I’m truly excited today to bring you this interview. When I first read about Volkswagen and D-Wave teaming up on traffic optimization, I wanted to get to the team to learn more about exactly what they were up to. Traffic optimization is exciting, to be sure. But high-performance batteries for electric vehicles? Vehicles with artificial intelligence? That’s some hot stuff, right? So I wanted to get this talented team to answer some questions, and boy did they deliver. Florian, Gabriele, Christian, and David are all awesome to work with. They’re so in sync, they responded as a team to almost all of my questions.
We have always been evaluating emerging technologies, so the first commercially available quantum computing devices, of course, sparked our interest. Although we maintain partnerships with universities and research groups, and some of our team members have an academic background in quantum computing, demonstrating how our customers or we can benefit from quantum computing in practical applications has always been our primary focus. The broader interest within Volkswagen emerged when we were able to show that problems potentially affecting millions of people in their daily life, such as the traffic flow, can be addressed using a quantum computer.
This initial project gave us confidence in the potential of these machines to solve efficiently a wide range of optimization problems, and motivated us to explore additional applications. Our current investigations focus on the topics of transportation and logistics, materials science, and include the study of quantum-assisted algorithms for machine learning applications.
Our first project focused on a simplified version of the traffic flow problem, where the optimization target was the minimization of road congestion in a real-world scenario based on the city of Beijing. This work was successfully presented at CEBIT 2017 and recently published in an article on Frontiers in ICT. Using a D-Wave QPU, we were able to calculate an individual optimal rerouting for each car that also takes into consideration the path of the other cars in the city.
However, this version of the quantum-assisted traffic flow optimization incorporates only a limited set of cars, no communication to infrastructure, no other traffic participants, and no other optimization targets except the minimization of road congestion. In our future work, we intend to consider all of these parameters, and we will also need to consider creative ways of formulating them as part of the optimization problem.
The main goal of this project was to map a real-world problem to a quantum annealing machine, and we demonstrated how this can be successfully done. When evaluating the solutions produced by the D-Wave QPU, the focus was on finding good quality solutions within short periods of calculation. To quantify the quality of a solution, we counted the number of congested roads after optimization. We are interested in an optimization interval of about 1 – 5 seconds, since this is the characteristic time frame in which connected cars send their position data. We found out that this is a sufficient time for the QPU to get good quality solutions, significantly reducing the overall road congestion. Based on this observation, we believe that future QPUs with increased computing power will play an important role in applications to time-critical optimization problems like this.
We are in a very early stage with quantum chemistry, and before we are able to simulate industry-relevant materials on a quantum-physical layer by using quantum computers, we will simulate smaller systems to show that it works, and approach more complex ones over time. We have teamed up with Google and are working on their universal quantum computer. No one simulated such complex systems with a quantum computer before, but what we are after are better materials for cathodes/anodes, enhancing durability, reducing charging duration, increasing how much charge the battery can store.
There is certainly a big potential, and our opinion is that quantum computers will strongly benefit materials science. The idea is that only with quantum systems we can accurately simulate quantum physical systems in full complexity.
Machine learning allows software to learn from experience or observe scenarios and then choose the best action in a given state, which is similar to how humans learn and behave. Behavior can be governed by static rules only when the environment and allowed actions are limited, but if an environment is dynamically changing and complex, such as in the SDC-scenario, this is not possible anymore. So what we do, for example, is clone the behavior of a skilled driver by collecting everything that the car senses in a given situation (LiDar, ultrasonic, radar, cameras), and feeding it into an algorithm that then tries to mimic the driver’s behavior in this situation (acceleration, breaking, steering angle, lane change, full stop, etc.).
This is a strongly simplified explanation, as we do a lot more to ensure an SDC behaves as it should. In terms of machine learning, we expect quantum computing to allow for the processing and optimization of more complex models and to enable access to problem classes that have not been approachable before.
