sign in sign up
<<
Home , Communication protocol

1

Toward a 6G AI-Native Air Interface Jakob Hoydis, Senior Member, IEEE, Fayc¸al Ait Aoudia, Member, IEEE, Alvaro Valcarce, Senior Member, IEEE, and Harish Viswanathan, Fellow, IEEE

Abstract—Each generation of cellular “cognitive” in the sense that various aspects such systems is marked by a defining disruptive technology of its as virtualized network function placement, slicing, time, such as orthogonal frequency division quality of service, , re- (OFDM) for 4G or Massive multiple-input multiple-output source management, and spectrum sharing will all (MIMO) for 5G. Since artificial intelligence (AI) is the defining technology of our time, it is natural to ask what rely on ML/AI to varying degrees. In fact, we expect role it could play for 6G. While it is clear that 6G must that ML/AI will significantly impact even the 6G air cater to the needs of large distributed learning systems, interface, which is the focus of this paper. it is less certain if AI will play a defining role in the While ML starts to be widely used in the industry design of 6G itself. The goal of this article is to paint a to enhance the of various compo- vision of a new air interface which is partially designed nents within the 5G radio access network (RAN) by AI to enable optimized communication schemes for any hardware, radio environment, and application. and core, it is fair to say that there is not a single component of 5G which has been designed by ML. The purpose of this article is therefore to raise and I.INTRODUCTION discuss the question: What if 6G was designed in a way that ML/AI could modify parts of the physical While 5G is rolled out globally and the stan- (PHY) and (MAC) layers? dardization discussions for its future evolution take place, researchers in academia and industry start to think about visions, use cases, and disruptive II.AI-NATIVE AIR INTERFACE key technologies for a possible 6G system. Publicly Let us start by presenting our vision and moti- funded 6G research projects in Europe [1], the vation of an AI-Native Air Interface (AI-AI) which United States, and China are under way, and also describes a disruptive change to the traditional way the ITU has begun their work on requirements communication systems are designed, standardized, for fixed networks in the 2030s [2]. A common and productized. We will first provide a summary theme in many 6G vision papers is that of creating of possible benefits of such an approach, then digital twin worlds for seamlessly connecting and detail three important but also necessary develop- controlling physical and biological entities to enable ment steps towards realizing our vision, and finally new mixed-reality super-physical experiences [3]. present a case study which exemplifies the respec- arXiv:2012.08285v2 [cs.NI] 30 Apr 2021 Apart from new spectrum technologies and the tive potential performance gains and advantages. support of simultaneous and sens- As illustrated in Fig. 1, the goal of the AI-AI ing as well as extreme connectivity requirements is to serve an application with the data it needs (among others), it is expected that machine learn- in the most efficient way by taking into account ing (ML) and AI will play a defining role in the the constraints of the available hardware and the development of 6G networks end-to-end across the radio environment. The AI-AI hence no longer design, deployment, and operational phases [4]. As decouples source and channel coding as well as the network evolves to programmable and flexible communication of data from the intended use by an cloud native implementation, ML/AI-based network application, and embraces hardware constraints and automation will be crucial to simplify network man- undesired effects of the communications channel agement and optimization. Networks will become rather than fighting them. While the last decades were used to implement the scientific breakthroughs The authors are with Nokia ({jakob.hoydis, fay- cal.ait aoudia, alvaro.valcarce rial, harish.viswanathan}@nokia-bell- by Shannon and Wiener (as well as many others), labs.com). we are now entering a new era for communications 2

