Call for change: a wider scope of automation
Self-Optimizing Network (SON) is not a sufficient answer to the new demands:
Increased complexity, e.g. mix of virtualization and physical workloads, increased network density, and advanced antenna systems
New business opportunity, e.g. network slices, multi-tenancy, dedicated networks
Increased demands for multi-vendor support and openness
The Ericsson Intelligent RAN Automation portfolio, shown in Figure 1, features end-to-end network automation that includes centralized and distributed SON solutions and new capabilities that support the transformation to a more open environment enabled for AI/ML, which empowers innovation and support for wide range of use cases, shorter time to market and is highly adaptable supporting existing and future networks.
Figure 1 Transformation of current offerings into a more open environment for innovation
The objective of RAN automation is to boost RAN performance and operational efficiency by replacing the manual work of developing, installing, deploying, managing, optimizing and retiring of RAN functions with automated processes. The AI’s role is to unlock more advanced network automation performance to make RAN network functions more autonomous and replace manual processes with intelligent tools that augment humans. Furthermore, it makes both AI/ML powered RAN network functions and tools more robust for deployment in different environments.
Creating competitive advantage with RAN automation
Ericsson AI and automation foundations gives service providers the platforms, and evolved life cycle management of RAN SW and services to evolve networks efficiently to successfully meet ever-changing demands. The aim to deliver improved network performance, accelerate time to market for new capabilities, target right investment for improved ROI and enhanced operational efficiency.
Ericsson Intelligent RAN Automation solutions provides the right automation where it makes sense, gives most “bang for the buck”. The Figure 2 illustrates how the task of efficiently operating a RAN to best utilize the deployed resources can be divided into different control loops acting all together according to different time scales and with different scopes. Intelligent RAN Automation solutions utilize AI/ML algorithms interacting and integrating with engineered algorithms and existing processes, where applicable, in all these control loops.
Figure 2 Holistic take on RAN Automation
The two fastest control loops are related to traditional Radio Resource Management (RRM). Examples include link adaptation in the fastest control loop and cell supervision in the second fastest) control loop. Functionality in these control loops is mostly autonomous, although mostly driven by engineered algorithms requiring complex configurations in a timeframe ranging from milliseconds (ms) to several hundred ms. In many cases AI/ML makes it possible to enhance the functionality in the fast control loops to make them more adaptive and robust for deployment in different environments. This, in turn, minimizes the amount of configuration optimization that is needed in the slow control loops.
The slower control loops shown on in Figure 2 are related to traditional Network design, optimization and management. Examples include RAN coordination and network power management. In contrast to the two fast control loops these slower loops are today to a large degree manual. The slow control loops encompass the bulk of the manual work that will disappear as a result of RAN automation, which explains why AI/ML is especially attractive in these loops.
In the near term, we expect the AI/ML powered solutions to be more accurate and efficient for certain use-cases or applications than the rule-based powered solutions. In the long-term AI/ML powered solutions will be ubiquitous in the RAN and AI/ML just another tool to achieve best performing and cost-effective network.
Evolved life-cycle management
Compared to traditional software, AI/ML technology introduces the elements of training, model concept drift, federated learning, and a stronger need for access to data. The life cycle management (LCM) processes set the roles of suppliers, integrators and CSPs in essence, who is responsible for what and who sells what to whom. As an industry, we must adjust the LCM of software to include AI/ML-LCM and technology to reach its potential as it evolves, maintaining a clear separation of concern and, with a minimum of variants, to avoid industry fragmentation. A very high-level AI/ML LCM process is captured in the figure below.
Figure 3 A high-level life cycle management (LCM)
We recognize four main LCM alternatives as shown in the Figure 3:
Trained global model, training is responsibility of the vendor in train step. Example: Power Optimization as application (future rAPP)
Local model, local training is responsibility of the vendor in train step, developer might deliver initially trained model with vendor data.
Initially trained global model with retraining capability, training or re-training is responsibility of CSP in adapt/accept step
Embedded model with automatic retraining, re-trained on local data autonomously in CSP network. Example: AI powered advanced cell supervision (Centralized Training System enabled)
AI/ML lends itself much better to choosing level of global vs local adaptation in comparison with traditional rule-based solutions. Globalization of AI/ML model can be described, as a model that is trained once, e.g. in vendor environment and deployed in many networks and situations or it has ability to adapt correctly to new previously unseen data, etc. There is also a need to do local adaptations and train or re-train the AI models with unique local data.
Ericsson Data driven development
Data collection is probably the biggest challenge to scaling the AI/ML. Public data (e.g. performance monitoring) is standardized exposed data available from a product or service supplied by a vendor to CSPs for the purpose of product operations and/or service delivery. Non-public data (e.g. AI model debug trace), on the other hand, is data containing sensitive information relating to Intelligent Property Rights (IPR) and is used by the vendor for innovation, and/or service development, verification and deployment. Generated non-public data is typically hundred thousand times larger in volume than the public data. Ericsson has therefore developed mechanisms to bring out just the data that is needed for the relevant use case from specific network elements.