We recommend having a quantitative background, such as in computer science, mathematics, or physics. Although linear algebra, the language of quantum computing, is not difficult, some concepts of quantum computing are fundamentally different from classical computing, and being able to deal with some of the more sophisticated mathematics of quantum physics definitely adds to the general understanding. It also depends on the area one is interested in. So if the interest goes towards solving quantum chemistry problems, then a background in that area is certainly helpful, as is a background in machine learning if one is interested in quantum-assisted machine learning.
We believe a good way to start with anything is always a mix of theory and practice, and there are some quantum simulators out there that certainly help to understand the theory. We would also recommend reading up on the differences between quantum annealing and gate model quantum computing, going through the basic algorithms, like Deutsch-Joszya, Bernstein-Vazirani, Simon, Grover, and searching arXiv for quantum machine learning/chemistry papers.
Honestly speaking, we feel that analogies in quantum physics are very often misleading, and we try to avoid them when possible. The most important effects in quantum computing are superposition, entanglement, tunneling, and interference. We can explain superposition as the fact that as long as an observer does not “look” at a quantum system, it can take all of its possible configurations at once. In the case of a quantum bit (qubit), this means that instead of just being either ‘0’ or ‘1’ as a classical bit, it can represent both values at once.
If a computer operates a calculation on this qubit, like a logic gate transformation, the outcome will be the superposition of the possible results of that calculation for the different input values ‘0’ and ‘1.’ With only one qubit, this is not impressive, but with 1000 qubits this means we can evaluate 2^1000 results of a single calculation at once. One could picture this as a form of very powerful parallel computation, and we believe this popular analogy contributed very much to the large interest towards quantum computing in the public.
Entanglement is probably the most difficult effect to explain in detail, as its discussion is tightly connected to the interpretation of quantum mechanics. We usually simply refer to it as a strong interconnection between quantum systems: if two qubits are entangled, then anything we do to one of these also affects the other.
With tunneling, there is an analogy we like: let’s assume we have a problem and distribute different possible solutions to this problem over a mountainous landscape. Let’s further assume that good solutions can be found in the valleys, and bad solutions on top of the mountains. A situation like this can arise, for example, when we translate an optimization problem into the minimization of a cost function related to one or more optimization targets. In this picture, the values of the cost function are spread around the landscape, and the objective of the optimization procedure is to find the lowest value. We could start from anywhere on that landscape and walk down to the next valley to find a good solution.
However, it may be that this valley is not the deepest valley in the landscape, and that the associated solution is not the best we can possibly find. We can think of tunneling as an effect that allows “flying” through all the mountains directly to the deepest valley. With a classical computer, and hence without tunneling, we would have to walk the whole surface to find the best solution.
Finally, when trying to explain interference, it helps thinking about waves. Interference can be constructive or destructive and most of the times we are concerned with correcting or getting rid of it, as it may negatively influence the results of a calculation. Imagine two waves hitting each other: if both are at the peak when they hit, they will positively interfere with each other and the result will be a proportionally higher wave. If one is at the lowest point and the other at the peak at the time they hit each other, they may cancel each other out. This is also what happens in quantum computing: any time qubits communicate with each other, charge is transported, and moving charges cause magnetic fields. These magnetic fields may interfere with magnetic qubits, and negatively influence the calculation.
Quantum computers will not replace classical computers, as many problems can already be solved efficiently using existing machines. However, they will augment the classical (non-quantum) world of computing, and we need to carefully decide which parts of a problem are suitable for being processed on a quantum computer and which not. All of the algorithms we are currently working on come with a significant classical component.
Yes, especially programmers who have never written assembly code (as most of us) may struggle a little with having to address every single bit, as it’s necessary for some quantum applications. Moreover, programming languages for quantum computers are still in their very early development stages and offer limited functionalities in comparison to traditional imperative languages used for classical computers. This often requires formulating problems in a creative and original way to make full use of the features of the hardware. In addition to that, it may take a while to understand and efficiently use the quantum effects that these machines make available, and for which there is no classical counterpart.