Data AI-AI Hardware Environment HardwareAI-AI Application

Environment, hardware & application dependency feedback

Fig. 1. The AI-Native Air Interface (AI-AI) adapts to different radio environments, hardware, data, and applications. Compared to previous air interfaces, it is not only designed to reliably transmit bits, but also to serve an application with the data it needs in an optimal way. where classical approaches must be revisited and design and hardware implementation anymore. They new theories developed to achieve the technological can be trained directly for the targeted hardware breakthroughs needed for a possible 6G system. We platform (or even on it, depending on the capabili- believe that our vision of an AI-AI could become a ties). With an increasingly rich diversity of expected reality within the next decade by optimizing the air 6G use-cases and the emergence of small-scale interface from end-to-end thanks to advances in the sub-networks [7], this versatility becomes essential field of AI for communications. to ensure that 6G can cater for each individual use-case and deployment scenario in the best way A. Possible benefits possible. Moreover, given the rapid speed at which ML hardware accelerators are developing, it is likely First, in contrast to a single classical waveform that learned transceiver will rapidly choice such as OFDM in 5G, the AI-AI could outperform their traditional counterparts in power enable learning of bespoke waveforms for different efficiency, latency, and cost. Advances in neuromor- frequencies which do not only make more efficient phic [8] could further amplify this trend. use of the spectrum but are also optimally adapted to practical limitations of the transceiver hardware Third, the more we follow the AI-AI principle and channel, such as non-linear power amplifiers, of learning based design and specification, the less hybrid analog-digital processing, low quantization needs to be standardized. The current 5G specifica- resolution, very short channel coherence time and tion boasts a very rich sets of options and parameters , phase and impulsive noise. Also new for different frequency bands and scenarios which schemes, pilot sequences, and codes pose a difficult challenge from an implementation can be learned or optimized with ML to squeeze point of view. It is undesirable to scale this approach even more performance out of the spectrum. Sev- to even more complex and diversified settings in 6G. eral research groups have recently demonstrated If, on the other hand, only a sufficiently flexible practical gains of such approaches, see, e.g., [5], framework for air interface learning was standard- [6]. A very interesting application arises for the ized, the system could auto-adjust to any kind of transmission of short messages, where the classical scenario. With a bit of wishful thinking, one could frame structure of preamble, pilots, coded bits, and hope that 6G could be the last communication cyclic-redundancy check could be replaced with system to be standardized. fully learned radio burst conveying a few bits of Fourth, the AI-AI allows integration of the data . Furthermore, learning of new wave- and the application consuming it into a single end- forms for simultaneous communication + “X” (e.g., to-end learning . Using the terminology from sensing or power transfer) has high potential. Shannon and Weaver’s seminal book [9], the AI- Second, fully learned transceivers have the benefit AI no longer only solves the problem of reliably that they do not need to undergo the very costly transmitting bits (Level A), but simultaneously ad- and time-consuming traditional process of dresses the problems of semantics (Level B) and 3

Transmitter Receiver Symbol Modulation / Channel Symbol Encoding Sync. Equalization Decoding Mapping Waveform Estimation Demapping

1. ML replaces/enhances individual processing blocks ML ML ML ML ML

2. ML replaces multiple processing blocks ML ML ML

ML 6G ML

3. ML designs part of the air interface itself

Fig. 2. Three phases towards the AI-AI: After gradually replacing most of the processing blocks, ML will design parts of the PHY layer. effectiveness (Level C) of communication (see, e.g., the receiver. Examples are physical random access [10]). While the latter aspects may not be applicable channel detection, channel estimation, or symbol to the generic communication scenario, demapping. Although seemingly simple, this step they become relevant for communication systems constitutes a paradigm change in the way the indus- which are tailored to specific purposes and under try designs and deploys radio transceivers. Even if the control of a single entity, such as industrial the ML models are likely to be rather small, several communication systems for sensing, surveillance, important problems such as data acquisition, model and robot control. updates, and online training need to be solved and Lastly, the idea of end-to-end learning naturally hardware accelerators must be integrated into the extends to the MAC layer where it would be de- PHY processing flow. The receiver processing will sirable to emerge optimized signaling schemes and contain a mix of ML and traditional blocks. channel access policies which fluently transition 2) ML replaces multiple processing blocks: In from contention- to schedule-based depending on the second phase, more functionality is given to the use-case and environment. Protocol learning ML models which take on the joint role of multiple could also address the problem of optimally mul- processing blocks. This could be, e.g., joint channel tiplexing resources for communication and sensing estimation, equalization, and demapping. In this (or other applications that radio waves can be used phase of the transition, the ML models will grow for). Ultimately, PHY and MAC layers could be larger, becomes increasingly jointly learned together. important, and vendors need to commit to an ‘ML- only/ML-first’ approach because it is not viable to B. Three steps towards the AI-AI implement ML and non-ML backup solutions in We currently see three important phases in the parallel in the same processing platform due to development and transition to the AI-AI, each increased power consumption and cost. This means of which requires sustained multi-disciplinary re- that ML is also trusted more although the inner search. These are schematically shown in Fig. 2. workings of large models are less interpretable, but The first two phases do not require any new sig- the potential gains are also higher. An example naling or procedures as they only impact the im- will be provided in Section II-C. In this phase, plementation of transceiver . They can we will also realize what possibilities such learned therefore be carried out on future 5G systems to transceiver components open up, e.g., need for less gather practical experience while the research on pilots, no cyclic prefix, less stringent synchroniza- 6G is progressing. tion. In other words, we will learn what are the 1) ML replaces single processing blocks: In the things ML allows us to do which we could not do first phase, which is already happening in the before (with reasonable effort). industry today, ML will be used to enhance or 3) ML designs parts of the air interface: In the replace some of the processing blocks, mostly in third phase, we will give even more freedom to 4

