent generation which is the stride solver.

[00:40:30]In terms of performance, um, we we can actually also revert back to the slide that I showed at the very beginning introducing the stride stride over.

[00:40:38]there was a satellite problem that we've depicted on there. So our own benchmarking test um approved the best performance on stride and it's very native in terms of those combinatorial variables that we introduced. So it's the state-of-the-art solver that I would encourage everyone to to be using right now.

[00:40:57]>> Okay, here's kind of a detailed question. So follow me here. I am preparing a two-stage project to conduct a comparative analysis in a smallcale but controlled problem space. In the first stage, I aim to develop a fast and feasible solution in the classical environment to the traveling salesman problem with 20 nodes and 153 asymmetric edges. The goal in this stage is to reliably obtain a valid Hamiltonian circle encompassing all nodes by relaxing the optimal solution. In the second stage, it is planned to solve the problem using a quantum computation approach with the same data set. By comparing the results obtained from these two stages, the aim is to evaluate the solution quality and computational performance of classical and quantum methods. I have a heristic and first 15 steps and greedy last five steps classical solution. Uh and now I want to perform the quantum stage with the s uh school bus. uh solution that you mentioned today be suitable for my problem and that was a lot so excuse me sorry if you didn't get it >> no no no worries I I was also trying to d theest problem in in in the meantime um absolutely you can adapt that and one point because you were talking about hemotonians and you know like natively programming or thinking in cubos you don't necessarily have to do that with with stride so it's a very native way to embed your problem as as already mentioned in that um and and of course I would encourage you to to do that.

[00:42:35]One point to consider as well is the scale of the problem. Which means if the toy box example is too small that you use for the classical heristics with 20 nodes or whatever it may not make the the quantum piece shine if that makes sense, right? Like it's got to be complex enough. So what I encourage you to compare both approaches. So first of all programming it it was right, right?

[00:43:00]And the kind of formulation is a bit of a switch in thinking. That's the first thing. And the second thing is that you scale the problem by doing so with both approaches. And eventually we're going to come up with a diagram showing exactly that gap that we saw on the satellite benchmark.

[00:43:19]>> Okay. Thank you. Uh do you have any sample applications of quantum fora transform QFT and tensor network methods in finance? I'm particularly interested in practical examples or proof of concept implementations such as option pricing, portfolio optimization, value at risk, XVA risk aggregation or scenario analysis.

[00:43:40]>> Great great questions. Um we do have some examples on that. Um I'm not sure right now which one of them are public examples. Uh probably it's worth exploring the wave examples in in this regard. One thing to consider is though um the kind of variables you're going to be choosing, right? What hybrid solvers or quantum solvers usually struggle with are large spaces of continuous variables, right? So imagine speaking of portfolio optimization where the solver won't perform great is if you give it a huge range of continuous variables. If you let it choose, am I buying this stock or that stock from a range from one to 100,000 or whatever, that's probably not the way to go in terms of formulation. So what what I would be thinking about in this regards and this is what we we've done in terms of price optimizations problems also with uh customers on production is to discretise the problem in certain steps that are meaningful to for for for the business on the first hand hand side and on the second hand side give it a range right is any range meaningful or can I limit that solutions play space after the discretizations we are creating those steps and you give it a certain um range, the solver is going to take care of the rest. But we got to make sure that we give it some some boundaries so that we can make it more performant and more efficient.

[00:45:09]Okay. Uh we had a quite a few questions about the recording of this webinar. So it is being recorded. It'll be up on the D-Wave YouTube channel in the next day.

[00:45:17]We'll also be sending out an email to everyone who registered with links to the slides as well that you can have. Uh we have a few more time for just a couple of more questions. Um how do qaoa based solutions compare with quantum analing for optimization? Is it reasonable to view quantum analing as a more practical approach today while expecting a gradual transition towards kuawa and gatebased quantum optimization as fault tolerant quantum hardware matures?

[00:45:46]short answer to that um it's just different problems that you're choosing for the different technologies and that's there's lots of research on that so I would also encourage you to review the the papers especially in kneeling and Q&A examples that we have in there so I would just point you to the papers and there are lots of examples and benchmarks in terms of that >> okay let's see one more question we have some um continue to be answered by our uh experts online.

[00:46:19]Uh hi, I plan to use a D-Wave quantum kneeler for power system contingency analysis. Because access to the QPU is limited, most of my experiments will use D-Wave's local simulated analing sampler. How should I expect the results to differ between classical simulated analing and the real QPU? Particularly in terms of solution quality, sampling distribution, constraint satisfaction, and scalability.

[00:46:45]Great question. Short answer once again very different in terms of performance and also in terms of formulation because the simulated kneeling is not the real quantum computer, right? And that's not the hybrid approach that we are we we've been talking about as well. So based on the problem and this is what commercial clients usually use. Um I would go for most of the optimization problems with a hybrid solver because it proved best in in practice and uh do my benchmark um over there. And once again, the shift of thinking, it's not necessarily in in cubos. You don't have to embed it yourself. It's taken care of by the solver.

[00:47:22]>> Okay. And lastly, uh someone is asking for a link to some of the papers you mentioned. Um we will add that into the email that gets sent out to everybody um because it's too difficult to put it into the chat right here. So with that, I'd like to thank everyone for coming.

[00:47:37]Thank you our panelists and thank you so much to Daniel. I hope we you found this useful. Uh D-Wave typically has webinars every month. So if you look at our website, you will see what's up next. Uh and if you have any questions, you can just send an email to salesdwaveys.com and we will get back to you. So thank you so much.

[00:47:57]>> Amazing. Thank you for your attention, everyone.