How AI and intelligent automation are being leveraged to boost network performance
Artificial intelligence (AI) and machine learning (ML) will be key in helping CSPs automate their networks for improved performance, efficiency and customer experience. But a strategic approach will be needed for the AI journey, to ensure early efforts don’t go to waste. Our AI experts explain.
Imagine networks that can self-adjust their coverage and capacity where it’s needed, when it’s needed. Networks that can heal themselves when things go wrong, and optimize for superior performance and service delivery, all of the time – with no intervention from humans needed. It’s more than possible – in fact, with our artificial intelligence (AI) technologies and cognitive solutions such as Performance Optimizers, we’re already well on the way. The journey to zero-touch operations, powered by AI, is one we set out upon well over a decade ago. AI technology has evolved a lot since – and so has the telecom industry.
With the global rollout of 5G, emerging use cases such as the Internet of Things (IoT) and extended reality requiring low-latency and high volumes of data will demand ever more complex network infrastructures. At the same time, the pressure is on communication service providers (CSPs) to deliver seamless coverage and customer experiences – without increasing energy consumption or operating costs. It’s clear that network automation will have a fundamental role to play in their future success. But automating and optimizing all-encompassing mobile network infrastructure will require superhuman capabilities – making AI and machine learning (ML) more crucial than ever.
Where are we now?
In general, automation is the focus for operators right now, with AI being seen as a crucial tool. Most telecom network automation functions to improve performance still utilize traditional rule-based methods. But we are seeing legacy Radio Access Network (RAN) algorithms starting to be enhanced or replaced with AI counterparts designed to generalize across different radio conditions and environments, and intended to be reusable across the network.
There are three ways AI is being incorporated in network components:
- By replacing the existing rule-based algorithm or component with an AI-powered one.
- By adding a new AI-based component with new functionality to the existing base component.
- By adding AI-based control to the existing legacy component.
We’ve already seen some success in boosting network performance using these methods. Our focus has been on the RAN operations with the highest potential to boost network performance. AI algorithms should always be designed so they can be executed where it makes the most sense, and in a manner that ensures they can deliver real value.
For example, multiple independent and orchestrated AI algorithms can be designed to jointly learn from and govern the most relevant RAN operations within the protocol layer. Traditional RAN protocol hierarchy can still be used to ensure the correct separation based on the timescale of RAN operations. Timescales are particularly important, as they determine the timeframe in which the algorithm can finalize a decision - ranging from milliseconds to weeks.
Executed where it makes sense. Ensures optimal efficiency and performance
At one end of the timescale, we have fast control loops, like scheduling and link adaptation, where decisions can be made locally, autonomously and very quickly, without the need for any network overview. Thanks to increased intelligence, AI and ML can increase network performance and enable zero touch automation or optimization tasks in fast control loops, while also making them more adaptive and robust to deployment in different environments. This in turn minimizes the amount of effort that is needed in the slow control loops.
Slower control loops mostly relate to traditional network planning, design, positioning and also optimization. Rule-based solutions have been applied here for over a decade, for example, using centralized self-organizing network (SON) solutions. But even today, automation efforts are predominantly still manual and human-centric, and changes generally require a long time to implement. Given the volume of manual work involved in slow control loops, centralized RAN automation with AI and ML is especially attractive, as it can replace the manual work and enable faster and more efficient planning and design.
We also expect AI- and ML-powered solutions, in certain use cases or applications at least –to be more accurate and efficient than their rule-based counterparts, while simultaneously (when used in a data-driven approach) unlocking whole new opportunities and enabling us to do things that would not have been possible before
Understanding intelligence – who is calling the shots?
When talking about optimizing and boosting network performance with AI, it’s important to understand the difference between automation and intelligence. While all intelligent networks are automated, not all automated networks are intelligent. Ultimately, it comes down to who or what is controlling the decision making, and how.
There are four ‘states’ in the evolution from ‘standard’ software automation to intent-driven intelligence, which will often exist in parallel as the AI journey progresses.
Intent Driven AI
-Evolution of leveraging ML
1. “Standard” rule-based functionality
The methods traditionally used for network optimization rely on rule-based functionality, where highly-skilled experts manually create the rules by which automation frameworks operate. For example, they might create a rule stating that if a certain KPI reaches a given limit, then a certain action should be taken. This form of network automation is not intelligent. The rules are human-designed, typically static and generalized across networks – meaning modifications to benefit a specific case or functionality are extremely challenging and limited.
2. AI-enhanced functionality to improve performance
AI is introduced in a supporting role to inform the (human-made) decision-making process with predictions based on ML algorithms. This could involve adding a component to solve a problem that is done well by AI, such as pattern recognition or prediction, or by replacing a static parameter with a dynamic one. This primarily leverages supervised learning (SL) and/or unsupervised learning (UL), with minor adaptations to the rule-based function to accept the AI component.
For example, in link adaptation, the dynamic prediction of the Block Error Rate (BLER) target (a previously static parameter) can allow for better adaptation of the control loop to radio channel conditions. In energy efficiency, the prediction of traffic patterns can enable adjustments to the number of active antenna branches in a cell running multiple-input and multiple-output (MIMO). In both examples, the prediction does not impact the outcome. For instance, the traffic (active users) in the adjusted MIMO cell will not change due to the reduced number of available antenna branches.
3. AI-based functionality and selection
AI is introduced to undertake the entire selection process based on set parameters, including taking critical decision, based on the ML decision. This requires much higher demands in terms of robustness, assurance and recovery actions of the AI. This will, to a larger extent, require reinforcement learning (RL).