−1 100 50 km h and the channel evolves in time accord- ing to Jakes’ model. We consider cyclic prefix- based OFDM with 72 subcarriers spaced 30 kHz 10 1 apart and assume transmission time intervals (TTIs) of 14 consecutive OFDM symbols which contain codewords of length 1024 bit at a coderate of 2/3, 2 10 generated by a 5G-compliant code. BER Our non-ML baseline assumes 64-quadrature am-

10 3 plitude modulation (QAM), pilots transmitted on ev- ery other sub-carrier on the third and twelfth OFDM symbols, least-squares channel estimation, equaliza- 10 4 tion based on the nearest pilot, exact demapping to

6 8 10 12 14 16 log-likelihood ratios (LLRs) assuming a Gaussian Eb/N0 [dB] post-equalized channel, as well as a standard belief

Baseline Neural Receiver propagation (BP) decoder. Neural Demapper (Symbol) GS + Neural Receiver The BER performance of the baseline and all Neural Demapper (Grid) Perfect CSI other schemes that will be subsequently introduced is shown in Fig. 3. One can see that there is ap- Fig. 3. BER performance of all compared schemes in the case study. proximately a 3 dB gap between the baseline and a receiver assuming perfect channel information (CSI). We now describe some ways to close this ML/AI and let it design parts of the physical and gap using ML-enhanced receiver processing before MAC layers itself. This represents another paradigm delving into the benefits of optimizing parts of the change in the way communication systems are de- , too. signed because not all aspects of the PHY and MAC Due to channel aging and imperfect channel es- layers might be fixed in advance. This approach timation, the quality of the post-equalized symbols requires new forms of signaling and procedures to which are fed into the demapper changes over the enable distributed end-to-end training. Rather than grid of resource elements (REs) within a TTI. A specifying, e.g., modulation schemes and wave- first possibility to cope with this problem is to learn forms, one would need to specify procedures that a bespoke neural demapper for each RE (Phase 1). can be used to optimize these aspects of the air inter- The BER performance of such a scheme is shown by face at deployment time. This is clearly something the red line with dot markers in Fig. 3. As expected, nobody has ever done before and that requires a it provides some 0.5 dB gain over the baseline by massive change in the way communication systems computing better LLRs, but cannot compensate for are standardized. It is of course also possible that channel aging which results in a rotation and scaling ML-designed solutions to specific problems will be of the equalized constellation. specified, which is, e.g., already the case in 5G for In order to address these shortcomings, one can the channel code design. use a larger neural demapper which does not operate symbol-by-symbol but rather produces LLRs for the full TTI. It was shown in [11], [12], that a C. Case study: From neural receivers to pilotless fully convolutional ResNet architecture with dilated transmissions separable convolutions achieves remarkable perfor- Next, we will present a case study that illustrates mance for this task (see Fig. 4). By having access to the progression through the three phases towards the full TTI of post-equalized symbols, the learned the AI-AI outlined above and demonstrates the re- demapper can compensate for some of the errors spective performance gains. We consider a doubly- made by the channel estimator and equalizer to selective single-input single-output (SISO) channel provide a 2 dB improvement over the baseline (see at a carrier frequency of 3.5 GHz with the TDL-A purple line with triangular markers in Fig. 3). power delay profile and a delay spread of 100 ns. Interestingly, it turns out that one can assign The receiver is assumed to move at a speed of the joint task of channel estimation, equalization, 5