One misconception is that quantum computers will replace classical computers. That will most likely never happen, as we just do not need to use a quantum computer for every problem. Another misconception is that quantum computing is of no use. It is important to emphasize that we are in a very early stage for practical applications of this technology, several problems are still unsolved, but over time we will see which areas will benefit the most from quantum computing.
We think there is nothing that can completely derail quantum computing. One thing that may slow it down though is hardware development. Only if we are able to produce error-corrected quantum processing units, scaling up the number of qubits makes sense – otherwise we suffer from an exponential increase of error probability.
Florian: I have always been impressed with John von Neumann, who contributed so much to quantum computing, mathematics, and physics. And Richard Feynman, of course. Both seemed to have an understanding of the world that goes a lot deeper than just describing it by formulae. I have had the chance to meet and work with so many brilliant people, so it is really difficult for me to pick one specific person, but I am definitely impressed by the work of the Google quantum AI lab, the QUTIS group, and the people at Rigetti.
Gabriele: I’ve always been interested in understanding how things work; I believe studying Galileo’s scientific method in high school was the first step that motivated me to focus on physics. Later in university, learning about the contributions of Italian scientists like Fermi, Cabibbo, and Maiani inspired me to get a Ph.D. in particle physics and pursue a career in research, that finally brought me to VW and gave me the chance to join the quantum computing team from its beginning. I’m fascinated by the work of the colleagues at D-Wave, I like their approach to quantum computing and was very impressed by their scientific research when we worked together.
Christian: That’s a difficult question . . . I guess I always was interested in science, otherwise I would have followed another path. During my studies, I found people like Donald E. Knuth, Noam Chomsky, Jon Kleinberg, or Esko Ukkonen very inspiring. Since I’d like to bring the Computer Science part of Quantum Computing more into the public focus (too often this is still seen as a pure physics topic), I work and talk a lot with computer science departments at various universities. The spirit in general right now, bringing quantum computers and computer science together, is very exciting. It’s pioneer work. It’s super cool!
David: Claude Shannon’s work as the founder of Information Theory is inspirational to me in the sense that it intersects so many fields, such as statistics, physics, mathematics, and computer science, with specific application within the sub-fields of machine learning and AI. Quantum information theory and machine learning present an opportunity to expand on many of the elegant ideas that originated with Claude Shannon. I also continue to be impressed with the research coming out of groups at the University of Toronto, The Barcelona Supercomputing Center, Sandia Labs, D-Wave, NASA, and Google.
Dr. Florian Neukart is a Principal Scientist at Volkswagen’s Continuous Optimization and Digital Engineering (CODE) Lab in San Francisco, where he works with David Van Dollen. Dr. Gabriele Compostella and Dr. Christian Seidel are Data Scientists in Volkswagen’s Data Lab in Munich.
Listen, I have to take a short break from the quantum commentary to address the rumors, e-mails, Facebook messages, tweets, and InMail (and that one snap chat someone sent; you know who you are). You all know by now I’m fascinated with quantum computing, because I never shut up about it. I figure it’s time to do something about that. Which is why I’m taking a couple weeks off to launch my own entry into the quantum computing fray.
I think we’ll soon see a quantum computer demonstrate quantum supremacy in a way that we can all take to the bank. However, physicists and engineers in labs around the world are still struggling to overcome the hardware challenges. Quantum software applications might as well be an endangered species. Richard Feynman, the physicist credited with the idea for a quantum computer, once quipped, “By golly, it’s a wonderful problem, because it doesn’t look so easy.”
Last week I shared with you that mathematician Gil Kalai doesn’t believe quantum computing is possible. He says the math just isn’t there to ever achieve reliable, practical error correction. Not that Dr. Kalai is alone in this belief, but there sure are a lot of major players that seem to think we’re on the verge of one of the biggest shifts in computing in the last 100 years.