Jiadai Wang, Jiajia Liu, Jingyi Li, Nei Kato
©SHUTTERSTOCK.COM/TANAWAT THIPMONTHA
6G networks are expected to provide instant global connectivity and enable the transition from “connected things” to “connected intelligence,” where promising network slicing can play an important role in network assurance and service provisioning for various demanding vertical application scenarios. On the basis of diversified massive data, artificial intelligence (AI)-assisted techniques are widely considered more suitable than traditional models and algorithms to deal with challenges faced by complex and dynamic slicing problems in 6G. In view of this, we provide a tutorial on AI-assisted 6G network slicing for network assurance and service provisioning, aiming to show the prospect of 6G slicing and the advantages of applying AI technology. Specifically, we propose six typical characteristics of 6G network slicing, analyze the feasibility of AI from different network domains and technical aspects, propose a case study on AI-assisted bandwidth scaling, and, finally, put forward the main challenges and open issues for its future development.
With the vigorous promotion and commercialization of 5G networks, the research and development of the next-generation communication system has also started. From the 6Genesis and Hexa-X projects in Europe, to the Next G Alliance in North America, to the first 6G satellite launched by China, 6G has attracted extensive attention worldwide. Featuring global network coverage, virtualization, and AI, the 6G vision aims to provide extreme connectivity and enable the transition from “connected things” to “connected intelligence” [1]. It also gives birth to various emerging verticals and application scenarios with specific capabilities, including holographic communication, fully autonomous driving, industrial automation, and so on, leading to strong demand for multipurpose and one-size-per-service networks [2], [3]. At this point, network slicing is widely regarded as a key enabler for 6G because it can customize and slice physical networks to accommodate various evolving application scenarios.
6G network slicing will play an important role in network assurance and service provisioning. For network assurance, it can create multiple logical networks over shared physical infrastructure to meet specific requirements, use its isolation feature to reduce the impact between logical networks, and set security-related indicators to achieve different levels of security assurance [4]. For service provisioning, it can not only meet the differentiated service-level agreements (SLAs) of various 6G scenarios but also greatly reduce the capital expenditure and operating expense by sharing network resources among tenants, bringing considerable economic benefits [1]. Meanwhile, it also faces lots of new challenges. For example, a large number of emerging services with stringent requirements can go far beyond the definition of the typical slice types in 5G [5], which further complicates the slice design. Multiple slices share infrastructures through virtualization technology that requires appropriate resource reservation and real-time performance analysis, which greatly increases the complexity of slice management. The addition of unmanned aerial vehicle (UAV) and satellite communications also makes it more difficult to coordinate different slice components. At this time, traditional approaches based on human cognition and modeling are hard to predict and may not satisfy the dynamic requirements of different services.
To deal with these challenges, AI-based approaches have attracted wide attention. The 3rd Generation Partnership Project (3GPP) has formally defined the AI-related network data analytics function as an important network function in 5G for intelligent analysis [6], which provides two service-based interfaces to the network slice selection function for requesting and subscribing to the slice-related key performance index, showing that AI is no longer just an accessory of network slicing. Since ubiquitous intelligence is one of the biggest characteristics of 6G, it is foreseeable that AI will exert more power for complex 6G slicing based on massive data, significantly improve the network assurance and service provisioning capabilities, and assist in automated slice management and slice performance optimization to ensure the quality of experience.
In this article, we not only present the 6G network slicing vision from a high-level view, including its architecture and six promising characteristics, but also highlight lots of detailed innovative AI-assisted slicing techniques in different network domains and make technical recommendations. In addition, we propose a case study on elastic bandwidth scaling, which demonstrates the advantages of AI applied to network slicing. Last but not least, we discuss the key trends and open issues of 6G network slicing.