2

LLRs for all REs 1

0

1

2

2 1 0 1 2

Depth-wise separable Residual connections Fig. 5. Learned constellation allowing pilotless transmissions to- dilated convolutions gether with a neural receiver. It has zero mean, unit power, and a single axis of symmetry. The optimal bit-labeling is also learned but Fig. 4. The neural receiver produces LLRs for an entire TTI of post- not shown for readability. FFT symbols. Key architectural components are depth-wise separable dilated convolutions and residual connections. The same architecture can also be used as a neural demapper, operating on a TTI of equalized symbols. III.THENEXTFRONTIER:PROTOCOLLEARNING FORTHE MAC The protocols above the PHY take bit-by-bit and demapping to a neural network with a similar transmission for granted to develop complex - architecture (Phase 2). It is fed with a TTI of ing schemes and orchestrate elaborate procedures post-FFT received signals from which it directly across the network’s nodes. As a result, computes LLRs for all bits. In addition to the gains coordinate harmoniously and provide more powerful of the learned demapper, this neural receiver is now services than what point-to-point links offer. able to carry-out data-aided channel estimation and The industry defines these detection, resulting in an additional 0.5 dB gain. By procedures through protocol standards which are increasing the model complexity and the size of the agreed upon in large meetings, where competing input (more sub-carriers and OFDM symbols), the technical and economical interests are debated year performance can be brought arbitrarily close to the after year. These efforts have a high cost and some- perfect CSI performance [11]. times result in ambiguous technical specifications Lastly, we would like to investigate the benefits (TSs). After a TS is released, the implementation of a learned constellation (i.e., geometric shaping and test phase begins, which is even more costly. (GS)) at the transmitter side (Phase 3), which is For this reason, it is interesting to question if this jointly optimized together with the neural receiver. burdensome undertaking could be somehow auto- Fig. 5 shows this constellation which is used on mated. And if so, would the result perform better every RE instead of the mix of pilots and 64-QAM than human-designed protocols? symbols sent by the baseline. As can be seen from protocols are sequences of messages Fig. 3, this system achieves the same BER as the exchanged between radio nodes to transmit service neural receiver with 64-QAM, but has the additional data units (SDUs). As such, protocols can also benefit that no pilots are transmitted. End-to-end be understood as a language between collaborative learning could hence remove the need and control machines. Learning a language is something not overhead for reference signals. only humans but also machines can do [13]. In This case study has only scratched the surface fact, the field of learning to communicate (L2C) of what will be possible in the future. Interesting is growing fast thanks to recent developments in directions for future research include end-to-end deep multiagent (MARL), learning for new waveforms, constrained hardware, see, e.g., [14]. While most research efforts in this very short messages, as well as joint source-channel field are targeted towards natural languages, we be- coding for a specific application (which is also lieve these techniques can also be used for training learned). Meta, transfer, and federated learning are wireless devices to learn communication protocols. key enablers to make such schemes practical. The 6G protocols of the AI-AI could be built this 6

way and there are two ways to achieve this: BLER = 0.0. Learners BLER = 0.1. Learners BLER = 0.2. Learners A. Learning a given protocol 0.20 BLER = 0.0. Expert BLER = 0.1. Expert Intelligent agents could be trained to 0.18 BLER = 0.2. Expert behave according to an a priori known protocol. ρ = 0.0 Instead of coding the standard, agents could be 0.16 ρ = 0.6 trained on it via . Such training 0.14 ρ = 0.5

