Intelligent Agents in Network Operations
Have you seen the movie Blade Runner? Otherwise, I highly recommend it. Blade Runner is a masterpiece for many reasons – but one of its greatest strengths is being technologically forward-thinking, especially in the direction of the evolution of artificial intelligence (AI). For this reason, I use it as a reference in this blog. Like in that sci-fi movie, I don’t want to immediately divulge who or what these intelligent agents are – what would be the fun in that? Instead, let’s first set the scene by putting things into perspective, exploring the requirements of new radio network operations and how these new requirements scale differently across communications service providers (CSPs), such as different strategies , use case requirements and model operations. A tier 1 CTO said to me a while back, “We can’t live without customizations.” Certainly, each CSP should be able to build their own competitive differentiators and unique solutions for their customers. The challenge is how to do it in the most efficient way possible.
Intelligent RAN Automation Creates Competitive Advantages
In the tech industry, the word “intelligent” usually refers to the application of AI techniques such as machine learning (ML) and deep reinforcement learning (deep RL). There are many so-called smart solutions, but this does not guarantee that they are optimal. The AI application guarantees optimal efficiency and performance only if it is executed where it makes sense. Two control loops must be distinguished: Real Time (micro and milliseconds) and non-Real Time (seconds, days, weeks).
These control loops are called distributed automation and centralized automation:
- Distributed automation requires a high volume of decisions in a very short time, and it is fully automatic. The feature is mostly self-contained and driven by designed algorithms.
- Centralized automation requires network-wide coordination, more time for complex decisions, and some human interaction. The use cases are mainly network design, optimization and management, which are not implemented in real time and would require a huge amount of manual work without automation.
AI/ML-based solutions based on self-trained algorithms are gaining in efficiency over rules-based and human-created solutions. AI algorithms, just like the Blade Runner movie replicants, are improving their performance and abilities in the latest versions. This is how intelligent RAN automation can create competitive advantages for CSPs.
The different alternatives for CSPs when using AI models
AI/ML technology introduces training, the concept of model drift, federated learning, and an increased need for access to large volumes of data. Life Cycle Management (LCM) processes define the roles of vendors, integrators, and the CSP. There are several models that can vary depending on the role of the service provider. Mainly, we are talking about four alternatives:
- The global trained model: the provider provides a model it was initially trained with global data.
- The local model: the provider provides the model and trains it with local data (data from the service provider’s network).
- The overall model initially trained with recycling capability: model retraining is performed by the service provider with local data.
- The embedded model with automatic relearning: the retraining of the model is done autonomously in the service provider’s network with local data.
“Who” are these intelligent agents, and what are their results?
These augmented humans are in the real world (not in science fiction) a set of AI models or algorithms.
The creator or provider of the algorithm provides a model which is initially trained to the service provider and then there is re-training with the actual network data. This allows continuous improvement of the models and their adaptation to an ever-changing environment. There are two scenarios, using simulators or emulators. Both strategies pursue similar goals, and the main differences lie in the configuration of the training environment, as you can see in the illustration below:
- Emulator Scenario: Typically duplicates all software (SW) and hardware (HW) functionality as a live network replica to provide more accurate results to the operations team.
- Simulator scenario: The simulator is a software program that models the behavior of a network. This is suitable for capturing general trade-offs and trends.
The application of AI techniques in networks has produced excellent results. For a European customer, Ericsson Power Optimizers reduced the overall power transmitted by 20%, with a saving of 3.4% on the electricity bill per base station. Enhanced enhanced antenna tilt with 5.5% downlink (DL) rate and 30% uplink (UL) rate.
AI and deep reinforcement learning techniques draw on human psychology to learn from the environment or data lakes. Network data is sent in real time in streams, creating a huge volume that needs to be stored and processed. To facilitate data management and operations, it is extremely important to select relevant field data for training. Therefore, it is mandatory to develop mechanisms to surface only the data needed for relevant use cases from specific network elements.
AI technology has sparked a lot of interest in sci-fi movies, as I mentioned at the beginning. But what is happening around us is no longer science fiction, it is already proven in the field with AI technologies in radio networks, we are creating the networks of the future! The world of tomorrow is now.
Do you want to dive deeper into intelligent agents? Read the guide: Intelligent Operation – How AI Plays a Critical Role in Network Operations
Learn about AI-based solutions in radio networks: visit our intelligent automation site: Intelligent RAN Automation to Manage 5G Complexity – Ericsson
Understand the main topics of intelligent RAN automation in our intelligent guide series: Intelligent RAN Automation Guide Series – Ericsson
Want to know more about performance optimizers? Find detailed information here: New Ericsson Performance Optimizers – Ericsson