Network slicing divides a shared physical network into multiple end-to-end logical networks as required, in which each logical network has tailored functions and features to meet differentiated SLAs of various 6G scenarios as well as provide independent, isolated, and integrated services. Following the architectural design of previous generations of communication systems, it is expected that 6G network slicing will continue to involve radio access network (RAN), transport network (TN), and core network (CN) subslices [7], as shown in Figure 1. Given the seamless global coverage feature of 6G, slice extensions with UAVs and satellites are also important [8]. Unified management can be realized by the slice management system, including the customer portal responsible for slice ordering and demand handling, the cross-domain module for the overall slicing management, and the subdomain controller for the management of the subslices. Through the multilayer management system, service requirements can be decomposed and delivered to each network domain, and end-to-end network slicing can be constructed and maintained.
Figure 1 The 6G network slicing architecture and application scenarios. SWIPT: simultaneous wireless information and power transfer; MIMO: multiple-input, multiple output.
We envision that 6G network slicing will have six typical characteristics, as in the following, including not only some enhanced existing features but also some new characteristics.
6G network slicing can customize diverse resources, such as computing, storage, transmission resources, and network functions, to achieve agreed differentiated network performance, meeting the requirements of various emerging 6G services.
Physical infrastructures and the underlying network resources shared by multiple network slices can be intelligently configured and adjusted based on service requirements, greatly improving resource utilization and reducing facility costs.
Coexisting network slices can be completely isolated and do not affect one another. On the one hand, this can guarantee the confidentiality of private information within slices, such as user data and service data. On the other hand, it can reduce the interference between slices and limit the scope of attacks and unexpected failures.
The addition of nonterrestrial networks in 6G, such as air networks, including various UAVs, and space networks, including all kinds of satellites, can provide global seamless coverage even in remote areas with no ground networks. In this way, 6G network slicing can be seamlessly extended to all corners of the world to provide customized services.
6G network slicing will inevitably involve terahertz spectrum to improve network capacity. Metamaterial-based antennas and radio-frequency front ends, such as reconfigurable intelligent surfaces [9], can be introduced to improve wireless communication performance. Other advanced technologies, such as simultaneous wireless information and power transfer and visible light communication, are also expected to be applied to the future 6G slicing. Also, all kinds of virtualization technologies will be further integrated into network slicing, including enhanced resource virtualization and digital twins.
Ubiquitous intelligence is one of the most important features of 6G. It will also penetrate various parts of 6G network slicing. Examples include intelligent prediction and resource allocation in the RAN domain, elastic bandwidth scaling and reliability guarantees in the TN domain, and flexible network function adjustment and anomaly detection in the CN domain. This characteristic is also the focus of this article.
Ubiquitous slice intelligence is one of the main advantages of 6G network slicing. In this section, we introduce AI-assisted slicing for network assurance and service provisioning from different network domains and technical focuses, including RAN, TN, CN, management system, and slice extensions.
RAN slicing can be implemented in various ways, such as carrier isolation, resource reservation, and quality of service (QoS) priority-based resource sharing, to meet differentiated service requirements. On this basis, AI can bring more flexibility and proactivity to 6G RAN slicing, as follows.
The problem of device association in 6G RAN slicing involves access control and handover. In terms of access control, base stations (BSs) can select the users allowed to access from QoS guarantee, service capacity guarantee, power optimization, and other perspectives. In terms of handover, since one BS can serve multiple slices and one slice can cover several BSs, the reselecting of both should be carefully considered. In the context of highly dynamic 6G networks, reinforcement learning (RL) can make intelligent decisions for device association through trial-and-error learning [10], which is a valuable prospect for AI-assisted RAN slicing, as demonstrated in Figure 2. Also, federated learning (FL), as a promising distributed learning mode, can help solve privacy issues and reduce data transmission costs, in which devices can learn models based on local data and then send model features to BSs for aggregation.
Figure 2 The typical AI-assisted 6G slicing techniques in different network domains: (a) AI-assisted CN slicing, (b) AI-assisted TN slicing, and (c) AI-assisted RAN slicing. VNF: virtual network function; SDN: software-defined networking; ML: machine learning; DL: deep learning.
In RAN slicing, the intelligent monitoring and prediction of the network condition facilitate proactive resource allocation and decision adjustment, which means preallocating resources and making policy adjustments in advance for certain services. The network condition includes not only the network parameters, such as channel quality, but also the request density. In the first case, machine learning (ML)/deep learning (DL)-based methods can be used to infer channel quality and stability [5]. In the second case, they can also be used to predict trends in user requests so as to deal with future service demands accurately.