would ideally be done only once during factory Performance [SDU/t] production. This would yield a protocol implemen- 0.12 tation and, although it does not replace protocol 0.10 standardization, could replace protocol interpreta- 0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.90 0.95 tion, implementation, and testing efforts. The cost- Instantaneous coordination savings and time-to-market reduction potential of Fig. 6. Relationship between performance and signaling-based co- 6G nodes built this way might be significant. ordination. Each point depicts the mean SDU rate and the IC of Already today, user equipments (UEs) could be different learned UE MACs. Two UEs were trained to transmit two randomly arriving SDUs to the BS without collision or channel loss, trained to learn the 5G New Radio (5G NR) MAC via a protocol known to the BS and unknown to the UEs (i.e., UEs protocol. This would include learning to interpret had no mapping between the protocol’s messages and PHY actions). the different control messages received from the (BS) (e.g., discontinuous reception (DRX), timing advance (TA), etc.), as well as learn- are fixed and known to the BS. Nevertheless, ML ing what to send in the uplink (e.g., buffer status training can yield UEs with widely different inter- report (BSR), power headroom report (PHR), etc.). pretations of these messages and consequently, also Learning can be formulated as a MARL problem, very different policies and performance (note the wherein the UEs’ MACs are deep reinforcement high variance of the results in Fig. 6). This variance learning (RL) agents with two action spaces: is a consequence of the vast size of the solution • Uplink signaling action space: All possible space, which strengthens the case for optimizing the uplink control messages a UE may ever send. protocols our radios use. • PHY action space: All channel access com- mands the MAC may invoke through the PHY B. Emerging a new protocol application programming interface (API). As there are many possible ways a UE can com- Protocol implementations trained this way may municate with a BS that implements a given proto- outperform expert systems thanks to the customized col, why should we constrain ourselves to human- signaling and channel access policy [15]. The sig- designed protocols? The next frontier in protocol naling is the vocabulary of messages the nodes have learning is to let UEs and BSs explore the entire at their disposal as well as the rules about how to space of possible protocols. This is challenging use them. The channel access policy decides how to because for two radios to coordinate, they first need make use of the PHY API, based on information it to find a state where they can interpret each other’s has thanks to the signaling. It is therefore fair to ask messages. Recent L2C research suggests that this whether the gains are due to the learned signaling, language discovery problem may be overcome by the channel access policy, or both. The impact of first training the nodes with . signaling on the UEs’ actions can be quantified This essentially endows radios with an initial pro- by metrics such as the instantaneous coordination tocol that they can later evolve through self-play. (IC), which is the mutual information between the Communication protocols emerged this way may downlink signaling messages and the next channel be hard to interpret, which is essential for fault de- access actions. This is illustrated in Fig. 6, which tection or performance monitoring. For this reason, shows that in unreliable channels, higher levels some use-cases may favor protocols that are close to of coordination lead to performance gains (note known ones. This requires metrics that measure the the positive Pearson correlation coefficients ρ). The distance between two protocols. Training techniques semantics of the MAC messages used in Fig. 6 minimizing this distance may improve intelligibility. 7

