The late 2022 launch of OpenAI’s ChatGPT and subsequent enhancements through GPT-4, changed the world’s perception of the maturity and potential of artificial intelligence (AI). Governments are now writing regulations, industry is developing technology and business models and academics are probing the latest research topics. Driven by a chatbot built on a large-language model (LLM) created by the transformer architecture, this tidal wave of activity is hiding many practical AI developments that are more relevant to radio systems in mobile wireless.

As early as 2019, the ITU-T’s Network 2030 Focus Group (FG Network-2030)¹ highlighted the necessity of AI from the physical to the application layer of the 6th generation of wireless networks to accommodate the demands of realizing their vision. Thus, the 6G vision has always included AI as a fundamental building block and tool. However, LLM’s AI models, trained on the vast database of text on the global internet, are not the solution for overcoming many technical challenges in wireless communications. Since FG Network-2030, there have been myriad journal articles, research papers, technical demonstrations, early standards work and commercial solutions illustrating that the intensity of this work is focused on machine learning (ML) unrelated to LLMs. In most cases, language models will not be adequate for wireless and particularly radio technology. Instead, they will require models trained on other sources of data such as radio I/Q pairs, signaling traffic or user (payload) data.

ML-based optimization is the subject of research and development at all layers of the wireless network. Allow me to examine just a snapshot of the work closest to the physical layer. This includes AI as applied to the “air interface.” An excellent early overview² delineated novel concepts using ML to:

• Create custom waveform and modulation optimized for transceiver hardware and real-time channel conditions
• Design and train transceivers themselves based on use-case specifics
• Allow these capabilities to moot PHY- and MAC-layer standards themselves by standardizing only how ML would be implemented for such bespoke and real-time determination of the air interface
• Take steps towards semantic communications rather than optimizing only how to transmit bits.

A panel of experts in 2021³ were asked if AI was already used in contemporary commercial wireless and the answer was a resounding “yes.” AI was already being used in 4G and 5G applications including traffic load balancing and signaling optimization, MIMO precoding algorithms, energy-use management and network planning. The intent is to expand the use of ML to drive improvements in system behavior that have become so complex that conventional solutions are constrained by their deterministic one-size-fits-all mathematical models.

For ML to drive large-scale reliable and viable improvements in performance, quality of service and even quality of experience, the challenges we must address fall into three categories:

Known weaknesses: The October 2021 issue of IEEE Spectrum⁴ featured a cover asking: “Why is AI so Dumb?” Charles Q. Choi’s article therein described seven “ways AI’s fail.” These included brittleness, embedded bias, catastrophic forgetting and perhaps the most challenging issue: a lack of both “explainability” and “common sense.” I have read about recent progress in addressing the former but neural-network ML suffers from a lack of explainability when answers are “right” or “wrong.” From an engineering perspective, understanding the “whys” is essential to reliable and viable solutions. One can see “common sense” manifest daily in the inanity of AI-generated news articles and some of the false and ridiculous answers to questions posed to LLMs.

Data and model validation: Training AI models requires tremendous amounts of data that fits the balance of being “random enough” (unbiased, uncorrelated) while also being “appropriate enough” (relevant to solving the problem at hand). Clean and controlled training data has thus become a premium commodity. While LLMs can capitalize on the vastness of internet data, data for solving more specific technical systems issues is less plentiful and often private and proprietary. These two words are operative in the implications of how the data can be used. And, once a dataset is available, how does one know whether it is adequate, appropriate, unbiased and secure? Once a model is trained, designers learn that the model itself requires continuous improvements and thus, model validation has become a critical step.

Standards: Perhaps the most relevant work for wireless is in 3GPP.⁵ 3GPP started AI standardization discussions as early as Release-17 with RAN3 initial study items⁶ related to data collection and focusing on energy saving, load balancing and mobility optimization. RAN1, in Release-18, added an extensive study item on using ML for improving channel-state information, beam management and positioning accuracy. Work has progressed to multiple normative work items in 3GPP as part of Release-19, now underway.

This all must happen in the context of governments developing associated policies related to AI technology. Related headlines include European Parliament’s landmark law⁷ related to proper use, security and consumer recourse, the U.S. Executive Order on AI Safety and Security⁸ and the subsequent founding of the U.S. AI Safety Institute.⁹ Much of this policy work is focused on the impact of LLMs on the internet and other media as well as on the security of critical communications and compute infrastructure. We can expect impacts on the deeper technical uses of AI in unpredictable ways.

For radio engineers, we are already seeing new approaches to using AI to not only manage wireless communications but even do some designing. Rather than becoming cynical or worried, I view these as technical challenges characteristic of any advance in technology that our engineering community has put to good use time and again. I have seen the interesting results of ML-designed filters and antennas and, while not always practical, they change one’s perspective as to how to meet the demanding key performance indicators of our industry. In a demanding technical environment like radio systems and wireless communications networks, there is much work ahead of us to not only validate models and datasets but also to validate and improve the results of the AI-optimized behavior and designs themselves. The combination of conventional and AI-enabled means of such measurement is an intriguing and exciting area of development and I am looking forward to working to make the most of this technology.

References

1. Focus Group on Technologies for Network 2030, Web: www.itu.int/pub/t-fg-net2030.
2. J. Hoydis, F. A. Aoudia, A. Valcarce and H. Viswanathan, “Toward a 6G AI-native Air Interface,” IEEE Commun. Mag., Vol. 59, No. 5, pp. 76–81, May 2021.
3. “Exploring the Role of AI in Wireless ,” Keysight, 2022, Web: www.youtube.com/watch?v=DW0iZyXcDr8.
4. “The Turbulent Past and Uncertain Future of Artificial Intelligence,” IEEE Spectrum, Sept. 2021, Web: spectrum.ieee.org/magazine/2021/october.
5. Finding AI in 3GPP, 3GPP, August 2022, Web: www.3gpp.org/technologies/finding-ai-in-3gpp.
6. “Study on Enhancement for Data Collection for NR and ENDC,” 3GPP, Web: portal.3gpp.org/desktopmodules/Specifications/SpecificationDetails.aspx?specificationId=3817.
7. “Artificial Intelligence Act: MEPs Adopt Landmark Law,” European Parliament, March 2024, Web: www.europarl.europa.eu/news/en/press-room/20240308IPR19015/artificial-intelligence-act-meps-adopt-landmark-law.
8. “Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence,” The White House, Oct.2023, Web: www.whitehouse.gov/briefing-room/presidential-actions/2023/10/30/executive-order-on-the-safe-secure-and-trustworthy-development-and-use-of-artificial-intelligence/.
9. U.S. AI Safety Institute, Web: www.nist.gov/artificial-intelligence/artificial-intelligence-safety-institute.