For example, for link adaptation, the link adaptation control loop (that in step two accepted the dynamic BLER predictions) might be replaced with an AI agent. The decision taken by this AI agent will impact the user equipment (UE) feedback – a parameter that will be used for future decisions. In energy efficiency, this could be done for decisions on active changes to the network, such as cell sleep, which in turn impacts the traffic in other cells. In both cases, the AI impacts the outcome and environment, so any feedback on the AI’s performance would need to take that into consideration.
4. Cognitive, intent-driven AI-based functionality
At this stage, the AI has the ability to look at multi-objective problems with (often) conflicting goals – such as network performance or energy efficiency, for example. The system is fully AI-controlled, needing only to be provided with the high-level intent for the operation, like a Service Level Agreement (SLA) which includes the business-related requirements such as cost or priorities.
An AI intent manager analyzes the intent and monitors the network to see if the intent is being fulfilled. With forecasting, it predicts when the intent is not going to be fulfilled in the near future, so that preemptive actions can be taken, orchestrating execution as needed to optimize the network based on the intent. Meanwhile, AI-powered automation agents throughout the system manage any conflict resolution that occurs. AI-native (where all components within the system are all designed to use AI, both within themselves and between one another and the system as a whole) will also be a key factor in achieving zero-touch operations.
What performance benefits have we seen with AI?
Spectral efficiency: One of the most interesting examples is an AI-powered feature to increase spectral efficiency. Spectrum is among the biggest asset expenses facing CSPs, and as such, increasing the efficiency of its utilization represents a very high value use case. Our initial demonstrations of using AI to adapt the modulation and coding scheme, have shown gains in spectral efficiency of roughly 10 percent, depending on the scenario. And while consumers themselves may not care much about their operator’s spectrum, they do care about their user experience – which will also be improved by more efficient utilization of the spectrum.
Downlink user throughput and transmission power: Reinforcement learning (RL), together with the use of digital twins to simulate and emulate network environments, has also been successfully validated for network performance boosting, this time in two live networks, as profiled in the Ericsson Mobility Report on AI enhancing customer experience.
In Malaga, Spain, MásMóvil managed to increase downlink user throughput by 12 percent in busy hours by using RL to optimize remote electrical tilt (RET), while keeping similar traffic volume and achieving a congestion rate close to zero. Swisscom also focused on optimizing RET, as well as lowering downlink transmission power, achieving a decrease of 20 percent on average while still achieving a 5.5 percent throughput gain. The output power decrease also resulted in a 3.4% reduction in base static power consumption.
Features like these, aligned to your business goals, can offer a great kicking-off point for boosting performance in your network with AI. But it’s important to remember that any AI automation or optimization exists within the context of your network ecosystem, and thus requires broader coordination and conflict resolution. If you only try to solve individual use cases or problems, or automate isolated functions with AI, you might see some improvements, but you certainly won’t get the full value from your efforts – and may even jeopardize the work of your project, or someone else’s. When planning your projects, make sure you’re taking an end-to-end approach with improvements that can act as stepping stones on the path toward a more intelligent, efficient and AI-native network.
Strategic insights for your AI journey
Many CSPs have, unsurprisingly, been swept up in the hype and excitement of AI technology and are eagerly jumping in to start implementing AI in various ways. And although now is definitely the time to act, we strongly recommend taking a strategic approach to your AI journey. Here are three key questions to consider, born from our own experiences, to help ensure you don’t head in the wrong direction.
1. Do you know the best place to start?
Think about the business outcomes you want to achieve, as well as the data needed and the end-to-end systems and processes involved in the area you want to automate. Start small and grow from there, while keeping the future in mind so you don’t run into any dead ends. For example, start with a project that suits your situation and needs, then treat that first use case or feature as a launchpad. It will take some initial investment, but once you have that first, well-designed feature established, the path forward becomes much clearer. Different elements can be reused in future projects, such as how you’re extracting your data, or the way you’ve set up automated feedback into your system. You gain more flexibility, as well as the end-to-end experience in how to train the model, retrain, update, build up new versions and so on. So having right systems e.g. MLOps and data strategy is necessary. It can be the vital first step to start building your AI capabilities.
2. Are you building the right AI foundation for a future-proof network?
As we mentioned before, building for AI-nativeness will be central for realizing zero-touch operations in the future. If you focus too much on just creating algorithms or models that only address a certain problem, without looking at the wider landscape the algorithm is operating in, you could build yourself into a dead-end. It’s important to consider the future role those components will have to play as part of the whole intelligent system architecture, or within the autonomous lifecycle of the network. Think ahead before you design, so you don’t end up having to replace your components again in a few years’ time.
3. Do you have the resources and know-how for success?
Many in the telecom industry don’t understand what AI can do in networks, how it can best be used, or even how it can be trusted. This isn’t helped by the lack of alignment or industry standards. Even with the support of experienced data scientists, it can be easy to fall into the trap of just focusing on a potential use case without considering the bigger picture. Along with AI knowledge, it is also crucial to have domain and system expertise. It can seem overwhelming, but you’re not alone.
Over the course of Ericsson’s long AI journey, we’ve already faced – and learned from – many of the challenges CSPs will come up against in the months and years ahead. In us you’ll find a trusted partner to help you move forward in your AI journey.
Benefits of AI in Networks blog series
We hope you enjoyed this episode on performance boosting with AI – you can read more of our ‘Benefits of AI in Networks’ blog series here, or sign up below to be notified of future posts.
Read more about our ‘Benefits of AI in Networks’ blog series.
Discover technical, industry and market insights in our intelligent RAN automation guide series.
Explore Ericsson’s AI-powered Performance Optimizers architecture.
Learn more about intent-based networks.
Like what you’re reading? Please sign up for email updates on your favorite topics.Subscribe now
At the Ericsson Blog, we provide insight to make complex ideas on technology, innovation and business simple.