The ability to learn new communication pro- [10] P. Popovski, O. Simeone, F. Boccardi, D. Gund¨ uz,¨ and O. Sahin, tocols opens the door to radio systems that are “Semantic-effectiveness filtering and control for post-5G wire- less connectivity,” J. Indian Inst. Sci., vol. 100, no. 2, pp. 435– highly tailored to their deployment environment, 443, May 2020. thus boosting 6G capacities for niche and vertical [11] M. Honkala, D. Korpi, and J. M. Huttunen, “DeepRx: Fully markets. The AI-AI will not only reduce today’s convolutional deep learning receiver,” arXiv:2005.01494, May 2020. signaling overheads, but also the standardization and [12] F. Ait Aoudia and J. Hoydis, “End-to-end Learning for development efforts for the highly complex radio OFDM: From neural receivers to pilotless ommunication,” technologies of the next decades. We foresee a 6G arXiv:2009.05261, Sep. 2020. [13] T. B. et al., “Language models are few-shot learners,” in Proc. future where parts of the radio stack development Int. Conf. Neural Inf. Proc. Sys. (NIPS), Vancouver, Spain, Dec. cycle could be replaced by the click of a button. 2020. [14] J. N. Foerster, Y. M. Assael, N. de Freitas, and S. Whiteson, “Learning to communicate with deep multi-agent reinforcement IV. CONCLUSION learning,” in Proc. Int. Conf. Neural Inf. Proc. Sys. (NIPS), While the next decade will reveal if our vision Barcelona, Spain, Dec. 2016, pp. 2145–2153. [15] A. Valcarce and J. Hoydis, “Towards joint learning of optimal of an AI-Native Air Interface provides sufficiently signaling and wireless channel access,” arXiv:2007.09948, Jul. compelling benefits to make it into 6G, we are 2020. certain that AI/ML will profoundly change the way communication systems will be designed and de- ployed in the future. We hope that some of the readers will join us on this exciting journey. Jakob Hoydis [SM’19] is head of a research department at Nokia Bell Labs, France, focusing on radio systems and artificial intelli- gence. He received his Ph.D. degree from Supelec,´ Gif-sur-Yvette, Acknowledgments France, in 2012, He is a co-author of the textbook “Massive MIMO Networks: Spectral, Energy, and Hardware Efficiency” (2017). He is We would like to thank P. Mogensen, S. ten currently chair of the IEEE COMSOC Emerging Technology Initiative Brink, S. Cammerer, S. Dorner,¨ M. Goutay, P. on Machine Learning for Communications as well as Editor of the Srinath, M. Honkala, D. Korpi, J. Huttunen (and IEEE Transactions on Wireless Communications. many others) for numerous discussions that helped to shape the vision outlined in this article. Fayc¸al Ait Aoudia [M’20] is a research engineer at Nokia Bell Labs, France, where he is working on machine learning for wireless REFERENCES communications. He received the M.Sc. degree in science engineering from INSA Lyon, France, in 2014 and the Ph.D. degree [1] 6G Waves 1, Spring 2020. 6G Flagship, University of Oulu, from the University of Rennes 1, France, in 2017. His Ph.D. work has Finland, Mar. 2020. focused on energy management and design [2] ITU Focus Group on Technologies for Network 2030 (FG-NET- for autonomous wireless sensor networks. 2030), “Network 2030: A blueprint of technology, applications and market drivers towards the year 2030 and beyond,” White paper, May 2019. [3] H. Viswanathan and P. E. Mogensen, “Communications in the Alvaro Valcarce [SM’20] is a research engineer at Nokia Bell 6G era,” IEEE Access, vol. 8, pp. 57 063–57 074, Mar. 2020. Labs, France, where he focuses on the application of reinforcement [4] K. B. Letaief, W. Chen, Y. Shi, J. Zhang, and Y.-J. A. Zhang, learning to L2 and L3 problems for the development of beyond- “The roadmap to 6G: AI empowered wireless networks,” IEEE 5G technologies. He was previously a system engineer with - Commun. Mag., vol. 57, no. 8, pp. 84–90, Aug. 2019. H GmbH, where he developed Self Organising Networks (SON) [5] S. Cammerer, F. Ait Aoudia, S. Dorner,¨ M. Stark, J. Hoydis, algorithms for LTE small-cells. He received his PhD in 2010 from and S. T. Brink, “Trainable communication systems: Concepts the University of Bedforshire and his background is on cellular and prototype,” IEEE Trans. Commun., Jun. 2020. networks, optimization algorithms, computational electromagnetics, [6] J. Downey, B. Hilburn, T. O’Shea, and N. West, “Machine and satcom. learning remakes radio,” IEEE Spectr., vol. 57, no. 5, pp. 35–39, Apr. 2020. [7] V. Ziegler, H. Viswanathan, H. Flinck, M. Hoffmann, V. Rais¨ anen,¨ and K. Hat¨ onen,¨ “6G architecture to connect the Harish Viswanathan [F’13] is head of the Radio Systems Research worlds,” IEEE Access, vol. 8, pp. 173 508–173 520, Sep. 2020. group at Nokia Bell Labs, USA, and a Bell Labs Fellow. He received [8] B. Rajendran, A. Sebastian, M. Schmuker, N. Srinivasa, and his B.Tech. from Electrical Engineering, IIT Madras, Chennai, India, E. Eleftheriou, “Low-power neuromorphic hardware for signal and the M.S. and Ph.D. from Electrical Engineering, Cornell Uni- processing applications: A review of architectural and system- versity, Ithaca, NY, USA. Since joining Bell Labs in 1997, he has level design approaches,” IEEE Signal Process. Mag., vol. 36, worked extensively on wireless research, ranging from physical layer no. 6, pp. 97–110, Nov. 2019. to and protocols. From 2012 to 2015, he was [9] C. E. Shannon and W. Weaver, The mathematical theory of a CTO Partner in Alcatel-, advising the Corporate CTO on communication. University of Illinois Press, 1949. technology strategy.