The growing number of interconnected intelligent devices puts existing wireless networks under immense pressure, demanding their development. As a result, next-generation beyond-5G and 6G systems, denoted by B5G herein, must achieve high efficiency and throughput. A novel strategy to achieve this is to apply AxC to network applications that do not require nine 9s reliability, for example, in non-critical mobile broadband. Examples of approximable applications within a standard 5G network are shown in Fig. 1. Applying AxC is expected to increase energy efficiency with little-to-no quality degradation for users and service providers.
Figure 1 Common resources used to implement a 3GPP-compliant 5G architecture with approximable applications highlighted.
We survey this intersection of AxC and B5G network algorithms, highlighting trends and tendencies in existing work and directions for future research. Our article has a particular emphasis on future technologies, which are yet to be integrated into real-world networks but are envisioned to enable previously unseen performance and efficiency. To improve reproducibility, we systematically survey existing work. The topics covered in the paper and summarized here are shown in Fig. 2.
Figure 2 The main topics covered in the paper and summarized here.
Within the existing network infrastructure, the main application points for AxC are highlighted in Fig. 1. The algorithms are all focused on optimization or management tasks for which there rarely exist optimal solutions as network usage is dynamic. Commonly, papers that address these topics outline a particular use case or scenario and describe it mathematically. Computing a solution for these problems is often NP-hard, so approximations are needed to reduce complexity. Frequently, the proposed algorithms’ performances are measured in terms of network latency, throughput, spectral efficiency, or energy efficiency.
Serving the increasing number of users requires densifying B5G networks to operate at high frequencies. Mitigating issues of signal attenuation arising thereof may be done by employing RISs, which can be programmed to reflect signal beams toward individual users. Alternatively, greater distribution of smaller-scale APs in so-called cell-free networks may be a practical alternative to the cell-based architectures of present-day networks. The dense deployment also enables using the existing network infrastructure for alternative purposes, particularly environmental sensing. So-called ISAC systems combine these features. Finally, as networks are increasingly virtualized, their resources may be sliced to dynamically prioritize certain applications or types of traffic. All these technologies are greatly dependent on optimization and ML algorithms to achieve satisfactory performance and offer vast approximation opportunities. They are illustrated in the survey.
Throughout our survey process, we have found that only very few of the selected publications have co-authors from the industry. Moreover, most papers propose algorithms that scale poorly with the network size or number of users. These factors lead us to question whether the algorithms are realistic for real-world deployment. A promising mitigation strategy for these issues, however, is the partitioning of them in the Cloud-Edge continuum. Within this space, critical applications are executed near the users, while non-critical ones are kept in centralized data centers. This enables interesting quality-latency trade-offs, which complement the ones that AxC introduces well.
In addition to exploring the aforementioned trade-offs in more detail and extending existing proposals and their evaluation to more or different use cases, there are plenty of other open research directions. We assign these directions three classes: design space exploration, ML, and AxC in B5G. Within the former two classes, most directions call for reduced-complexity alternatives to established algorithms – popular examples from the survey include successive approximation and RL – and more open data. Within the latter, examples of directions are: 1) exploring hardware- or architecture-level AxC techniques in wireless networks, 2) applying neural approximation in place of other algorithms than the ones highlighted in Fig. 1, and 3) evaluating and identifying benefits and drawbacks of AxC to B5G networks with Cloud-Edge task offloading. In summary, there is plenty of work needed to establish a solid foundation for future B5G networks!
The paper is a joint work between TAU and IS-Wireless. The full text is available at: https://www.sciencedirect.com/science/article/pii/S1389128623003171
Acronyms:
- AP: Access Point
- AxC: Approximate Computing
- B5G: Beyond-5G and 6G networks
- IoT: Internet of Things
- ISAC: Integrated Sensing and Communication
- ML: Machine Learning
- RIS: Reflective Intelligent Surface
- RL: Reinforcement Learning
- 3GPP: 3rd Generation Partnership Project