Objective

We are investigating the suitability of quantum processors from the DLR Quantum Computing Initiative (QCI) for the implementation of quantum algorithms for reinforcement learning.

Machine learning, and in particular reinforcement learning, is becoming increasingly important in the digital world. In reinforcement learning, what is referred to as an agent performs various actions to solve a given problem. The agent receives a reward for successful actions. The agent can use the knowledge thus acquired to develop optimal solutions to problems in areas such as navigation, biology and medicine, security and energy. Quantum algorithms can accelerate this search for rewarded actions and thus speed up the learning process. This is why we are investigating the suitability of quantum processors from the DLR QCI for implementing quantum algorithms for reinforcement learning.

Motivation

Both machine learning algorithms and quantum computing are revolutionising information processing. While the techniques referred to as ‘supervised learning’ and ‘unsupervised learning’ primarily aim at finding patterns in large amounts of data, ‘reinforcement learning’ is more about finding a constructive solution to a given problem. However, problem solving using reinforcement learning is usually very time consuming.

Quantum algorithms accelerate this learning process – we have shown the working principle for this theoretically and experimentally in recent publications. However, many challenges still need to be overcome before reinforcement learning with quantum algorithms can be implemented for solving real problems, such as those in the field of navigation.

Challenge

For reinforcement learning with quantum algorithms, we need low-error quantum processors and high-quality quantum algorithms. However, currently available quantum processors are still severely limited in the number of quantum gates that can be executed in succession until the resulting errors become overwhelming. Optimal implementation strategies for quantum algorithms are therefore required in order to take the next step towards real applications.

In addition, many questions about the application of quantum algorithms for reinforcement learning are still open. Quantum algorithms are subject to restrictions such as wave-particle duality and the no-cloning theorem. We have to take these restrictions into account when developing problem-specific quantum algorithms for reinforcement learning. Through the DLR QCI, we have the unique opportunity to develop quantum-assisted reinforcement learning with hardware-software codesign.