The Levels of Intelligence of Mobile Networks and Consideration of Architecture Evolution

Yan Wang, Hua Huang, Yingzhe Li, Wei Zhou, Wireless Network Research Department, Huawei Technologies, Shanghai, China
{jason.wangyan, hua.huang,, will.zhou}

Chih-Lin I, Qi Sun, Siming Zhang, China Mobile Research Institute, Beijing, China
{icl, sunqiyjy, zhangsiming}

IEEE Future Networks Tech Focus: Volume 2, Number 3, December 2018 


The rejuvenation of AI technology provides a new way to solve the increasingly complex and difficult problems in the 5G network, which makes the network more intelligent and autonomic. This field has gradually become a hot topic both in the academia and industry. But the industry has not yet reached a unified definition of an intelligent mobile network, and how to measure and judge the level of intelligence. This paper attempts to give the definition of the levels of mobile network intelligence and analyzes the influence of intelligence on the evolution of wireless network architecture, hoping to help the industry to reach consensus.

1. The state of the art of the mobile network with AI
In the evolution process from 4G to 5G, the performance and flexibility of wireless networks have changed fundamentally. For the performance, to support the three typical services of enhanced Mobile Broad Band (eMBB), massive Machine Type Communications (mMTC) and Ultra-Reliable and Low Latency Communication (URLLC), 5G network introduces advanced technologies such as large-scale antenna array, flexible air interface and Non-Orthogonal Multiuser Access technology to meet the more stringent technical requirements in terms of peak rate, spectrum efficiency, low delay, high reliability, connection density. For the flexibility, various decoupling are happened in the 5G architecture, such as software and hardware decoupling for the Network Function Virtualization (NFV), control and forward decoupling for the gateway, control plane function decomposition, Central Unit (CU) and Distributed Unit (DU) separation of Radio Access Network (RAN), etc. This makes network functions easier to deploy on the cloud computing platform, enabling automated orchestration and deployment of network functions to provide efficient network slicing services for different vertical application requirements.

Although 5G has brought a qualitative leap in performance and flexibility, the richer KPI dimensions, flexible air interface, virtualization of network function and the introduction of slicing technology lead to the utmost complexity and challenges of the design, deployment operation and optimization of the 5G network. Artificial Intelligence (AI) technology, which is rejuvenated by the great development of big data, deep learning and cloud computing, provides a data driven methodology worthy of exploration for solving the complicated problems that 5G network faces [1]. The application of AI to wireless networks has attracted more and more attention in the academic field. There are many research literatures in AI assisted New Radio (NR) resource allocation, cloud resource management, receiver design, channel parameter estimation and so on[2][3][4][5].

However, mobile network empowered by AI is not a single point technology or internal implementation problem but requires systematic thinking on the architectural level. 5G Americas proposed that orchestration, analytics and automation enabled by AI or Machine Learning (ML) will play a key role in 5G network [6]. Some research projects of 5G Infrastructure Public Private Partnership (5GPPP) tries to use AI and ML to a achieve real time autonomous 5G network management [7][8][9]. Many pre-standard and standardization organizations have carried out discussions and research items on network intelligence based on Big data and AI technology. For example, the European Telecommunications Standards Institute (ETSI) set up the Experiential Network Intelligence Industry Specification Group (ENI ISG) and the Zero touch network and Service Management Industry Specification Group (ZSM ISG) focusing on intelligent closed loop policy and fully automatic network management, respectively[10][11]. ITU-T set up Focus Group on Machine Learning for Future Networks including 5G (FG-ML5G) investigating on valuable use cases, data model and algorithms, and network intelligent architecture [12]. 3GPP has approved data driven related study items, e.g., Enables of Network Automation in 3GPP SA2 working group [13] and RAN-Centric Data Collection and Utilization in 3GPP RAN3 working group [14].

2. The Lack of unified definition for mobile network intelligence
Currently different organizations and research institutions have different views on the application of AI into the mobile network. They pay more attention to the specific layer or domain in which the AI or Big data can be used to improve the efficiency and performance, making the research relatively fragmented. For instance, ETSI ENI ISG mainly focuses on policy architecture, ETSI ZSM ISG focuses on management, 3GPP focuses on control plane and Self-Organized Network (SON), and academic community pays more attention to the physical layer and new application scenarios (such as cache in wireless network).