For resource allocation in RAN slicing, on the intraslice level, resources, such as power and resource blocks, can be flexibly and effectively allocated to users according to different requirements. On the interslice level, resource allocation can refer to resource management across slices, where the goal is to share resources effectively. To solve the problems of service demand and network environment fluctuation, various RL techniques that have been widely studied in the resource allocation field have significant advantages [7].
6G TN slicing virtualizes resources, such as physical ports, nodes, and links, to provide network transmission performance on demand, in which different resource isolation features can be achieved through preconfigured ports and network resource allocation. Flexible bandwidth adjustment and certain network recovery capabilities are the basis of realizing multigranularity network assurance and highly reliable service provisioning.
6G TN slices will carry massive data of emerging applications with high dynamics. When user subscriptions and requests suddenly increase and the QoS cannot be met, the initial bandwidth owned by the TN slice must be scaled to satisfy the performance requirements. RL-based strategies are beneficial in this regard by interacting with the dynamic environment and improving its elasticity, where a software-defined networking (SDN) framework can be used to grasp the global view and act as an intelligent agent to make decisions [11]. ML/DL-based traffic prediction methods can also be combined for bandwidth adjustment.
TN slicing is the bearer of massive data in 6G, and even a brief network interruption caused by an attack or failure can cause great harm. Therefore, to ensure the reliability of 6G TN slicing, selective multipath backup can be provided for quick switchover, and intelligent path splitting and reuse using AI-assisted methods, such as RL, can be carried out to adaptively compress the bandwidth of some services and ensure the QoS of most services.
6G CN slicing is mainly implemented based on virtualization technology and microservice architecture to achieve the goals of flexibility, resilience, and cost efficiency. Virtual network functions (VNFs) can be selected and combined based on different specifications and deployed at different locations to meet various service requirements. AI-assisted VNF scheduling, resource reallocation, and anomaly detection are typical 6G CN slicing techniques.
VNFs are the main component of 6G CN slices, which can decouple the network function from the traditional dedicated devices and be deployed in a shared or exclusive way so as to achieve flexible modular assembly. The operation of CN slicing involves VNF placement and scaling. For both, ML/DL-based algorithms can be used to predict traffic and resource demand and then determine the migration and adjustment strategy of VNFs. The algorithms’ main goals include ensuring connectivity, reducing deployment costs, and guaranteeing the QoS.
Network traffic and slice requests from various 6G applications may fluctuate due to human activities and social events, resulting in the loss of the original optimized strategies. To cope with this dynamic nature and maximize the utilization of network resources, CN slicing can be proactively reconfigured, including the reconfiguration of virtual nodes and links. At this point, RL-based intelligent slice reconfiguration can be constructed for such uncertain situations. However, it also faces the challenge of the dimensional curse in complex network environments, which requires special attention.
Since the CN is the management center of the 6G system, it also becomes a coveted target for attackers. If the CN slice is attacked or one of its components crashes, widespread failures will occur. To assure network reliability, the intelligent anomaly detection system for CN slicing is very important. Due to the variety of fault forms and changing attack methods, transfer learning (TL) can be introduced to adjust the existing model and adapt to the new anomaly forms. The reuse of the model can not only accelerate the speed of anomaly detection but also improve its adaptability.
The realization of 6G end-to-end network slicing is inseparable from the enhanced slicing management system, which has the functions of slice design, slice deployment, and full lifecycle management to match differentiated service requirements with network resources, as detailed in Figure 3. AI-assisted SLA decomposition, slice capacity forecasting, and slice lifecycle management are all promising techniques in 6G slicing management.
Figure 3 Enhancing network assurance and service provisioning with AI-assisted slicing management.
6G end-to-end network slicing runs across multiple domains and faces multidimensional SLAs in terms of latency, throughput, reliability, and so on. Therefore, SLAs need to be decomposed and allocated to each participating domain to facilitate operators’ planning. AI-assisted SLA decomposition is the key to automating 6G complex business processes, in which unsupervised learning (UL), such as clustering algorithms, can first be used to judge the feasibility of a given SLA and filter its features to be decomposed [12]. Then, the SLA decomposition can be predicted by supervised learning (SL) using slice history data. A model library can also be built for TL and model sharing.
