How to connect the dots of future network AI
Artificial intelligence will be a game changer for next-generation wireless networks. But are there any AI integration pitfalls which operators should be mindful of? And what exactly is needed to realize a fully-deployable AI in 5G networks and beyond? Read below to find out.
The different types of AI in the network
Artificial intelligence (AI) will inevitably have a significant role in reshaping next-generation wireless cellular networks - from intelligent service deployment to intelligent policy control, intelligent resource management, intelligent monitoring, and intelligent prediction.
The journey to tomorrow’s autonomous networks will begin with an integration of such intelligent functions across the wireless infrastructure, cloud, and end-user devices with the lower-layer learning agents targeting local optimization functions, while higher-level cognitive agents pursue global objectives and system-wide awareness. In this regard, it is important to differentiate between autonomous node-level AI, localized AI, and global AI.
- Autonomous node-level AI is used to solve self-contained problems at individual network components or devices, where no data is required to pass through the network.
- Localized AI is where AI is applied to one network domain. Localized AI requires data to be passed in the network, however, is constrained to a single network domain, for example, radio access network or core network.
- Global AI is where a centralized entity requires knowledge of the whole network and needs to collect data and knowledge from different network domains. Some examples of global AI are network slice management and network service assurance.
In the figure below, you can see just how intelligence is distributed across different domains and network devices in the network.
Figure 1: Distribution of intelligence in the network: global AI, localized AI, and autonomous node-level AI with autonomous node-level AI (i) at the cloud or (ii) at the edge.
The challenges to AI integration in the network
To reap the benefits of integrating AI in wireless networks, AI tools must be tailored to the unique features and needs of the wireless networks. These are evidently quite different from the traditional applications of AI, such as in robotics and computer vision.
Below, I have listed some of the main areas which I believe must be further investigated to realize the synergistic integration of AI in future wireless networks.
- Data: Acquiring and labelling data is fundamental. The process needs to consider the privacy of some features, the requirement of measurements across large-scaled infrastructure, and the need for domain knowledge expertise.
- Security: The success of integrating AI in next-generation wireless networks will not only depend on the capability of the technology but also on the security provided to the data and AI models. It is crucial to guarantee obtaining accurate datasets and AI models by avoiding data from faulty base stations (i.e. false base stations) or compromised network devices. Moreover, secure schemes are necessary for sharing data and network intelligence across different network devices and domains.
- Confidential computing: A confidential computing multiparty data analytics with secure enclaves is an interesting technology with the potential for security and privacy improvements for AI applications. Confidential computing can increase trust in AI applications to the wireless network domain by ensuring that operators can be confident and that their confidential customer and proprietary data is not visible to other operators.
- Learning at the edge: Centralized AI schemes can be challenging for some wireless communication applications due to the privacy of some features and limited bandwidth and energy for transmitting all the data to a cloud server. Meanwhile, distributed machine learning techniques, such as federated learning, have the potential to provide enhanced user privacy, if correct measures are taken. Such schemes enable network devices to learn global data patterns from multiple devices without having access to the whole data. This is realized by learning local models based on local data, sending the local models to a centralized cloud, averaging them and sending back the average model to all devices, as illustrated in Figure 2 (below). Nevertheless, the effectiveness of such schemes in real networks should be further studied considering the computational complexity and energy consumption of the mobile devices. Moreover, it is vital to design a common distributed and decentralized paradigm to make the best use of local and global- data and models.
- Reinforcement learning in cellular networks: Reinforcement learning algorithms try to maximize the expected future reward by exploiting already existing knowledge and exploring the space of actions in different network scenarios. However, exploration in real environments might cause performance degradation, so new approaches such as pre-training, transfer learning, shared learning, semi-supervised reinforcement learning, and the use of simulation-in-the-loop techniques are being investigated. One can also identify network conditions for the underlying use case under which exploration can still guarantee the promised quality-of-service to the connected devices. Partial observability and multi-agent reinforcement learning are also being investigated as these aspects are key components for enabling the application of reinforcement learning techniques in cellular networks.
How can we align the behavior of AI networks to human goals and intents?
- AI alignment: It is often complicated to design a reward-function for some wireless applications. For such scenarios, it would be interesting to enable the AI agent to interact with the user, thus taking into consideration the user’s goals and intentions during the learning phase. This is essentially known as the AI alignment problem which can be defined as “how to align the behavior of AI networks to human goals and intents?” and is indispensable for applications where a built-in reward function is not available.
- Device-based AI model: Generally, AI models are trained on data collected from different types of mobile devices. Still, it is possible to get different behaviors when the models are run on different devices. Therefore, it is essential to develop downloadable AI device-based models as opposed to having one unique downloadable AI model for all devices.
- Co-existence of AI-based schemes with conventional algorithms: Network devices can implement both reactive (non-AI-based) and proactive (AI-based), approaches for the different types of applications. Data driven algorithms should only replace or complement traditional design algorithms if there is an overall performance gain.
- Co-existence of different network operators with AI functionalities: Future wireless networks must ensure the co-existence of various AI legacy network devices belonging to different network operators. In this regard, distributed learning schemes are of interest since network operators typically do not want to expose radio access network information.
- Co-existence of AI and non-AI legacy network devicesis essential for enabling a smooth transition from a network of connected things to a network of connected intelligence. For instance, misleading behavior (intentional and non-intentional) of non-AI devices could impact the learning process of AI devices during online learning. This can be the consequence of human-based actions for non-AI devices which can potentially result in unpredictable behavior and therefore constrain closed-loop automation. Moreover, it is crucial to assess the performance of AI/non-AI devices when connected to non-AI/AI base stations.
In a nutshell, it is clear that there is a need for a scalable and deployable AI for 5G networks and beyond. It is important to consider the whole communication system including simple devices, smartphones, base stations, routers, core network, etc. as part of the AI-network to achieve closed-loop automation. Moreover, it is crucial to efficiently use AI i.e., to reuse data and AI models, transfer network intelligence across different network domains, and investigate efficient training schemes, in order to reap the benefits of integrating AI in next-generation wireless networks.
Got a few more minutes?
Visit our AI and automation page to find out how we have helped our customers to integrate AI across their networks, or read our latest AI report to explore the latest industry insights, challenges and expectations regarding AI adoption.
Visit our future autonomous networks page to find out more about our work at Ericsson Research.