Source-linked AI summary
Toward a 6G AI-Native Air Interface
Jakob Hoydis, Fayçal Ait Aoudia, Alvaro Valcarce, Harish Viswanathan
TL;DR
The paper asks whether AI can help define 6G’s air interface, not merely support distributed learning. It proposes an AI-native interface that jointly considers applications, hardware, and radio conditions, with case studies showing improved receiver performance and pilotless transmission. [Not supported by schema; omit citations]
Problem
It is uncertain whether AI will play a defining role in designing 6G’s air interface, beyond supporting distributed learning systems.
Method
The paper presents a vision in which ML designs or replaces parts of PHY and MAC processing and optimizes communication procedures for applications, hardware, and radio environments.
Results
The case study reports a 0.5 dB additional gain from data-aided channel estimation and detection, performance approaching perfect CSI with larger models, and pilotless transmission at the same BER as neural reception with 64-QAM.
Takeaways & Limitations
An AI-native air interface could integrate application needs and system constraints while enabling learned signaling, processing, and end-to-end removal of demodulation reference-signal overhead.
Takeaways & Limitations
Emergent protocols may be difficult to interpret, so some use cases may require protocols close to known ones and metrics measuring protocol distance.
Abstract
from arXiv · showhide
Each generation of cellular communication systems is marked by a defining disruptive technology of its time, such as orthogonal frequency division multiplexing (OFDM) for 4G or Massive multiple-input multiple-output (MIMO) for 5G. Since artificial intelligence (AI) is the defining technology of our time, it is natural to ask what role it could play for 6G. While it is clear that 6G must cater to the needs of large distributed learning systems, it is less certain if AI will play a defining role in the design of 6G itself. The goal of this article is to paint a vision of a new air interface which is partially designed by AI to enable optimized communication schemes for any hardware, radio environment, and application.
I. INTRODUCTION
6G research increasingly expects AI/ML to shape networks end-to-end, including the air interface. The paper asks whether ML could design parts of 6G’s PHY and MAC layers rather than merely enhance existing components.
- AI/ML is expected to influence 6G across design, deployment, and operation, including network automation and optimization.
- 6G networks may become cognitive by applying ML/AI to functions such as slicing, quality of service, mobility management, radio resource management, and spectrum sharing.
- Although ML enhances components in current 5G systems, no 5G component has been designed by ML.
- The paper therefore examines whether 6G could let ML/AI modify parts of the PHY and MAC layers.
II. AI-NATIVE AIR INTERFACE
The AI-Native Air Interface integrates application needs with communication, hardware, and channel constraints. Its vision includes learned waveforms, transceivers, signaling, access policies, and end-to-end PHY/MAC designs.
- The AI-AI serves applications with needed data efficiently by adapting to hardware constraints and radio environments.
- ML could learn bespoke waveforms, modulation schemes, pilots, and codes adapted to frequency, hardware limitations, and channel conditions.
- Fully learned transceivers can be trained for targeted hardware platforms, supporting diverse use cases and deployment scenarios without traditional algorithm-design and implementation processes.
- A flexible learning framework could reduce the need to standardize numerous air-interface options across frequency bands and scenarios.
- AI-AI learning can jointly optimize communication data and the application consuming it, addressing reliability, semantics, and effectiveness for purpose-specific systems.
- End-to-end learning could produce signaling schemes and channel-access policies that transition between contention- and schedule-based operation, with PHY and MAC jointly learned.
B. Three steps towards the AI-AI
The transition to the AI-AI is organized into three development phases. The first two can begin on future 5G systems because they change implementations without requiring new signaling or procedures.
- The three phases progressively replace processing blocks and eventually allow ML to design parts of the PHY layer.
1) ML replaces single processing blocks:
The first phase uses ML to enhance or replace individual transceiver processing blocks, especially in receivers. This introduces practical requirements for data, updates, training, and hardware acceleration.
- Phase 1 applies ML to individual receiver functions such as random-access detection, channel estimation, and symbol demapping.
- Phase 1 receiver processing combines ML and traditional blocks while requiring data acquisition, model updates, online training, and hardware accelerators.
2) ML replaces multiple processing blocks:
In the second transition phase, ML models jointly replace multiple receiver-processing blocks, increasing model size and the importance of hardware acceleration.
- ML can jointly perform channel estimation, equalization, and demapping.
- As ML assumes multiple processing roles, the models grow larger and hardware acceleration becomes increasingly important.