The resource management of 6G slicing is a challenging task that needs to allocate resources to each slice and meet time-changing service requirements. Since continuous oversupply will reduce resource utilization efficiency [13], AI-driven short-term and long-term cost-aware slice capacity forecasting can be considered in addition to the traditional traffic demand prediction methods, thus maximizing economic benefits.
Managing 6G network slicing that involves operations in multiple domains is a great challenge. Therefore, an automated slice management platform integrated with AI technology can be built, which can guarantee the whole lifecycle of network slicing. Here, service assurance is the design goal of smart slice automation, which, in turn, provides an implementation guarantee for service assurance. Specifically, service assurance data can help to fully perceive the frequent changes of slices and make optimal decisions to achieve a specific SLA. These data can be obtained by the slicing monitoring module and processed by the smart analysis module. The analysis results are then passed to the reporting module to help dynamically adjust the orchestration policy, thus achieving end-to-end network assurance that includes all domains.
Since UAVs and satellites are essential to support global access, they can act as slice extensions to achieve seamless service coverage. Due to the highly dynamic nature of air and space network topologies and the complexity of collaborative management, network slicing in this situation will face lots of problems, such as dynamic node management, complex slice control, heterogeneous network collaboration, and so on. Recent advances in AI-assisted slicing provide an alternative approach for learning complex networks and making intelligent decisions adaptively, which can help improve the performance of slice extensions.
UAVs are considered one of the most important extensions of 6G slicing, due to their flexibility and fast response capability, closing communication gaps in natural disaster scenarios, busy urban areas, and remote areas. In addition, as many emerging 6G applications require real-time processing, to ensure ultralow delay, UAVs equipped with edge computing facilities can be used to expand 6G slicing [14]. Here, AI-based techniques are suitable for predicting the distribution of mobile users, planning the path of UAVs, and performing intelligent task offloading so as to realize network assurance and provide users with better services.
To achieve the seamless service coverage of 6G slicing, the satellite network is essential, and it is suitable for providing and maintaining network access in areas with no ground infrastructures and achieving seamless service coverage. When the workload of the ground network is too large, the satellite network can also help relieve the communication load. Considering the characteristics of satellite and ground communications, RL-based schemes can be applied for joint resource allocation in ground and satellite networks [1]. In addition, other AI-based techniques can predict user request patterns and adaptively assign them to ground and satellite links.
There are many model training enforcements to satisfy diversified network assurance and support the requested service provisioning in 6G. For SL that plays an important role in network condition and slicing requirement prediction, a model can be derived from massive training data and then used to make predictions. For UL that is beneficial to user classification, by learning statistical rules from unlabeled data, users can be subdivided from multiple dimensions so as to provide customized services accordingly. RL is suitable for adaptive slice resource allocation and management, in which the RL agent can interact with the unknown environment to self-optimize the system performance. FL helps to solve the privacy problems related to privacy protection and data islands, where the local devices train local models and distribute the parameters to the central server to implement a secure aggregation. Also, TL has advantages in accelerating the training process and can apply the experience learned from previous tasks to other related scenarios. Typical AI techniques for different domains mentioned in this section are illustrated in Figure 2. Also, the recommended AI assistance for 6G network slicing is summarized in Table 1.
Table 1 The recommended AI assistance for 6G network slicing in different domains.
Since monitoring, prediction, and resource allocation from the interslice level often require centralized training, privacy and security protection becomes an important issue. At present, homomorphic encryption and differential privacy are the mainstream techniques for handling it. Homomorphic encryption allows direct operations on ciphertext, and the obtained decryption result is consistent with the result of the plaintext operation. In this way, slices can upload training data to the centralized server in the form of ciphertext, while the server does model training but does not know the original training data, thus achieving privacy and security. Then, differential privacy aims to randomize the original data in a certain range and add some acceptable noise data to the original data. This technique makes it difficult for adversaries to infer sensitive information and is generally used in ML to protect sensitive data.
