A new, revolutionary era of data processing is about to dawn in the mathematical optimization of various parameters and the optimization of subsequent processes. Is this pure science or who stands to benefit in the future?
Axel Plinge: As part of BayQS, the Bavarian quantum hub, we are working to optimize existing AI machine-learning approaches by means of quantum algorithms. Armed with the potential of quantum computing, AI-supported optimization finds more precise solutions – and is significantly faster.
But, given how differently a quantum computer processes data compared to a classical computer, it’s also essential to translate classic AI algorithms properly into quantum-AI algorithms. This allows them to be performed on a quantum computer and enables scientists to develop completely new algorithmic approaches.
How can the different data processing method of a quantum computer be used for certain technology optimizations?
Axel Plinge: Our goal is to improve mobile applications, say, by enhancing their positioning accuracy and their use of suitable frequencies or methods in the case of sensor fusion, for example. In mobile networks, 5G technology – and, in due course, 6G technology – offers new programming and software options, yet also poses new challenges. These include positioning methods when measuring the duration and angle of radio signals. Positioning parameters are particularly challenging in scenarios with shadowing effects and multipath propagation. In this case, multi-antenna systems are a promising approach.
Many subproblems are extremely complex in higher frequency ranges and distributed systems. This applies to measurement, configuration and signal processing in equal measure. Owing to the size of the system and large number of dynamic parameters, previous solution processes have proved to be computationally intensive, making them unsuitable for time-critical applications.
Your team is focusing particularly on optimizations using algorithms for reinforcement learning, in other words, learning in dynamic environments. How do you plan to deploy quantum computers here?
Axel Plinge: Compared to standard algorithms, the underlying optimization problems can be resolved more efficiently using artificial intelligence techniques. Our team is therefore looking at self-learning AI methods for dynamic scenarios, especially reinforcement learning.
Reinforcement learning is automated learning from experience in dynamic environments. An adaptive decision algorithm – the learning AI agent – interacts with its environment, which adequately represents the problem that needs to be addressed. A problem-specific reward signal enables the agent to gradually improve its decision-making strategy.
How can we understand this AI-based strategy in more vivid terms?
Axel Plinge: Some may be familiar with Go-playing programs or AI that can beat any human at old computer games hands down. Quite simply, I always say that if the AI agent spent billions of hours practicing Formula One races in a simulator, no racing driver would stand a chance.
Preliminary approaches to quantum reinforcement learning are already in place. Nonetheless, the projected runtime speedups of many current approaches are moderate, while the demands placed on existing quantum hardware still tend to be too high.
That’s why we are also testing hybrid quantum algorithms and approaches to quantum machine learning. Today’s quantum hardware is characterized by low numbers of qubits and high error rates in executing algorithms. It’s already possible to explore the potential for improvement with hybrid algorithms; that is, the combination of conventional AI/machine learning and quantum computing.
What use cases are you currently analyzing that enable the combination of AI/ML and quantum computing to really come into its own?
Axel Plinge: We’re studying this combination in 5G wireless technology applications and automated industrial environments. Teaming up with industry partners, our goals are to optimize production planning, optimize control based on reinforcement learning in process manufacturing and implement distributed automation systems in the smart factory.
One example in production planning is defining the manufacturing chain in real time, perhaps by determining the sequence of finished cars to pass through the painting process to fulfill incoming orders.
To put it another way, the idea behind control optimization is that the AI agent helps adjust all the variables during system operation. Practical examples range from oil rigs right to the production of peanut puffs. By facilitating the optimal manufacture of different products according to customer specifications, the smart factory goes a long way toward embracing this concept.