DeepMind’s AlphaZero has taught itself how to control powerful quantum computers

Quantum computing and Artificial Intelligence (AI) are born bedfellows, but despite this chatter about Quantum Artificial Intelligence (QAI) so far has been very muted. Researchers and companies throughout the world are constantly working on developing this technology, which can solve extremely complicated problems that are too advanced for classical computers, such as how to use AI to control plasma in a state of the art Fusion Reactor ...

One such group working on developing a quantum computer is at Aarhus University, a research group led by Professor Jacob Sherson who, following on from the use of AI to control a fusion reactor, used the same computer algorithm - AlphaZero from DeepMind - to actually control an entire quantum computing system to improve its efficiency.

The Future of AI, by AI Keynote Speaker Matthew Griffin

Quantum computers utilize quantum mechanics, which is a branch of physics that focuses on the smallest building blocks of our universe. One of the fundamental rules is that a system can exist in more than one state at a time. These rules get translated into computer language, and a quantum computer is able to perform multiple calculations at the same time. This means that a quantum computer can perform much faster than regular computers.

The theory of quantum computers has been established, but there has yet to be a full-scale quantum computer created.  AlphaZero is capable of learning on its own without any interjection from humans. Because of this, the algorithm has been able to defeat both humans and complex computer programs in difficult games like Go, Shogi, and Chess. AlphaZero was able to do this by competing against itself and improving over time.

The algorithm was able to beat the leading chess program Stockfish after playing against itself for just four hours. After that impressive performance, Danish grandmaster Peter Heine Nielsen compared AlphaZero to a superior alien species.

The research group at Aarhus University has used computer simulations to demonstrate how AlphaZero can be applied to three different control problems which could possibly be used in a quantum computer.

“AlphaZero employs a deep neural network in conjunction with deep lookahead in a guided tree search, which allows for predictive hidden-variable approximation of the quantum parameter landscape. To emphasize transferability, we apply and benchmark the algorithm on three classes of control problems using only a single common set of algorithmic hyperparameters,” according to the study.

The research done by the team was published in Nature Quantum Information.

Lead Ph.D. student Mogens Dalgaard spoke about how the team was impressed with AlphaZero’s ability to quickly teach itself.

“When we analyzed the data from AlphaZero we saw that the algorithm had learned to exploit an underlying symmetry of the problem that we did not originally consider. That was an amazing experience.”

The real breakthrough came from pairing AlphaZero, which is an extremely impressive algorithm on its own, with a specialized quantum optimization algorithm.

According to Professor Jacob Sherson, “This indicates that we are still in need of human skill and expertise, and that the goal of the future should be to understand and develop hybrid intelligence interfaces that optimally exploits the strengths of both.”

The group wants to quicken the pace of development within the field, so they released the code and made it openly available. The move generated a lot of interest.

“Within a few hours I was contacted by major tech-companies with quantum laboratories and international leading universities to establish future collaboration” Jacob Sherson said, “So it will probably not be long until these methods will find use in practical experiments across the world.”

DeepMind is a UK-based Google sister-company that is responsible for both AlphaZero and AlphaGo. These systems are now showing their importance in other areas, including quantum computing.