China Unicom experts: 5G should embrace AI from five aspects

5G uses large-scale antenna systems and ultra-dense networking technology, and will introduce complex wireless transmission technologies such as spectrum sharing and D2D. Compared with previous mobile network technologies, the overall network architecture is more flexible, with richer functions and more diversified services. All of this makes network planning, deployment, management and maintenance extremely challenging, and 5G networks are born with the mission of providing users with intelligent and optimal experience services. Therefore, future 5G networks will have a high degree of autonomy and full flexibility.

[[257084]]

After more than 60 years of evolution, artificial intelligence technology is also accelerating, especially driven by new theories and technologies such as mobile Internet, big data, supercomputing, sensor networks, brain science, and strong demand for economic and social development, showing new features such as deep learning, cross-border integration, human-machine collaboration, group intelligence openness, and autonomous control. Brain-like intelligence inspired by the research results of brain science is ready to go, and the trend of chipization, hardwareization, and platformization is more obvious, and the development of artificial intelligence has entered a new stage. At present, the development of new-generation artificial intelligence-related disciplines, theoretical modeling, technological innovation, and software and hardware upgrades are being promoted as a whole, which will soon trigger a chain breakthrough and promote the acceleration of various fields of the economy and society from digitalization and networking to intelligence.

Five suggestions for 5G to embrace AI

In the 5G era, the combination of network and artificial intelligence will become an inevitable proposition. Operators should seize the historic opportunity brought by the national artificial intelligence development plan, make full use of the technology, products and operation strength of all parties, and promote the transformation of the communications industry to network intelligence, business personalization, industry application intelligence and management intelligence. Operators should also use artificial intelligence technology to improve the efficiency of network planning, construction, maintenance and other aspects, enhance the ability of network intelligent networking, flexible operation, efficient support business, etc., reduce network construction, maintenance and management costs, enhance the competitiveness of their own industry, personal and family businesses, and realize network intelligent transformation. In this process, 5G networks should start from the following five aspects and be ready to embrace AI.

Network data normalization

Data acquisition and processing is a major challenge for applying AI to 5G networks. Mobile communication data has high dimensions, multiple data types, large data volumes, a lot of missing data, and inconsistent data formats from different equipment manufacturers, making wireless data acquisition and processing difficult.

The entire communications industry needs to work together to address the data acquisition and processing issues of AI applications in 5G networks. First, a unified data standard should be formed. For wireless network data, authoritative associations, alliances or national departments should formulate unified data standards, covering data formats, parameter definitions, calculation methods and other aspects to reduce the complexity of data processing; second, extract high-value data to reduce the hardware resources required for data storage and calculation; third, data desensitization, encrypting and encoding data containing user privacy or involving information security, which will effectively protect personal privacy and will not affect the analysis of data by AI algorithms; finally, it is necessary to strengthen distributed parallel processing. For large-scale wireless data sets, a distributed system should be established to process data in parallel and improve processing efficiency.

Capability Openness and Integration

Operators are relatively weak in AI technology, and face problems such as hardware deployment, software development, talent shortage, and cost shortage. In the face of these problems, operators need to combine the strengths of the AI ​​industry, on the one hand, give full play to their own advantages in "cloud, pipe, terminal" and big data applications, and on the other hand, actively cooperate with external partners with deep technical accumulation such as the Internet industry and AI product companies, continuously accumulate AI technical knowledge, and learn from the Internet industry's experience in AI applications, so as to apply artificial intelligence to 5G networks faster and better, and promote the development of networks towards intelligence.

If network-related capabilities can be opened up and AI technology can be introduced for integration to form a network + AI capability opening platform, then AI and the network will fit in very well. The data, transmission, information and other capabilities and resources opened up by the network can enable AI technology to be quickly integrated into the network, laying an important foundation for operators to improve their AI service capabilities, and is also an effective way to make up for shortcomings in AI technology. The adoption of cooperative sharing and "borrowing troops to fight" methods can improve AI service capabilities and build the telecommunications industry's own AI team.

For example, China Unicom Network Technology Research Institute is currently cooperating with AI "unicorn" - Fourth Paradigm. With the feature processing idea of ​​"dimensionality upgrading" of Fourth Paradigm's AutoML product, AI algorithms used in the Internet industry are introduced into the network platform of operators. At the same time, network transmission, scheduling, routing and other capabilities as well as desensitized network, user and business data are delivered to the AI ​​engine through the capability opening platform, realizing the potential connection between individual characteristics and combined characteristics of networks and users and target results through the "dimensionality upgrading" algorithm, thereby improving the accuracy of prediction results in network development, user experience and business needs. Drawing on the idea of ​​"dimensionality upgrading", it can also solve the problems of network quality, user experience evaluation, network fault location, problem tracing and so on. Through simulation tests, compared with classic machine learning algorithms such as decision trees and expert systems used in traditional mobile communication networks, "dimensionality upgrading", a new algorithm adopted by the Internet industry, has brought unexpected results. The accuracy of analysis results has increased from 66% to 79%, breaking through the accuracy bottleneck of traditional methods.

