The development of mobile wireless technology, whether 4G or 5G, as well as the introduction of the concept of Industry 4.0, have all added to the design complexity associated with wireless systems.
Not only have wireless networks become more complex but they have also become more difficult to manage, due to requirements that require the optimum sharing of resources.
“Consequently, Artificial Intelligence (AI) is being used across a growing number of verticals – wireless being just one of them,” explained Houman Zarrinkoub, the principal product manager at MathWorks, who is responsible for wireless communications products.
“But when you consider wireless, it is confronted with several significant challenges with which AI is well suited to help. The number of devices, people and satellites that are now connected mean that resources need to be optimised – scale, for the industry is a massive problem.
“Another issue is controllability. The dynamic nature of wireless needs to be better managed especially as expectations grow and users expect a constant, reactive and dynamic system. In terms of the Internet of Things, AI also has a very important role to play.”
These numerous challenges are forcing engineers to think beyond traditional rules-based approaches and many, despite their concerns, are now turning to AI.
“It is very much seen as the go-to solution to address the challenges introduced by modern systems,” according to Zarrinkoub.
Whether it’s managing communications between autonomous vehicles or handling the allocation of resources to better manage a simple mobile phone call, AI is providing a level of sophistication that is necessary for modern wireless applications.
“As the number of devices connecting to the wireless network increases, so engineers will have to be prepared to introduce AI into increasingly complex systems,” explained Zarrinkoub. “They will need to know the benefits and current applications of AI in wireless systems, as well as the best practices necessary for optimal implementation. This will be key for the future success of the technology.”
However, the apparent removal of the human factor has triggered a degree of resistance at ground level.
“According to critics a big issue with AI is that it tends to lack what they refer to as explain-ability,” suggested Zarrinkoub. “We need to find better ways of undertaking and explaining internal state monitoring so that AI techniques are more explainable and visible – turning the black box into a white one!”
According to Zarrinkoub, once these kinds of issues with AI are addressed, so the anxiety or reluctance to use the technology will be removed and, “people will become more comfortable with using it.”
Embracing AI
Data size and quality are certainly critical if AI models are to be effectively deployed, as those models will only be as good as the data they are trained with.
“AI needs to be capable of coping with a range of real-world scenarios, and models will need to be trained using a broad range of data,” explained Zarrinkoub. “That means looking at static and dynamic data as well as considering outlier data, not just the average.
“When it comes to AI the focus has to be on the quality of data used, but both security and privacy issues will also need to be considered. What’s so positive about the situation at present is the way in which the broader wireless community is working together.”
For engineers the transition to 5G has brought about the optimisation of speed and quality of mobile broadband networks, as well as the need for ultra-reliable low rates and massive machine-type communication for time-sensitive connections.
“Consequently, we are seeing AI adopted more broadly. As devices compete for the resources of the network, formerly linear patterns of designs once understood by human-based rules, cease to be adequate,” said Zarrinkoub.
AI techniques are certainly better at solving non-linear problems by extracting any pattern automatically and efficiently, beyond the ability of human-based approaches. AI can recognise patterns within communications channels that link devices and people and enables systems to then optimise the resources given to that link to improve performance.
“Simply put, running a network for those disparate use cases without exploiting AI methodologies would be a near impossible task,” according to Zarrinkoub.
AI also brings project management benefits. Incorporating simulated environments into an algorithmic model through estimating the behaviour of source environments will help engineers to quickly study a system’s dominant effect using minimal computational resources.
“This leaves more time to explore design and carry out more iterations faster, reducing cost and development time,” said Zarrinkoub.
By synthesising new data or by extracting them from over-the-air signals, applications like MathWorks’ 5G Toolbox provide the data variability that’s necessary for 5G network designers to then train AI.
“As I explained earlier, it’s critical that we explore a large training data set and iterate on different algorithms based on that data, otherwise you can end up with a narrow local optimisation instead of an overall global one. As such a robust approach to testing AI models in the field should be considered critical to success,” according to Zarrinkoub, who concedes that the variability in signal needed for testing AI techniques can be a problem.
“Signals captured in a narrow, localised geography may adversely impact how an engineer might optimise the quality of their design.”
A Wireless World
Digital transformation in areas like telecommunications necessitate the use of AI as applications like smart cities, telecommunication networks and autonomous vehicles (AV) are connected.
“As they are connected so the resources of the network joining these applications will become increasingly stretched,” said Zarrinkoub.
When it comes to telecommunications, AI is deployed at two levels – at the physical layer (PHY) and above PHY - the application of AI for improving performance in a line connecting two users is referred to as operating at PHY.
The application of AI techniques to the physical layer includes digital pre-distortion, channel estimation and channel resource optimisation, as well as automatic adjustments to transceiver parameters during a call otherwise known as autoencoder design.
Channel optimisation is the enhancement of the connection between two devices, notably network infrastructure and user equipment. Often, this means using AI to overcome signal variability in localised environments through techniques such as fingerprinting and channel state information compression.
“AI is being used to optimise positioning and localisation for wireless networks by mapping disruptions to propagation patterns in indoor environments, caused by individuals entering them. AI then estimates based on these individualised 5G signal variations the position of the user,” explained Zarrinkoub. “AI is vital in finding the most efficient and effective path between a base station and a user.”
Meanwhile, channel state information compression is the use of AI to compress feedback data from user equipment to a base station, ensuring that the feedback loop informing the station’s attempt to improve call performance does not exceed the available bandwidth which can result in causing a dropped call.
Above-PHY uses are mainly in network management and resource allocation.
Applications, such as scheduling, beam management and spectrum allocation, are functions that manage and optimise the resources of core systems for the competing users and use-cases of the network and network designers are now turning to AI techniques to respond to allocation demands in real time.
“As the use cases for wireless technology expand, so too will the need to implement AI within those systems,” concluded Zarrinkoub. “Without it, systems such as 5G, autonomous vehicles and IoT applications would not have the sophistication necessary to function effectively.”
While AI’s place in engineering, particularly wireless system design, has been increasing in recent years as the use cases and the number of network users grow, so the use of AI will only increase going forward.