The TN slice can be regarded as a virtual network whose bandwidth, reliability, and other characteristics are qualified by the service demander to meet various requirements. In this regard, elastic optical networks (EON) can support fine-grained spectrum slot allocation, transmission format selection, and elastic bandwidth scaling through bandwidth variable transponders and flexgrids, making it one of the most potentially enabling technologies for TN slicing [15]. Since the available resources vary among different links, both spectrum allocation and routing path selection should be scheduled. In view of this, we propose an EON-enabled TN slicing scheme, which utilizes RL’s advantages of environment state sensing, interactive learning, and reward reinforcement to achieve elastic bandwidth allocation.
First, a bandwidth scaling request can be represented as a quadruple, including the index of the requested slice, the start point and endpoint of the bandwidth scaling, and the scaling amount. To satisfy the request, we need to compute a routing path from the start point to the endpoint and assign consecutive spectrum slots on each link of this path according to the request. Note that the scaling amount W needs to be translated into the number of frequency slots required N, which can be obtained by ${N} = {\left\lceil{W} {\big{/}} {\lambda}_{x}\,{\cdot}\,{D}_{\text{BPSK}}\right\rceil}$, where ${D}_{\text{BPSK}}$ is the data rate that one spectrum slot of the binary phase-shift keying (BPSK) signal can support and ${\lambda}_{x}$ is a multiple of the data rate generated by different modulation formats. For example, when x is set as BPSK, quaternary phase-shift keying, and eight-ary quadrature amplitude modulation, ${\lambda}_{x}$ is one, two, and three, respectively.
As can be seen from the “AI-Assisted 6G Slicing in Different Domains” section, RL is applicable to sequential decision making in real situations, and lots of research has applied RL to intelligent slicing methods. In our case, deep Q-learning (DQN) is adopted to find a path in a TN slice by matching links and spectrum slots for bandwidth scaling requests step by step. Here, the TN domain controller can act as a proxy for this scheme to observe the network state, interact with the environment, and maximize the long-term reward. In the following, we briefly describe the three key elements of the RL-based scheme.
The occupancy condition of spectrum slots in all links and the current completion status of the bandwidth scaling request are included in the state. The latter further includes the contents of the aforementioned quadruple except that the start point of bandwidth scaling changes depending on the action selection.
Based on the observed state, action has two parts. The first part is the next hop selection at the start point of the current path, while the second part is the free spectrum slots selection.
The goal of the elastic bandwidth scaling scheme is to find a path that meets the bandwidth requirements. Therefore, if the endpoint is successfully reached under the condition of satisfying the requirements, the reward is set as one; otherwise, it is zero.
Here, we consider only the case of a bandwidth increase. If a bandwidth decrease is required, resources can be directly released, and the occupancy condition of spectrum slots in the state will change accordingly.
We used the Abilene topology with 11 nodes and 14 links as the substrate TN topology and set up three TN slices, all with 12 links. Each link was set with 10 spectrum slots. For the bandwidth scaling request, values for its quadruple were randomly generated, and its arrival intervals followed an exponential distribution with an average of 500 episodes. For the DQN setup, its learning rate, batch size, replay buffer, and target network update frequency were 0.001, 32, 1,000, and 200, respectively. The greedy level started at 0.1 and decreased as the training progressed.
First, we compare our proposed RL-based bandwidth scaling scheme with the shortest path (SP)-based scheme. The latter is to find the SP and traverse the free continuous spectrum slots of each link of the path to meet the bandwidth requirement. In Figure 4, we can see that under different scaling ratios (the ratio of the bandwidth scaling requirement to the total spectrum resources) and initial spectrum slot occupancies of substrate TN, the proposed RL-based scheme has a higher request satisfaction rate than the SP-based scheme. This is because the links of the SP might not have satisfactory continuous spectrum slots, and our scheme can learn through interaction with the environment and reward incentives to meet the requirements. In addition, with the increase in the scaling ratio, the request satisfaction rate of both schemes decreases significantly since there are fewer eligible continuous spectrum slots in the network.
