Leveraging AI and machine learning in RAN automation – Ericsson

The left side of Figure 3 illustrates how the task of efficiently operating a RAN to best utilize the deployed resources (base stations or frequencies) can be divided into different control loops acting according to different time scales and with different scopes. A successful RAN automation solution will require the use of AI/ML technologies [6] in all of these control loops to ensure functionality that can work autonomously in different deployments and environments in an optimal way.

The two fastest control loops (purple and orange) are related to traditional RRM. Examples include scheduling and link adaptation in the purple (layer 1 and 2) control loop and bearer management and handover in the orange (layer 3) control loop. Functionality in these control loops has already been autonomous for quite some time, with the decision-making based on internal data for scheduling and handover in a timeframe ranging from milliseconds (ms) to several hundred ms, for example. From an architecture perspective, these control loops are implemented in the RAN network function domain shown in Figure 3.

The slower control loops shown on the left side of Figure 3 represent network design (dark green) and network optimization and assurance (light green). In contrast to the two fast control loops, these slower loops are to a large degree manual at present. Network design covers activities related to the design and deployment of the full RAN, while network automation covers observation and optimization of the deployed functionality. Network optimization and assurance is done by observing the performance of a certain functionality and changing the exposed configuration parameters to alter the behavior of the deployed functionality, so that it assures the intents in the specific environment where it has been deployed. From an architecture perspective, these control loops are implemented in the RAN automation application domain [7].

The green control loops encompass the bulk of the manual work that will disappear as a result of RAN automation, which explains why AI/ML is already being implemented in those loops [8]. It would, however, be a mistake to restrict the RAN automation solution to just the green control loops. AI/ML also makes it possible to enhance the functionality in the purple and orange 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 light-green control loop.

While the control loops in Figure 3 are all internal to the RAN domain, some of the functionality in a robust RAN automation solution will depend on resources from other domains. That functionality would be implemented as part of the RAN automation application domain. The RAN automation platform domain will provide the services required for cross-domain interaction.

One example of RAN automation functionality in the RAN automation application domain is the automated deployment and configuration of ERAN. In ERAN deployments, AI/ML is used to cluster basebands that share radio coverage and therefore should be configured to coordinate functionality such as scheduling [8]. To do this, data from several network functions needs to be clustered to understand which of them share radio coverage. This process requires topology and inventory information that will be made available to the rApps through the services exposed by the network automation platform over R1.

The outcome of the clustering results is a configuration of the basebands that should coordinate as well as a request for resources from the transport domain. This information can also be obtained by services provided by transport automation applications exposing services through the R1 framework. When designing the rApp for clustering, it is beneficial to have detailed knowledge about the implementation of coordination functionality in the RAN network function to understand how the clustering analysis in the rApp should be performed.

An example of RAN automation functionality in the network function domain is AI/ML-based link adaptation, where AI/ML-based functionality optimizes the selection of the modulation and coding scheme for either maximum throughput or minimum delay, removing the block error rate target parameter and thereby the need for configuration-based optimization. Another example is secondary carrier prediction [8], where AI/ML is used to learn coverage relations between different carriers for a certain deployment. Both of these examples use data that is internal to the network function.

As the objective of RAN automation is to replace the manual work of developing, installing, deploying, managing, optimizing and retiring RAN functions, it is certain to have a significant impact on the way that the LCM of RAN software works. Specifically, as AI/ML has proven to be an efficient tool to develop functionality for RAN automation, different options for training and inference of ML models will drive corresponding options for the LCM of software with AI/ML-based functionality.

Figure 4 presents a process view of the LCM of RAN components, ranging from the initial idea for a RAN component to its eventual retirement. A RAN component is defined as either a pure software entity or a hardware/software (physical network function) entity. As the different steps in the LCM structure include the manual work associated with RAN operations, it is a useful model to describe how RAN automation changes the processes, reduces the manual effort and improves the quality and performance of the RAN.

See the original post:

Leveraging AI and machine learning in RAN automation - Ericsson