Can the mobile network be considered as intelligent when the AI is only used in mobile networks to solve certain specific problems at certain specific layer or domain? Let's first examine the following questions:

  • Is AI enabled operation and maintenance an intelligent network? The intelligent operation and maintenance is the primary value of the AI for network. AI helps to improve the optimization of parameter configuration and the efficiency of fault prediction and diagnosis, minimize the manual intervention and reduce OPEX. However, intelligent operation and maintenance only maintains the existing network to its best level and does not make the network itself to have intelligence such as context awareness, service awareness, and automatic policy control and resource scheduling.
  • Is AI enabled SON an intelligent network? Self-Organizing Network (SON), defined by 3GPP, includes self-configuration, self-optimization and self-healing. In fact, it has embodied the characteristics of network intelligence. But 3GPP's definition of SON is limited to certain specific features, such as the Automatic Neighbor Relation (ANR), Mobility Load Balancing (MLB), Mobility Robustness/Handover Optimization (MRO), Inter-Cell Interference Coordination (ICIC), Coverage and Capacity Optimization (CCO), Cell Outage Detection (COD), Cell Outage Compensation (COC) and so on [15]. These SON features usually designed independently and heavily rely on the standardization resulting in a bundle of chimney-like features. AI enabled SON may be just a better SON, but it won’t change its limited scenarios and dependence on the standard process.
  • Is AI enabled Radio Resource Management (RRM) or Radio Transmission Technology (RTT) an intelligent network? Recently, academic community have tried to study AI technology in radio resource management level and physical level, such as Adaptive Modulation and Coding (AMC), Massive MIMO beam forming, etc., to exploit potential performance with reduced complexity [16]. For these research, AI is more of an auxiliary means of existing algorithm or an alternative to the traditional algorithm, which cannot be regarded as a system-level intelligence from the whole network perspective.

In summary, the use of AI technology at any layer or domain of the mobile network can solve some complex problems and bring a certain degree of intelligence, which unfortunately does not demonstrate the intelligence of the whole network. Network intelligence should be a system-level concept, not a single function. This leads to several further questions:

  • What is the ultimate goal of realizing wireless network intelligence with AI?
  • How to evaluate the level of wireless network intelligence?
  • What is the impact of different level of intelligence on the existing wireless network architecture?
  • How will the network architecture evolve towards the goal of intelligence?

3. The definition of mobile network intelligence
The grading definition of automated driving provides a good reference for us to understand and judge the levels of intelligence of mobile networks. To unify the understanding of the concept of automated driving in the whole industry, the SAE (International Automotive Engineering Society) J3016 document provides a taxonomy with detailed definitions for 6 levels of driving automation [17]. The proposal has become a widely accepted standard in the vehicle industry and is used to guide the vehicle industry to carry out five phases of work for driving automation in stages [17].

Introducing intelligence into vehicles achieves fully automated driving, while the goal of wireless network intelligence is the "network autonomy". The network can automatically deploy, configure, and optimize by itself to achieve target KPI according to the intention of the operators, and can automatically avoid or solve abnormal events to ensure the security and reliability of the network. But the ultimate objective of full autonomy cannot be achieved overnight. It needs to be implemented step by step. Borrow the idea from SAE, we also need to define several features related to the level of mobile network intelligence and determine the level of network intelligence by analyzing the degree of substitution for AI subsystem with the defined features.

Here we try to give 7 features for investigating network intelligence levels:

  1. Context awareness and analysis. It is to tell what happened inside the network and what is the root cause;
  2. Non-real time prediction and inference. It is to judge what will happen in the network management plane in future (maybe several minutes or hours later);
  3. Decision-making and execution. It is to make policy and take control measures automatically based on the prediction and reasoning;
  4. Real-time prediction and inference. It mainly aims at prediction and inference at milliseconds or microsecond scale at L1-L3 layer of RAN with real-time control and scheduling;
  5. Exception handling. It refers to the handling of sudden or extreme events and recovery from abnormal outage;
  6. Human-Network Interface. It refers to whether operator interact with the network through traditional specialized signaling or command interaction, or through intention-based interface;
  7. Applicable scenarios. It refers to whether the intelligent scenarios cover specific functionalities, specific services, or integrated scenarios in the complete network life cycle.

Based on the degree of substitution for the AI subsystem in the 7 defined features, we can classify the intelligence of the mobile network into 6 levels, as shown in Figure 1.

