When Machine Learning Meets Wireless Cellular Networks: Deployment, Challenges, and Applications Ursula Challita, Henrik Ryden, Hugo Tullberg Ericsson Research, Stockholm, Sweden Email: fursula.challita, henrik.a.ryden, [email protected] Abstract—Artificial intelligence (AI) powered wireless net- such networks to dynamically adapt to the changing network works promise to revolutionize the conventional operation and context in real-time enabling autonomous and self-adaptive structure of current networks from network design to infras- operations. Network devices can implement both reactive and tructure management, cost reduction, and user performance improvement. Empowering future networks with AI functional- proactive approaches, where data driven algorithms would ities will enable a shift from reactive/incident driven operations only replace or complement traditional design algorithms if to proactive/data-driven operations. This paper provides an there is an overall performance gain. In essence, AI techniques overview on the integration of AI functionalities in 5G and can be used to augment existing functions by providing beyond networks. Key factors for successful AI integration such useful predictions as input, replace a rule-based algorithm, as data, security, and explainable AI are highlighted. We also summarize the various types of network intelligence as well as and optimize a sequence of decisions such as radio resource machine learning based air interface in future networks. Use case management and mobility. examples for the application of AI to the wireless domain are then summarized. We highlight on applications to the physical Existing literature have investigated the application of ma- layer, mobility management, wireless security, and localization. chine learning (ML) techniques to wireless networking. The authors in [1]–[3] describe different applications of ML to I. INTRODUCTION wireless communication problems such as radio resource man- Evolution to the 5th generation cellular technology (5G) and agement and channel estimation. Nevertheless, such papers do beyond networks will see an increase in network complexity not investigate deployment issues and network design chal- - from new use cases to network function virtualization, lenges for aligning ML techniques to applications in wireless large volumes of data, and different service classes such as networks. The main scope of this paper is to summarize some ultra reliable low latency communications (URLLC), massive of the main requirements needed for efficiently integrating AI machine type communications (mMTC), and enhanced mobile functionalities in real networks while also highlighting on the broadband (eMBB). The latest standard, new radio (NR), was benefits that AI brings to various wireless applications. designed to be flexible in order to meet the new service requirements. Nevertheless, the increased flexibility implies The main contribution of this paper is to review the role of a growing number of control parameters and an increased AI in wireless networks. First, we summarize some of the key complexity which is forcing a fundamental change in network factors for successfully deploying AI functionalities in future operations. Meanwhile, the recent advances in artificial intelli- networks. We discuss the distribution of network intelligence gence (AI) promise to address the emerging complex commu- - in the device, base stations, and cloud - and elaborate arXiv:1911.03585v2 [cs.IT] 1 May 2020 nication system design. AI promises to combine simplification on ML-based air interface. Second, we highlight on various with improved performance and spectral efficiency and can components, such as secure data exchange and confidential therefore be regarded as a key component for increasing the computing, that are crucial for the successful integration of AI value of 5G and beyond networks, if properly integrated into functionalities in cellular networks. Finally, we present various the system. applications of AI to the wireless domain such as AI-based At a technical level, AI will have a significant role in mobility and AI-assisted localization. shaping future wireless cellular networks - from AI-based service deployment to policy control, resource management, The rest of this paper is organized as follows. Section II monitoring, and prediction. Evolution to AI-powered wire- elaborates on the distribution of network intelligence and ML- less networks is triggered by the improved processing and based air-interface. Section III provides a summary of the key computational power, access to massive amount of data, and factors for the successful integration of AI tools in future enhanced software techniques thus enabling an intelligent networks. Use case examples for the application of AI in radio access network and the spread of massive AI devices. wireless networking are summarized in Section IV. Finally, Integrating AI functionalities in future networks will allow conclusions are drawn in Section V. II. KEY FACTORS FOR SUCCESSFUL AI DEPLOYMENT Centralized AI schemes can be challenging for some wire- less communication applications due to the privacy of some Future wireless networks must support flexible, pro- features such as user location, limited bandwidth, and energy grammable data pipelines for the volume, velocity, and variety availability for data transmission for training and inference. of real-time data and algorithms capable of real-time decision This in turn necessitates new communication-efficient training making. Future communication networks must be designed algorithms over wireless links while making real-time and to support exchange of data, models, and insights, and it is reliable inferences at the network edge. Here, distributed the responsibility of the AI agents to include any necessary machine learning techniques (i.e., federated learning [4]) have user data. In this section, we provide an overview on the dis- the potential to provide enhanced user privacy and energy tribution of network intelligence and ML-based air interface, consumption. Such schemes enable network devices to learn which are key components for realizing such vision in future global data patterns from multiple devices without having networks. access to the whole data. This is realized by learning local A. Distribution of Network Intelligence models based on local data, sending the local models to a centralized cloud, averaging them and sending back the Future wireless networks will integrate intelligent functions average model to all devices. Nevertheless, the effectiveness across several layers of the network such as the wireless of such schemes in real networks should be further studied infrastructure, cloud, and end-user devices with the lower-layer considering the limitations of processing power and mem- learning agents targeting local optimization functions while ory of edge devices. As such, configurations for centralized, higher-level cognitive agents pursuing global objectives and distributed, and hybrid architectural approaches should be system-wide awareness. Such intelligent distribution can be supported. Moreover, it is vital to design a common distributed categorized into three main types, namely autonomous node- and decentralized paradigm to make the best use of local and level AI, localized AI, and global AI. global data and models. • Autonomous node-level AI is used to solve self-contained problems at individual network components or devices, B. ML-based Air Interface where no data is required to be passed through the In addition to the distribution of network intelligence, future network. wireless networks might comprise a fully end-to-end machine • Localized AI is where AI is applied to one network learning air-interface. Here, the challenge is to train an inter- domain. It requires data to be passed in the network, face that can support efficient data transmission and reduce en- however, is constrained to a single network domain, for ergy consumption while also fulfilling the latency requirements example, radio access network or core network. Localized of different applications. While an ML air-interface can be AI can also refer to scenarios where data is geographi- trained for optimizing data transmission, it can be challenging cally localized. to handle typical control channel problems such as energy • Global AI is where a centralized entity requires knowl- efficiency in scenarios where no data is transmitted or received. edge of the whole network and needs to collect data Moreover, an ML-based air interface system should be able to and knowledge from different network domains. Network adopt to the different requirements of each network slice in slice management and network service assurance are terms of data throughput, energy efficiency, and latency. As examples of global AI. such, initial AI deployments can focus on an ML air-interface Table I summarizes the benefits and challenges of au- targeting data transmissions improvement only. This approach tonomous node-level AI, localized AI, and global AI. Global is similar to the first NR non-standalone deployments where AI in future networks can be deployed either per slice whereby NR was introduced for eMBB services while being aided by each global AI entity manages a single slice independently of existing 4G infrastructure. In NR non-standalone, the network the other slices or per system whereby a single global AI man- uses an NR carrier mainly for data-rate improvements, while ages all the network slices simultaneously.