- Vendors must adopt an ML-only or ML-first approach because parallel ML and non-ML backup implementations are not viable on the same platform.
3) ML designs parts of the air interface:
The third phase gives ML/AI freedom to design parts of the PHY and MAC, replacing fixed specifications with deployment-time optimization procedures.
- ML/AI can design parts of the physical and MAC layers rather than operating only within predetermined blocks.
- This approach requires signaling and procedures that support distributed end-to-end training and deployment-time air-interface optimization.
- Standardization may specify procedures for optimizing modulation schemes and waveforms instead of fixing those components in advance.
C. Case study: From neural receivers to pilotless transmissions
The case study progresses from a conventional OFDM receiver to neural demapping, neural receiver processing, and a jointly learned constellation that enables pilotless transmission.
- The case study evaluates a doubly selective SISO channel at 3.5 GHz with TDL-A, 100 ns delay spread, and a receiver moving at 50 km h−1.
- The baseline uses 64-QAM, pilot symbols, least-squares channel estimation, nearest-pilot equalization, Gaussian LLR demapping, and belief propagation decoding.
- The baseline is approximately 3 dB worse than a receiver with perfect channel state information.
- A per-resource-element neural demapper improves BER by approximately 0.5 dB over the baseline but cannot compensate for channel aging.
- A full-TTI neural demapper uses convolutional residual processing to compensate for some channel-estimation and equalization errors, improving BER by 2 dB over the baseline.
- A jointly optimized learned constellation achieves the same BER as the 64-QAM neural receiver while transmitting no pilots, potentially removing demodulation-reference-signal overhead.
III. THE NEXT FRONTIER: PROTOCOL LEARNING FOR THE MAC
The paper proposes learning MAC protocols by treating wireless message exchanges as a language between collaborative radio nodes, potentially automating protocol design.
- Higher-layer protocols assume bit-by-bit transmission while coordinating signaling and procedures across network nodes.
- Protocol standardization and subsequent implementation and testing are costly and can produce ambiguous technical specifications.
- Wireless protocols are sequences of messages exchanged between radio nodes to transmit service data units and can therefore be viewed as a machine language.
- Deep multiagent reinforcement learning could train wireless devices to learn communication protocols.
A. Learning a given protocol
Agents can be trained to implement a known protocol while learning customized signaling interpretations and channel-access policies. In unreliable channels, greater signaling-based coordination is associated with better performance, but learned UEs can still vary widely in policy and results.
- A. Learning a given protocol: Agents can learn a known protocol instead of receiving a conventionally coded implementation.The proposed training could replace protocol interpretation, implementation, and testing efforts while preserving protocol standardization.
- A. Learning a given protocol: The uplink signaling action space contains all control messages a UE may send, while the PHY action space contains all channel-access commands available through the PHY API.
- A. Learning a given protocol: Instantaneous coordination is the mutual information between downlink signaling messages and the next channel-access actions.In unreliable channels, Fig. 6 reports positive Pearson correlations between coordination and performance.
- A. Learning a given protocol: Different learned UEs can assign different meanings to the same fixed messages, producing substantially different policies and performance.The resulting high variance reflects the large solution space and motivates optimizing radio protocols.
B. Emerging a new protocol
The next stage is to let UEs and base stations discover protocols rather than merely learn a human-designed one. This creates opportunities for deployment-tailored radio systems, while interpretability remains a practical constraint.
- B. Emerging a new protocol: Protocol learning can progress from implementing a known protocol to exploring the full space of possible protocols between UEs and base stations.The main challenge is discovering a shared state in which the radios can interpret one another’s messages.
- B. Emerging a new protocol: Self-play can let radios evolve an initially provided protocol after they acquire a starting communication language.
- B. Emerging a new protocol: Emergent protocols may be difficult to interpret, which matters for fault detection and performance monitoring.Metrics measuring distance between protocols, together with training that minimizes this distance, may improve intelligibility.
- B. Emerging a new protocol: Learned protocols could tailor radio systems to deployment environments and boost capacities for niche and vertical markets.
IV. CONCLUSION
The paper concludes that AI/ML may profoundly change how communication systems are designed and deployed. Whether the proposed AI-Native Air Interface becomes part of 6G remains to be determined over the next decade.
- IV. CONCLUSION: AI/ML is expected to profoundly change the future design and deployment of communication systems.
- IV. CONCLUSION: The paper does not yet establish whether the AI-Native Air Interface will provide sufficiently compelling benefits for adoption in 6G.The authors frame that outcome as a question for the next decade.