 Figure 1. The levels of mobile network intelligence

 Figure 1. The levels of mobile network intelligence

As shown in Fig. 1, the higher the level of network intelligence, the more work of human operators are replaced by AI subsystems in the 7 features. It is worth noting that a certain level of intelligence is reached only when all the features of this level are implemented by the AI subsystem. For example, to achieve level 2 intelligence, the AI subsystem must simultaneously support feature 2 and feature 3. However, if the AI subsystem only supports part of the feature required, it cannot be regarded as Level 4 intelligence.

The benefits of such intelligence grading are as follows:

  • It helps the industry to reach a consensus of the definition of intelligent wireless networks;
  • It provides some criteria for judging the development level of the wireless network intelligence;
  • It provides decision-making basis for the government, operators, equipment vendors and other related industry partners to do technology selection, product planning, etc.

4. The impact on the mobile network architecture evolution

Introducing Big data and AI technology into wireless networks to achieve network intelligence will definitely impact the existing mobile network architecture. We believe that AI subsystems will be introduced in the three domains of the network, i.e., the Operations Support System (OSS) plane, the Core Network (CN), and the Radio Access Network (RAN.) 3GPP has approved two study items related to network data analysis, i.e., Enabler of Network Automation in SA2 and RAN-Centric Data Collection and Utilization in RAN3, which will accelerate the standardization pace. The core network side has defined the Network Work Date Analysis (NWDA) function as the hosting entity of big data and AI subsystem. The RAN side study item which approved in June may also study the possibility of introducing a Radio Data Analytics (RDA) function (i.e. NWDA-like functional entities) on the RAN side.

 Figure 2. The Introduction of AI subsystem in the 3GPP network architecture

 Figure 2. The introduction of AI subsystem in the 3GPP network architecture


The application scenarios and intelligent level that each domain requires are different, as shown in Figure 2. At the OSS, it usually covers the network planning related use cases, including coverage prediction, fault diagnosis, cell edge throughput enhancement, and virtual grid based multi frequency parameter optimization and so on. The prediction is non-real-time and usually beyond minutes’ level. According to the aforementioned intelligent level definition, the OSS intelligence can reach level 3 at the most because it cannot achieve near-real-time and real-time prediction and inference needed at the control layer and the physical layer. 5G CN with introduced AI subsystem can achieve control plane related intelligent scenarios, such as the intelligent QoS control, the personalized mobility management, the load-balance of the Virtualized Network Function (VNF), network traffic prediction, etc. The prediction timescale is usually larger than second level. Since the core network cannot handle the real-time resource control at the TTI level in RAN, the core network intelligence can reach level 3 or at most level 4. If the AI subsystem is introduced in the RAN, it can handle the near real-time and real-time RRM/RTT related intelligent scenarios, including Smart AMC, intelligent multiuser Pairings, AI decoders, deep learning based digital pre-distortion, etc., and the prediction timescale can be even reduced to the microsecond level. It is possible to realize intelligence of level 4 at most since only the scenarios within RAN are considered.

Only when the AI subsystems of all three domains collaborate, the network intelligence of level 5, i.e., the fully autonomous network, can be reached.

 Figure 3. The architecture evolution map according to the network intellgent levels

Figure 3. The architecture evolution map according to the network intelligenct levels

Based on the above analysis, Figure 3 shows the relationship between the architecture evolution and network intelligent grading. If only the network intelligence or autonomy of L1~L3 level is needed, there is no impact on the 3GPP network architecture, and the intelligence is mainly embodied in the OSS, or the management and orchestration layer. At this stage if the AI technology is utilized to improve the performance and efficiency within base station, since this is only an internal implementation issue, there is still no need for architectural support.

However, if we need to achieve intelligence above level 3, network architecture is required to be upgraded. To achieve Level 4 intelligence, we need to further introduce AI subsystem in CN and RAN to increase near-real-time and real-time prediction and reasoning capabilities. In addition to the interaction between the three AI subsystems, it may be necessary to interact between the wireless network AI system, cloud platform and UE to achieve    cross-layer or cross-domain coordination, thereby significantly expanding the network intelligent application scenarios. For the Level 5 intelligence, the AI will be already a ubiquitous capability of the network, and it will bring about a fundamental change in the human-network interface, realizing an intent-driven network, free of manual operation and maintenance. To realize the network intelligence of Level 4-5, the network architecture will have great changes in functions, interfaces and procedures, as well as innovations in chips and algorithms.

