In the age of cloud, big data, IoT, and mobile devices, wireless networks are faced with massive capacity and coverage requirements, and increasingly complex usage scenarios. The network industry has responded to this challenge by designing novel wireless technology into the radio access network layer, e.g., cognitive radio/software-defined radio (CR/SDR), as well as new architectures such as software-defined networking (SDN) and network function virtualization (NFV). Machine learning (ML) has been recognized as a powerful tool to address these requirements and thus has gained significant attention in 5G and upcoming 6G systems. This article will focus on the potential use cases of AI in wireless networks, as well as some challenges that remain to be addressed before widespread adoption of AI solutions in wireless communications becomes possible.
One of the reasons for the buzz around AI for wireless networks is that fifth generation (5G) and sixth generation (6G) networks bring about substantial technological innovation at scale. On the other hand, the network management architectures, operation methods, mechanisms of user plane protocols, device types, service types, or interference scenarios are not fully defined or field-tested yet.
On top of that, there are currently limited standards documents to evaluate implementation results either. Thus, developers must be flexible enough to absorb their own experiences with algorithms as time goes by and solve problems one step at a time. Network complexity raises another difficulty because it is hard to find an optimal solution even if we adopt an advanced algorithm. For example, it might be impossible to optimize every part of a communication system due to limited computational power. And it might also be difficult to estimate parameters accurately since they can vary from place to place or time period to time period.
Artificial intelligence offers solutions for such problems because AI is capable of learning from data automatically and can self-optimize based on its own analysis. This article introduces several use cases for AI in wireless communication systems: spectrum cartography using statistical learning; resource optimization and resource management using deep learning; network optimization using reinforcement learning; self-optimizing transceivers; etc., along with challenges related to AI development for 5G and 6G communication systems (e.g., availability of large-scale datasets).
Deep learning for Spectrum cartography and resource optimization
Deep learning uses a model architecture of interconnected, configurable computational units (represented via artificial neural networks). These units take input signals, compute weighted sums of their inputs, and provide an output based on that sum. Deep neural networks (DNNs) can learn to extract features from input data with multiple levels of abstraction. This characteristic makes DNNs applicable to many problems.
For example, they can be used for spectrum cartography: generating statistical models that summarize how radio signals propagate through space under different propagation conditions. Because propagation is nonlinear, it is difficult to characterize by analytical formulas; but it is easy for a DNN to learn its characteristics by observing examples.
Similarly, deep learning has been applied to wireless resource optimization. In particular, deep reinforcement learning (DRL) has been applied to optimize link scheduling in wireless networks using Q-learning techniques. A 5G network may use hundreds or thousands of antennas at each base station; so managing them efficiently will require AI techniques such as DRL mentioned earlier.
Network optimization using reinforcement learning
There are several wireless communication systems that are dominated by a few large nodes, e.g., AP (access point) in cellular networks or base stations (BSs) in WLAN networks. In these cases, it is difficult to fully utilize all available transmission opportunities for high throughput due to the limited amount of capacity.
For example, when multiple APs send packets over a single channel, they usually transmit at sub-optimal power to avoid interference between each other. However, it is possible to improve overall network performance using reinforcement learning techniques if we can calculate actual throughput data on given links or wireless channels.
The key idea behind reinforcement learning is to learn an optimal policy from experience without any prior knowledge about system dynamics. In particular, two major challenges remain: how to efficiently collect throughput data for training purposes and how to use it effectively for policy optimization.
To address these challenges, BS can learn an optimal policy with respect to its own throughput targets based on past experience with neighboring BSs’ policies via local interaction with them during normal operations without interfering with their communications. The underlying algorithm consists of three main steps: firstly, BS collects past measurements from neighboring BSs; secondly, it constructs a model based on collected data; finally, it optimizes its own policy according to model parameters while considering the potential impact caused by neighbors’ policies on its own targets.
