Why is Artificial Intelligence (AI) critical to operating and monetizing 5G networks? What kinds of AI and machine learning (ML) models can apply to 5G, and how are CSPs taking advantage of them? Let’s take a closer look.
Across every industry, businesses are adopting powerful new digital tools to streamline operations, lower costs, and fuel innovation. But invariably, these efforts carry a price: escalating complexity. Nowhere is this dynamic more apparent than in the evolution of communications service provider (CSP) organizations to 5G.
5G introduces groundbreaking digital innovations like Network Slicing, Open Radio Access Networks (Open RAN), ultra-low-latency services, and more. These technologies can help CSPs break away from yesterday’s “one-size-fits-all” connectivity model and start delivering customized services tailored to the needs of diverse enterprise and consumer applications. But there’s a problem: to unleash these next-generation business models; operators need to meet the exacting requirements of diverse high-performance and low-latency 5G services under stringent service-level agreements (SLAs). All this must be done in a network environment that is more complex than ever before.
5G innovations exponentially increase the number of dynamic real-time components in a CSP network. Operators relying on traditional manual approaches can’t observe everything in these environments, much less identify opportunities to optimize them. There’s too much information, too many decision points, and too many dependencies for the human brain to process. This problem is tailor-made for AI.
Understanding AI in 5G
AI will play a central role in the next evolution of 5G networks, helping CSPs enable greater customization and automation while driving down complexity and costs. For example, AT&T uses AI and ML to reduce fraud and robocalls, identify areas to improve its 5G network coverage, and much more. Vodafone uses ML to identify and fix network problems before they affect subscribers. Verizon uses ML to help predetermine the optimal design for new 5G cell sites. Looking ahead, AI can also assist in:
- Automating the coding of large numbers of parameters to deliver customized services (such as via private networks) to diverse enterprise customers
- Managing the allocation, distribution, and load-balancing of resources across 5G radio networks in response to changing conditions
- Managing real-time traffic flows end-to-end across network slices and services to meet bandwidth, latency, and reliability requirements under service-level agreements (SLAs)
Wherever operators need to manage vast volumes of network traffic across non-linear mobile and wireless communications networks or perform real-time polynomial calculations to optimize decision-making, AI can play a role.
Today, CSPs are applying a variety of AI and ML techniques to solve real-world operational challenges, including:
- Supervised learning, which trains an algorithm to categorize input data towards a predetermined output accurately
- Unsupervised learning, which trains an algorithm to identify previously undiscovered patterns and relationships in a data set
- Reinforcement learning, which trains an ML agent to achieve the desired outcome by providing feedback in the form of rewards and punishment
Supervised Learning
Supervised learning accurately trains an algorithm to predict outcomes based on input data. The algorithm analyzes large, labeled data sets to learn which outputs are associated with different input values. Over time, it learns to recognize underlying relationships between inputs and outputs to apply to unlabeled data. This method includes two variants:
- Classification, where the algorithm is trained to categorize input data into specific labels or classes accurately
- Regression, which aims to predict numerical relationships between input and output data without associating the data with any particular class
CSPs use supervised learning to enable self-organized networks (SON) that automatically optimize 5G networks (minimizing bandwidth utilization) without requiring human intervention. They also apply supervised learning techniques like linear regressions, decision trees, and support vector machines (SVMs) to optimize dynamic frequency and bandwidth allocation (DFBA). Here, the AI analyzes and adjusts network frequency and bandwidth parameters to improve performance.
Unsupervised Learning
Where supervised learning trains algorithms to identify patterns to achieve a specific desired outcome, unsupervised learning lets ML agents train without a roadmap. Unsupervised learning uses data sets that have not yet been labeled. The algorithm analyzes the unlabeled data sets to discover novel patterns and underlying relationships. The trainer specifies the number of groups the model should create without knowing the nature of the relationships the AI will identify ahead of time.
In 5G networks, CSPs use unsupervised learning to analyze traffic patterns to optimize network resources. For example, the K-Means Clustering algorithm can help operators determine the best way to distribute relay devices between base stations. This approach proves particularly valuable for optimizing connectivity to highly mobile devices such as smartphones and autonomous vehicles.
Figure 1 depicts supervised versus unsupervised learning, illustrating how these methods train algorithms to categorize or cluster training data.
Reinforcement Learning
Reinforcement learning (RL) involves training an AI to interact with its environment and learn through trial and error. Like supervised learning, RL applies mapping inputs to outputs. But here, the agent is trained via rewards and punishments as it interacts with the data. By constantly seeking to maximize rewards, the agent ultimately identifies the optimal approach to a task.
The RL model consists of three elements:
- A policy function that decides the agent’s behavior
- A value function that interprets the agent’s actions that contributed to achieving the objective
- A model of the environment that contains the agent’s already-learned knowledge
Figure 2 depicts how RL can be used in a telco network. Here, an agent interacts with a network environment and is rewarded based on its ability to optimally control traffic (reducing packet loss, reducing latency, etc.)
In the future, CSPs will likely perform random testing with RL in 5G networks to reduce latency and packet loss. Over time, this model could help operators continually optimize traffic flows and reduce congestion in their networks. As of today, however, this model is still unproven; we will need to see what CSPs can achieve with RL as 5G networks mature.
Looking Ahead
5G gives CSPs the tools to deliver a wide variety of more customized (and lucrative) enterprise and consumer services. To capitalize, however, operators must find ways to operate these complex network environments efficiently and achieve the stringent performance and latency requirements of next-generation 5G services and slices.
AI will play a vital role in this effort. By enabling accurate predictions, optimal resource distribution, faster anomaly detection, and more, AI can help CSPs unleash the full power of 5G networks to transform their businesses.