5. Conclusions

The introduction of AI into mobile networks has been highly expected by the industry for solving complex problems at various network levels, and ultimately enabling an intelligent and fully autonomous network. This paper attempts to give a definition and grading of mobile network intelligence from a research perspective. We propose a taxonomy with detailed definitions for 6 levels of intelligence and 7 key features to be fulfilled. Hopefully such discussion will serve as an anchor in reaching a unified understanding of the definition of intelligent mobile networks and its evolution path to the ultimate intelligence and autonomy. Although this paper proposes intelligent grading for mobile networks, the conclusion is also applicable to the definition of general network intelligence. The current thinking is still relatively elementary. We welcome contribution and discussion from academic community and industrial organization to further improve the taxonomy of the intelligence grading for mobile networks.


  1. You X H, Pan Z W, Gao X Q, et al. The 5G mobile communication: The development trends and its emerging key techniques. Sci Sin Inform, 2014, 44 (5): 551-563 [. 5G. 2014, 44 (5):551-563
  2. Wang T, Wen C K, Wang H, et al. Deep learning for wireless physical layer: Opportunities and challenges[J]. China Communications, 2017, 14(11): 92-111.
  3. Tang P, Li F, Zhou W, et al. Efficient auto-scaling approach in the telco cloud using self-learning algorithm[C]//Global Communications Conference (GLOBECOM), 2015 IEEE. IEEE, 2015: 1-6.
  4. Zhou W. TCRM: Telco Cloud Resource Management Using Real-Time Data Analysis[C]//Cloud Computing Technology and Science (CloudCom), 2016 IEEE International Conference on. IEEE, 2016: 480-481.
  5. O’Shea T, Hoydis J. An introduction to deep learning for the physical layer [J]. IEEE Transactions on Cognitive Communications and Networking, 2017, 3(4): 563-575.
  6. 5G Americas white paper. 5G Network Transformation.
  7. SELFNET. Framework for Self-Organized Network Management in Virtualized and Software Defined Networks. [Online]. Available:
  8. CogNet. Building an Intelligent System of Insights and Action for 5G Network Management. [Online]. Available:
  9. SESAME. Small cEllS coordinAtion for Multi-tenancy and Edge services. [Online]. Available:
  10. ETSI ENI White paper, Improved operator experience through Experiential Networked Intelligence,
  11. ETSI ZSM ISG White paper, Zero-touch Network and Service Management,
  12. ITU-T FG ML5G ToRs,
  13. 3GPP Study of Enablers for Network Automation for 5G (Release 16), 3GPP TR 23791
  14. 3GPP New Study Item Proposal, RAN Centric Data Collection and Utilization, RP-181456
  15. 3GPP, “Technical Specification Group Services and System Aspects; Telecommunication Management; Self-Organizing Networks (SON); Concepts and requirements (Release 13),” 3GPP, TS 32.500, v13.0.0,
  16. N. Samuel, T. Diskin, and A. Wiesel, "Deep MIMO detection," IEEE 18th Int. Workshop Signal Process. Advances Wireless Communication (SPAWC), pp. 690–694, 2017.


YanWangYan Wang received his Ph.D. degree from the department of Electronic Engineering, Shanghai Jiao Tong University, China, in 2009.  He is a principal engineer at Huawei Technologies in Shanghai, China. He has led many research projects on the mobile network evolution, including EPC enhancement, NFV, SDN, CU separation, service chaining, MEC mobility, Mobile LAN etc. Many of the research outputs have contributed to the 3GPP 4G and 5G standards. He is now leading a team of future network evolution research and prototype, and he is also the leader of architecture group of Wireless Artificial Intelligence Alliance. His current interesting of study includes intelligent mobile network enabled by big data and Artificial Intelligence, as well as mobile deterministic networks



HuaHuangHua Huang is the director of Huawei Mobile Broadband Network Research Department. He is also the chief expert of the wireless architecture research in Huawei. Mr. Hua graduated from Zhejiang University with a Master degree in telecom science in June 1996 and joined Huawei in 2000. From 2000 till 2003, Hua is a senior engineer in Huawei 3G product, responsible for the system design of packet domain equipment. From September 2003 till April 2006, Mr. Hua was the leader of Huawei 3GPP SA2 Standard team and responsible for the system and architecture evolution standard research. He also joins the 3GPP TSG-SA plenary as delegate. Since 2006, Mr. Hua was responsible for Huawei wireless research as a director in access network and architecture areas. His current research is focused on the future wireless architecture, including AI, Big data, SDN, NFV, Cloud Computing, Open source, and other technologies, he is in charge of 5G architecture researches, and also lead the wireless AI research team in Huawei. 


