How Will RIC Leverage AI/ML to Improve User Experience?

By :

AdobeStock 626051817 1240x600

RIC poses some unique new challenges in ensuring its success in a production network

Long before ChatGPT made artificial intelligence and machine learning (AI/ML) a topic of conversation at the family dinner table, the wireless industry had started using AI/ML to manage and optimize network performance and user experience. The growing complexity of networks in terms of topologies, types of devices (handsets, sensors, automobiles, industrial robots, etc.), applications, and the rapid introduction of new services has made it nearly impossible to effectively manage networks manually. Therefore, timely and optimal management necessitates machine intelligence assistance, which is why the O-RAN Alliance has identified the radio intelligent controller (RIC) as a network function to optimize RAN performance.

While radio access network (RAN) optimization has been done in previous generation networks, the software-defined RIC takes it to a whole new level. The open architecture of Open RAN (O-RAN), in general, and RIC, in particular, encourages innovation by enabling third parties to develop new and unique ways of optimizing the RAN and the user experience. Amazingly, this optimization can be down to the granularity of an individual subscriber. The RIC helps network operators reduce costs, optimize network performance and user experience, and generate new revenue opportunities.

RIC overview

Architecturally, the RIC is divided into the near-real-time (near-RT) RIC and non-real-time (non-RT) RIC. The non-RT RIC is collocated within the SMO (Service Management and Orchestration) framework and operates at a scale greater than one second. The near-RT RIC is typically deployed at the RAN edge and enables control even closer to real-time, at a scale of 10ms to one second.

O-RAN Alliance SMO RIC

The RIC targets specific use cases for optimization, such as energy savings, traffic steering, and QoE optimization. The RIC architecture facilitates the deployment of applications on the RIC platform to target these individual use cases. For instance, one application could target energy savings, while another targets QoE optimization. These deployed applications can be from multiple third parties, thus encouraging competition and innovation.

Non-RT RIC and near-RT RIC applications

The applications deployed on the non-RT RIC are referred to as rApps and those on the near-RT RIC, xApps. Typically, xApps and rApps use AI/ML models to achieve their objectives. These models ingest data from various network functions, such as the CU (Central Unit) or DU (Distributed Unit). The xApps/rApps use this data to track the state of the RAN and ongoing status of the subscribers. Based on this input, the applications may make recommendations towards dynamic allocation of scarce and shared resources, load balancing, and configuration updates optimizing the use case they are targeting. The RAN acts on these recommendations, thus realizing the optimization objectives of the xApps/rApps. This fully automated closed loop is the holy grail!

RIC testing in the lab

Testing the network functions in the lab before deploying them in a live network is standard industry practice to avoid network disruptions. This includes verifying specification compliance of the interfaces, performance, and capacity to ensure it can handle the live network load. All of this applies to the RIC as well. However, the RIC poses some unique new challenges in ensuring its success in a production network deployment. They include:

  • Training the AI/ML models

  • Verifying the AI/ML models

  • Keeping the AI/ML models current

Let’s look at each.

Training the AI/ML models

Every network is unique. The topology, subscriber behavior, traffic and data patterns, etc., vary from network to network. This is one of the challenges with effectively training the AI/ML models in xApps and rApps. While a certain base level of training can be achieved with generic training data, there must be mechanisms to fine-tune the training for specific networks. The richness of the training data in volume, variety, and authenticity is critical to effectively train these models. They need to be extensively trained on every possible type of scenario they are likely to encounter in a live network and in the context of the specific use cases that the xApp/rApp is designed for. Training data is gold when it comes to AI/ML.

Verifying the AI/ML models

Comprehensive lab verification of the models is a must. Once again, the richness of the verification data is the key. They must be verified against a full range of scenarios they will likely encounter in a live network. This is accomplished by recreating these scenarios in the lab with high fidelity. Similar to training the AI/ML models, verification must be in the context of specific use cases.

In addition, the way the output of the RIC is scored offers another challenge. Unlike traditional testing, a simple pass/fail label is inappropriate when it comes to judging the output of the RIC. Often, there isn’t one correct “answer” to optimize network performance in any given scenario. More nuanced ways of scoring are needed.

Keeping the AI/ML models current

Training the models is not a one-and-done type of situation. Modern networks are complex and rapidly evolving with ever-changing conditions and topologies. As new apps and games become popular and new types of devices and services are introduced, it leads to a shift in subscriber behavior and traffic patterns. This means the xApp/rApp AI/ML models can lose their effectiveness over time. Additionally, feedback from the performance of the xApps/rApps in a live network is needed to pinpoint areas of improvement in the AI/ML models.

How can RIC meet these challenges? Continuous education of the models is the answer. Eventually, a virtuous cycle of CI/CD/CT (where CT represents Continuous Testing and Training) needs to be established between lab and live networks to keep the AI/ML models current and constantly improving based on feedback and evolving conditions. Automation is essential to keep this cycle running smoothly and the models current and effective in a timely manner.

While the RIC promises to leverage the power of AI/ML to optimize the RAN, improve user experience, and reduce costs, it presents new challenges. It will be a rocky road getting to a point where it can do it reliably and predictably over time. These require new and innovative solutions to get the RIC and the xApps/rApps production ready. Spirent’s RIC test solution offers strategies to address these unique challenges, providing more test coverage in less time and at a lower cost, designed to seamlessly integrate into CI/CD/CT pipelines.




Nisar Sanadi

Senior Product Manager

Nisar Sanadi is a Senior Product Manager in Spirent’s Lifecycle Service Assurance business unit, where he is responsible for Landslide, the lab performance & compliance testing platform. He has extensive experience in the wireless test and measurement space, including product design & development, services, and product management. He has a passion for the intersection of AI/ML with telecom and his areas of expertise include 3G, LTE, 5G, and IMS. Nisar holds a Bachelor of Technology degree in electrical and electronics engineering from the Indian Institute of Technology, Bombay.