AI-driven Self-optimizing transceivers
Self-optimizing transceivers are able to learn from their own radio behavior and surroundings, allowing them to improve their performance over time. For example, a self-optimizing transceiver might use information about other local radios’ interference levels to adapt its own multi-dimensional settings, using machine learning algorithms, for maximum data rates or minimum interference.
It might also learn about its physical environment by analyzing radio behavior when passing through different parts of a building or down different streets in order to dynamically adjust its settings based on where it’s located. The resulting gain in overall network capacity can be significant as devices take advantage of lower interference regions without relying on human configuration.
Self-optimizing resource management: With artificial intelligence, wireless systems can now optimize resource allocation automatically. Previously, networks would only make adjustments to resource allocation (e.g., frequency band) if they received explicit commands from humans—not so with AI! Now networks can decide how best to allocate resources based on real-time conditions such as traffic load and spectrum availability.
AI for NP-Hard Network Problems
One of AI’s strengths is that it can help find solutions to NP-hard problems through various methods, including sampling. For example, if you want to route a network traffic flow across a wireless network while ensuring some quality of service (QoS) requirements are met, it might take days or even weeks for a human operator to solve.
In contrast, AI algorithms can compute an optimal solution by taking samples of what happened in thousands or millions of similar real-world situations over a few hours. For example, when AT&T wanted to optimize their device-deployment pipeline with AI so that new cell towers could be built more quickly and efficiently, they found they could build six times as many towers using AI than without it.
AI can also reduce operating costs. For example, AI technologies have been used to predict which devices will need maintenance before any humans have noticed anything wrong—and then schedule maintenance visits automatically. This means using ML algorithms we can spend less time on mundane tasks like scheduling and focus on work that requires our creativity and problem-solving skills. The same goes for other industries like manufacturing: AI has been used to predict when machines need maintenance before any humans have noticed anything wrong—and then schedule maintenance visits automatically, reducing downtime from 30 percent to just 2 percent in certain cases.
Adoption of AI in 5G and 6G Networks: Challenges and Opportunities
There are four main challenges to the widespread adoption of AI for communication networks:
- Communication networks present different kinds of data compared to other domains like e-commerce, games, or video, making operationalizing and integrating AI into the network management stack a challenging endeavor.
- In wireless communication systems, we need real-time data and AI processing with low latency. This is still challenging at present because AI frameworks, including computing resources with optimized power consumption, aren’t designed for such high-speed use cases yet.
- Security is a big issue here. If deployed incorrectly, a self-optimizing network could threaten security. Automation without human intervention might be too dangerous to make sense at first—we may need some time before people trust automation enough that they rely on it without human oversight.
- Finally, artificial intelligence requires huge amounts of training data for machine learning and deep learning algorithms to work well. It’s not clear where 5G and 6G will get their training sets from. Training sets for AI can be incredibly large; for example, Google’s DeepMind learned how to play Atari 2600 video games using 16 million frames from each game as its training set! Getting training sets together won’t be easy in the constantly changing wireless systems.
On the flip side, there are also many benefits to AI for communication networks:
- We’ll see improved performance thanks to increased intelligence in our communications infrastructure.
- AI offers more fine-grained control over network resources, meaning we’ll be able to better optimize usage of those resources by monitoring traffic in real-time and adjusting our strategies accordingly. With proper development, AI could help us achieve much greater efficiency than ever before possible with manual controls alone.
- We’ll see improvements to things like emergency response times through better prediction algorithms that anticipate demand spikes during natural disasters or emergency situations. And there are plenty more ways AI can improve communications systems overall.
Final Thoughts
In this article, we have introduced a few use cases for artificial intelligence in wireless networks. We have also discussed some of the challenges related to developing AI-driven wireless communications applications. Despite the challenges, AI is a promising technology for wireless communications and has the potential to enable a number of new and innovative applications. In the next generation of communication networks, AI will play a key role in meeting the ever-increasing demands placed on these networks. Further reading: Case studies of AI in the networking industry.
What do you think about the use of artificial intelligence in wireless communications? Do you think it can help address some of the challenges that these networks are currently facing? Let us know in the comments below. Thanks for reading.