Yingzhe LiYingzhe Li received the Ph.D. degree from Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, China, in 2011. He is a principal engineer at Huawei Technologies in Shanghai, China. He has research on the algorithm of Self-Organizing Networks, such as network optimization, Mobility robust optimization. Now He focuses on the applications of Artificial Intelligence in wireless networks.




Wei ZhouWei Zhou (M’15) received his Ph.D. degree from the department of Electrical Engineering and Information Systems, University of Science and Technology of China, in 2009. Currently, he is a Principal Engineer in Huawei Technologies Co., Ltd. His research interests include wireless communication, wireless Intelligence, machine learning and big data in next generation mobile network.




chih lin I croppedChih-Lin I received her Ph.D. degree in electrical engineering from Stanford University. She has been working at multiple world-class companies and research institutes leading the R&D, including AT&T Bell Labs; Director of AT&T HQ, Director of ITRI Taiwan, and VPGD of ASTRI Hong Kong. She received the IEEE Trans. COM Stephen Rice Best Paper Award, is a winner of the CCCP National 1000 Talent Program, and has won the 2015 Industrial Innovation Award of IEEE Communication Society for Leadership and Innovation in Next-Generation Cellular Wireless Networks.

In 2011, she joined China Mobile as its Chief Scientist of wireless technologies, established the Green Communications Research Center, and launched the 5G Key Technologies R&D. She is spearheading major initiatives including 5G, C-RAN, high energy efficiency system architectures, technologies and devices; and green energy. She was an Area Editor of IEEE/ACM Trans. NET, an elected Board Member of IEEE ComSoc, Chair of the ComSoc Meetings and Conferences Board, and Founding Chair of the IEEE WCNC Steering Committee.

She was a Professor at NCTU, an Adjunct Professor at NTU, and an Adjunct Professor at BUPT. She is the Chair of FuTURE 5G SG, an Executive Board Member of GreenTouch, a Network Operator Council Founding Member of ETSI NFV, a Steering Board Member and Vice Chair of WWRF, a Steering Committee member and the Publication Chair of IEEE 5G Initiative, a member of IEEE ComSoc SDB, SPC, and CSCN-SC, and a Scientific Advisory Board Member of Singapore NRF. Her current research interests center around “From Green & Soft, to Open & Smart”.


Qi SunQi Sun received the B.S.E. and Ph.D. degree in information and communication engineering from Beijing University of Posts and Telecommunications in 2009 and 2014, respectively. After graduation, she joined the Green Communication Research Center of the China Mobile Research Institute. She has published over 20 conference and journal papers and over 20 patent applications. Her research focuses on 5G and 5G+ technologies, including wireless big data driven intelligent RAN optimization, network resource management, non-orthogonal multiple access, etc.



siming zhang croppedSiming Zhang received her Ph.D. degree in wireless communications from the University of Bristol (UK) in 2016. She currently works at the Green Communications Research Center in China Mobile Research Institute (Beijing). She is one of the co-leads on NGMN Trial and Testing Initiative. She is also co-leading WG1 of the Wireless Artificial Intelligence Alliance. She is the Associate Managing Editor of the IEEE 5G Tech Focus Journal. Her current research interests include PHY layer design on Massive MIMO and mmWave, especially on channel modeling and prototyping, wireless big data analysis, and AI application in the RAN domain.


Editor: Panagiotis Demestichas  


Pageviews: 8751

Subscribe to Tech Focus

Join our IEEE Future Networks Technical Community and receive IEEE Future NetworksTech Focus delivered to your email.

Subscribe Now

Article Contributions Welcome


Submit Manuscript

Submit Manuscript via Track Chair

Author guidelines can be found here

Other Future Networks Publications 

Archived Issues 

Contact Us

IEEE Future Networks Tech Focus Editorial Board

Rod Waterhouse, Editor-in-Chief
Mithun Mukherjee, Managing Editor
Imran Shafique Ansari
Anwer Al-Dulaimi
Stefano Buzzi
Yunlong Cai
Zhi Ning Chen
Panagiotis Demestichas
Ashutosh Dutta
Yang Hao
Gerry Hayes
Chih-Lin I
James Irvine
Meng Lu
Amine Maaref
Thas Nirmalathas
Sen Wang
Shugong Xu
Haijun Zhang
Glaucio Haroldo Silva de Carvalho