Figure 4 The performance of the RL-based bandwidth scaling scheme under different scaling ratios and spectrum slot occupancies (OPs) of substrate TN.
Subsequently, we observe the elasticity and convergence performance of the RL-based bandwidth scaling scheme. As graphed in Figure 5, when a new request arrives, our proposed scheme can quickly adapt to the environment and meet the requirements of the request. Here, the number of steps also represents the length of the path found. To avoid excessive attempts, we set the maximum step size of each episode to 50. The adaptivity of the proposed scheme lies in the self-learning ability of RL, which can be trained by obtaining samples during interaction with the environment, saving the process of obtaining a large amount of labeled data and being closer to the learning process of human beings.
Figure 5 The convergence performance of the RL-based bandwidth scaling scheme with changing requests.
With the deepening of network slicing research, there are still some problems worth noting and solving:
In this article, we carried out a comprehensive study on AI-assisted 6G network slicing for network assurance and service provisioning, including promising characteristics as well as AI-assisted techniques for RAN/TN/CN slicing, management systems, and slice extensions. In addition, we proposed an RL-based elastic bandwidth scaling scheme, which has significant advantages in increasing the request satisfaction rate and adapting to environmental changes. Finally, we presented the potential challenges and open issues faced by 6G network slicing. In future work, we will further consider detailed AI-assisted automated slicing mechanisms for 6G, hoping to make more contributions to its development.
This work was supported, in part, by the National Key R&D Program of China (grant 2022YFB3104200), National Natural Science Foundation of China (grant 62202386), Basic Research Programs of Taicang (grant TC2021JC31), Fundamental Research Funds for the Central Universities (grant D5000210817), Xi’an Unmanned System Security and Intelligent Communications ISTC Center, and Special Funds for Central Universities Construction of World-Class Universities (Disciplines) and Special Development Guidance (grants 0639022GH0202237 and 0639022SH0201237).
Jiadai Wang (wangjiadai@nwpu.edu.cn) received her Ph.D. degree from the School of Cyber Engineering, Xidian University, in 2021. She is currently an associate professor with the School of Cybersecurity, Northwestern Polytechnical University, Xi’an 710072 China. Her research interests include network slicing, software-defined networking, and 5G/6G communications. She is a Member of IEEE.
Jiajia Liu (liujiajia@nwpu.edu.cn) is a full professor and vice dean with the School of Cybersecurity, Northwestern Polytechnical University, Xi’an 710072 China. His research interests include intelligent and connected vehicles, mobile/edge/cloud computing and storage, and Internet of Things security. He has published more than 180 peer-reviewed papers and currently serves as an editor of IEEE Network and IEEE Transactions on Wireless Communications. He is a Senior Member of IEEE.
Jingyi Li (jyli_xd@163.com) received her B.S. degree in electronic information science and technology from Henan Agricultural University in 2018. She is currently working toward her Ph.D. degree in the School of Cyber Engineering, Xidian University, Xi’an 710071 China. Her research interests include aerial–ground communications, nonconvex optimization, and intelligent reflecting surfaces. She is a Student Member of IEEE.
Nei Kato (kato@it.is.tohoku.ac.jp) is a full professor at and the dean of the Graduate School of Information Sciences, Tohoku University, Sendai 980-8579 Japan. His research interests include computer networking, wireless mobile communications, and satellite communications. He has published more than 500 peer-reviewed papers and is a past editor-in-chief of IEEE Transactions on Vehicular Technology and IEEE Network. He is a Fellow of IEEE, the Engineering Academy of Japan, and the Institute of Electronics, Information, and Communication Engineers.
[1] T. K. Rodrigues and N. Kato, “Network slicing with centralized and distributed reinforcement learning for combined satellite/ground networks in a 6G environment,” IEEE Wireless Commun., vol. 29, no. 1, pp. 104–110, Feb. 2022, doi: 10.1109/MWC.001.2100287.
[2] Z. Zhang et al., “6G wireless networks: Vision, requirements, architecture, and key technologies,” IEEE Veh. Technol. Mag., vol. 14, no. 3, pp. 28–41, Sep. 2019, doi: 10.1109/MVT.2019.2921208.
