Quantum speed-up
Quantum speed-up for AI.

Arti­ficial intelli­gence meets quan­tum phy­sics

Scientists at the University of Innsbruck have married two future technologies in a recent research paper. The researchers led by Hans Briegel were able to show how quantum computers can improve the performance of deep reinforcement learning methods, especially in large and complex environments.

Quantum computers can solve certain tasks much faster than normal computing machines. With first quantum computers expected to be ready for the market soon, many research groups are exploring possible applications. One area that is particularly exciting is the combination of two future technologies: Artificial Intelligence and quantum algorithms. One of the varieties of machine learning is reinforcement learning, in which an agent learns to solve a task through trial and error. The agent is nor directly told which action is best in which situation. Rather, it receives rewards at certain points in time and learns on its own the best strategy to collect them. By combining this method with deep learning approaches, such agents are also able to operate with very large and unstructured data. Deep reinforcement learning is used, for example, in the computer programs that first beat world champions at the Asian board game of Go. Now, a team led by quantum physicist Hans Briegel has investigated for the first time whether and how quantum algorithms can further improve the learning performance of this class of AI algorithms.

AI gets quantum speed-up

Drawing on ideas from statistical physics and combining them with known and new quantum algorithms, the researchers found "quantum enhancements" for several deep reinforcement learning algorithms. They discovered classes of neural-network based RL algorithms that achieve better learning performance than conventional methods when the tasks have large state and action spaces. “We achieve this by using computationally intensive learning models, which – as we have been able to show – can be processed much faster by quantum computers,” says Sofiene Jerbi from the PhD program Atoms, Molecules and Light at the University of Innsbruck. “Some of the quantum algorithms we developed could be executed on quantum computers that will be available in the near future.” 

The study in the journal PRX Quantum was financially supported by the Austrian Science Fund FWF and the Austrian Academy of Sciences, among others. 


Nach oben scrollen