Introducing innovative technology

Existing AI algorithms may not be applicable in complex communication scenarios, and they need to be improved or innovated according to the characteristics of the communication network. For example, in terms of applying AI technology to solve the problems of service experience evaluation and network dynamic optimization, some existing AI methods can well solve the problems of Internet service user experience evaluation and APP function optimization, but they cannot adapt to the multi-factor correlation and environmental complexity of mobile communication networks.

In order to overcome the dynamic characteristics of network status and services and cope with the challenges of diversified multimedia services, China Unicom Network Research Institute and Tsinghua University AI Research Team have improved and innovated the existing AI algorithms, combined with human factors engineering and mobile communication network power, and proposed a communication and service collaborative optimization method for QoE based on reinforcement learning, which maps the user's psychological and physiological perception to the mobile service experience, and then associates the mobile communication KPI with QoE. The model is established through reinforcement learning and feedback learning mechanism to obtain the optimal solution in high-dimensional space. At the same time, the real-time network status and service quality of the output end are fed back to the input end, so as to obtain the highest network resource utilization under the current service demand, so that the user experience is the best, and the dynamic joint optimization of complex services in the mobile network and the ultimate goal of improving QoE are achieved.

AI applications at the margins

5G networks will provide services for a wide range of vertical industry applications, bringing more edge service demands. Multi-access edge computing (MEC) is one of the important technologies of 5G. By providing an information technology service environment and cloud computing capabilities close to mobile users, it can better support the low latency and high bandwidth business requirements of 5G networks. At the same time, MEC naturally has the gene to combine with AI. It is closer to network nerve endings such as data sources and base stations, so it can cooperate with 5G base stations and edge big data systems. AI technology will play an important role in the intelligence of edge business scenarios and the openness of wireless networks.

For example, in response to the surge in requests for media services such as video in communication networks, network congestion, and delays in the distribution of existing video content, artificial intelligence technology can be applied to 5G network MEC cache decisions to improve user experience quality, and intelligently determine the content in the cache device based on the network data collected by each base station. MEC cache solutions based on deep learning can enhance the MEC cache hit rate, so that video requests can be responded to quickly.

Modeling of network environment

The path loss calculation, coverage planning, beamforming, etc. of traditional networks all involve the calculation of the network environment. In the context of the complex 5G network environment, it is necessary to introduce AI to solve problems such as planning optimization related to the network environment. At this time, the traditional algebraic calculation method needs to be modeled based on AI. Accurate modeling in the AI ​​algorithm is crucial to the actual application effect of the algorithm.

On the one hand, communication networks have the characteristics of many scenarios. For different scenarios in communication networks, such as pilot power adjustment, edge throughput improvement, M-MIMO beam adjustment, D-MIMO intelligent cluster allocation, multi-antenna characteristic gain and other scenarios, precise modeling is required. On the other hand, communication networks have the characteristics of strong time-varying. For sudden changes caused by abnormal behaviors of network transmission (such as malicious attacks) or changes in the external environment (such as channel mutations caused by bad weather), it is necessary to establish dynamic learning and continuous learning algorithm models to deal with sudden problems in communication scenarios. For example, establishing an accurate large-scale model of wireless channels is crucial for network design. It can determine the coverage size of the cell, thereby reducing interference from neighboring cells and optimizing the network.

However, the current channel modeling methods mainly rely on channel measurement. The channel model established based on various statistical characteristics of the wireless channel has the disadvantage of being difficult to give accurate channel responses for specific environments and has certain limitations. Using artificial intelligence methods, according to the characteristics of wireless channel data, tasks such as large and small scale fading prediction can be abstracted and classified into regression classification and other problems that machine learning is good at solving. Through machine learning and data mining, more accurate channel fading prediction and simulation methods can be obtained.

The application of AI in the network is still in its infancy

Communication networks are developing in the direction of diversification, broadband, integration, and intelligence. Wireless transmission uses higher and higher spectrum, larger and larger bandwidth, and more and more antennas, so traditional communication methods are too complex and performance is difficult to guarantee. At the same time, with the explosion of smart terminals and various apps, wireless communication network behavior and performance factors are more dynamic and unpredictable than in the past. Operating increasingly complex wireless communication networks at low cost and high efficiency is a challenge currently faced by operators. In addition, social media activities can affect users' network behavior. As the focus of network operation and optimization shifts from network performance to user experience, traditional KPI optimization methods and network planning optimization tools can no longer meet the needs of 5G networks.

There is a large amount of measurement information in network transmission, and the communication network itself also has a large amount of big data such as terminals, services, users, network operation and maintenance, and wireless transmission performance. Making full use of these communication big data, using artificial intelligence methods such as machine learning and deep learning, conducting in-depth mining, and dynamically reconfiguring wireless networks in real time are the core and key to improving network performance and user experience, reducing labor cost investment, and adapting to various new applications.

However, the application of artificial intelligence in the communications field is still in its infancy. The intelligent evolution path of 5G networks faces both challenges and opportunities. Operators need to promote the integration of the two in stages based on the current network status, cloud transformation progress, and 5G technology maturity, and build a new intelligent ecosystem with equipment vendors, Internet companies, research institutions, etc.