[3] M. H. C. Garcia et al., “A tutorial on 5G NR V2X communications,” IEEE Commun. Surveys Tuts., vol. 23, no. 3, pp. 1972–2026, thirdquarter 2021, doi: 10.1109/COMST.2021.3057017.
[4] J. Wang and J. Liu, “Secure and reliable slicing in 5G and beyond vehicular networks,” IEEE Wireless Commun., vol. 29, no. 1, pp. 126–133, Feb. 2022, doi: 10.1109/MWC.001.2100282.
[5] S. Bakri, P. A. Frangoudis, A. Ksentini, and M. Bouaziz, “Data-driven RAN slicing mechanisms for 5G and beyond,” IEEE Trans. Netw. Service Manag., vol. 18, no. 4, pp. 4654–4668, Dec. 2021, doi: 10.1109/TNSM.2021.3098193.
[6] “Architecture enhancements for 5G system (5GS) to support network data analytics services (3GPP TS 23.288 version 16.4.0 release 16),” ETSI, Valbonne, France, ETSI TS 123 288 V16.4.0, 2021.
[7] J. Mei, X. Wang, K. Zheng, G. Boudreau, A. B. Sediq, and H. Abou-Zeid, “Intelligent radio access network slicing for service provisioning in 6G: A hierarchical deep reinforcement learning approach,” IEEE Trans. Commun., vol. 69, no. 9, pp. 6063–6078, Sep. 2021, doi: 10.1109/TCOMM.2021.3090423.
[8] M. Mozaffari, X. Lin, and S. Hayes, “Toward 6G with connected sky: UAVs and beyond,” IEEE Commun. Mag., vol. 59, no. 12, pp. 74–80, Dec. 2021, doi: 10.1109/MCOM.005.2100142.
[9] X. Lei, M. Wu, F. Zhou, X. Tang, R. Q. Hu, and P. Fan, “Reconfigurable intelligent surface-based symbiotic radio for 6G: Design, challenges, and opportunities,” IEEE Wireless Commun., vol. 28, no. 5, pp. 210–216, Oct. 2021, doi: 10.1109/MWC.121.2100126.
[10] Y.-J. Liu, G. Feng, Y. Sun, S. Qin, and Y.-C. Liang, “Device association for RAN slicing based on hybrid federated deep reinforcement learning,” IEEE Trans. Veh. Technol., vol. 69, no. 12, pp. 15,731–15,745, Dec. 2020, doi: 10.1109/TVT.2020.3033035.
[11] J. Huang, Y. Qian, and R. Q. Hu, “Secure and efficient privacy-preserving authentication scheme for 5G software defined vehicular networks,” IEEE Trans. Veh. Technol., vol. 69, no. 8, pp. 8542–8554, Aug. 2020, doi: 10.1109/TVT.2020.2996574.
[12] M. Iannelli, M. R. Rahman, N. Choi, and L. Wang, “Applying machine learning to end-to-end slice SLA decomposition,” in Proc. 6th IEEE Conf. Netw. Softwarization (NetSoft), 2020, pp. 92–99, doi: 10.1109/NetSoft48620.2020.9165317.
[13] D. Bega, M. Gramaglia, M. Fiore, A. Banchs, and X. Costa-Perez, “Deepcog: Optimizing resource provisioning in network slicing with AI-based capacity forecasting,” IEEE J. Sel. Areas Commun., vol. 38, no. 2, pp. 361–376, Feb. 2020, doi: 10.1109/JSAC.2019.2959245.
[14] H. Guo and J. Liu, “UAV-enhanced intelligent offloading for internet of things at the edge,” IEEE Trans. Ind. Informat., vol. 16, no. 4, pp. 2737–2746, Apr. 2020, doi: 10.1109/TII.2019.2954944.
[15] N. Shahriar et al., “Disruption minimized bandwidth scaling in EON-enabled transport network slices,” IEEE J. Sel. Areas Commun., vol. 39, no. 9, pp. 2734–2747, Sep. 2021, doi: 10.1109/JSAC.2021.3064643.
Digital Object Identifier 10.1109/MVT.2022.3228399