A simulated environment is often used as our first development step with AI/ML-based algorithms, regardless of whether we use public and non-public field data or simulated data to train the final model.
The AI/ML algorithm may be improved over time, or complemented with other algorithms, to make the predictions more accurate, or by re-training the model with local data in the network where it operates. In a longer perspective, this iterative development may result in centralization of certain AI/ML resources as the system architecture and capabilities evolve. Data-driven development is important components in evolution of life cycle management of RAN SW.
We make a distinction between initial training of AI/ML algorithm, here defined as creating and training an ML algorithm in design phase, or training and re-training in maintenance phase. Once the AI/ML features are identified for the initial training of the AI/ML model, we know what data is needed for re-training when the model starts to drift, which might impact efficiency of network function or a process. The re-training can either be done off-line in data-driven development at Ericsson or within the operator’s network. In the latter case, the re-training is done with customer-unique data and often with the purpose of adapting to local environment that are hard to generalize with the data available off-line.
Standardization as enabler for multi-vendor ecosystem
The industry has recognized that in order to transition to an industrialization phase and enable mass adoption of AI/ML, industry alignment is required. This results in all the major industry bodies trying to work out how they can leverage the technologies and claim their stake in the AI/ML landscape, leading to multiple and somewhat diverging directions being taken. To accelerate the coming industrialization phase and mass adoption, the industry must do more to align standards between 3GPP, ORAN, ONAP and ETSI by:
Focusing on data collection using Service Based Intents and data management using standardized interfaces of both public and non-public data.
Focus on aligning functional architecture for AI/ML functions and LCM
Focus on and align different network analytics specifications while enabling optimization for different domains.
Applying AI/ML in use cases
Objective of AI is to unlock more advanced network performance and automation and ultimately it is about delivering the value to customers. Data sciences are combined with telecom knowledge to create use-case driven and business driven approach to implementing AI where it makes most sense.
Table 1 Few of the latest use cases being industrialized
AI powered DL link adaptation
AI powered advanced cell supervision
AI powered inter-DU coordination
AI powered DL link coordination
AI powered link adaptation is network optimization solution targeting improved spectrum efficiency. The feature introduces a neural network Ericsson compute to enhance link properties giving an improved spectral efficiency and increased throughput. Current link adaptation is optimized for high loaded systems. By utilizing information from adjacent cells, we enhance link adaptation for medium loaded systems with significant improvements in the spectrum efficiency.
AI powered advanced cell supervision
AI powered advance cell supervision is network healing solution. Locally executed, self-learning, on RAN Compute, and self-retraining on the Centralized Training System, Machine Learning algorithms allows for continuous full network supervision that continuously looks for anomaly in cell’s performance. Capable of instantaneous and predictive detection and immediate recovery actions with minimal impact.
This offers Instantaneous detection and recovery of cells with degraded KPI, resulting in improved In-Service Performance (ISP). At model drift RBS triggers AI model re-training and deployment of new model by Centralized Training System in the cloud.
AI powered inter-DU coordination
AI powered inter-DU coordination is network deployment solution where NR carrier aggregation (CA) between e.g., low-band/ high-band, provides enhanced peak rate and coverage extension. Selecting and configuring the most optimal DU partners, on a network-wide basis, can be challenging and time-consuming. Advanced RAN Coordination optimizes and automates this task, via a central application for optimal partner selection over the entire network. Machine Learning algorithms are used to predict the cell load to secure an optimal selection removing the need or manual selection and configuration.
Downlink Power Optimization is network optimization solution that uses Deep Reinforcement Learning technology to identify if cell TX power can be reduced without compromising coverage or performance. Equally the solution identifies cells where power increase is required for performance improvement. Power optimization saves energy and allows maximizing radio capacity in markets with strict RF emissions regulations. Continuous closed-loop optimization automatically maintains the optimum settings as the network evolves and traffic distributions change. Resulting in DL power reduction on coverage layer while maintaining traffic volume and improving DL and UL performance.
Ericsson is well on the way to innovate and significantly change the way the automation of RAN is done. Leveraging thought leadership in the intersection between data-driven AI/ML principles and RAN automation expertise, AI/ML is applied in the Ericsson Intelligent RAN Automation solution. The Ericsson Intelligent RAN Automation portfolio, which is the next step in SON, features end-to-end network automation and new capabilities that support the transformation to a more open environment enabled for AI/ML, empowering innovation, support for wide range of use cases, shorter time to market and is highly adaptable supporting existing and future networks. This innovative solution transforms RAN SW life cycle enabling AI operations and provides AI functions where it makes most sense. Ericsson AI and automation foundations gives our customers the platforms, and evolved life cycle management of RAN SW and services to evolve networks efficiently to successfully meet ever-changing demands. Applying AI in RAN enables to industrialize a wide range of use cases working across various control loop time frames. The use cases will enable our customers to create business value in terms of improved performance, higher efficiency, enhanced customer experience and ultimately create new revenue streams.