AI/ML-Driven Network Automation and RIC

Many network operators have launched their commercial 5G services worldwide, offering much higher energy efficiency than 4G networks. However, the higher energy efficiency of deployed 5G networks does not mean overall lower energy consumption costs. The development of RAN Intelligent Controller (RIC) plus corresponding application use cases creates a great opportunity and is well-positioned to solve the problem. Powered by network virtualization, cloud-native architectures, and open Application Programming Interfaces (APIs), the RIC framework allows network operators, partnered with various software and hardware vendors, to develop innovative application use cases for dynamic radio network control in cost-effective and intelligent ways.

As defined by the O-RAN Alliance, RIC mainly consists of a non-Real-Time (non-RT) RIC (supporting greater than 1s task latency) and a near-RT RIC (supports less than 1s task latency). From a business perspective, the market focus is on two aspects: RIC platform development and the corresponding application (xAPP/rAPP) case studies. To test the feasibility of the solution, many players have already conducted live network trials and tests. For example, Nokia, partnered with AT&T, conducted RIC trials over the latter’s 5G mmWave network to test RAN E2 interface and certain xAPPs, such as measurement campaign and admission control. Samsung collaborated with KDDI to demonstrate the 5G end-to-end network slicing capability with RIC in Tokyo, Japan. Nokia, together with CMCC, also trialed AI-powered RIC on CMCC’s live 4G and 5G networks to forecast users’ traffic bandwidth and detect network anomalies in Shanghai and Taiyuan, China, respectively.

The programmability of a network allows operators to manage their end-to-end system configuration, monitor, and upgrade automatically without much human intervention. However, such an automation strategy requires skilled programmers to handle the initial installation and debugging process. Once the system is installed and programmed, it also needs appropriately trained maintenance personnel to monitor network productivity and intervene when assembly line failure is predicted. All these actions could result in sky-high deployment and management costs. The alternative is to automate these processes, where analytical and data-driven approaches can allow network operators to take a more proactive role to orchestrate radio resources and improve user experience.

When considering data-intensive network management, it is envisioned that the future network system should enable much intensive interaction between network infrastructures and end users. Moreover, network automation will most likely shift the focus from passive/static management to proactive/dynamic resource allocation for granular service assurance. In this case, AI/ML technology supported by RIC platform and corresponding xAPP/rAPP plays a pivotal role in providing computer systems with the ability to learn network conditions and enforce proper actions without being explicitly programmed. Figure 1 gives a scope of three types of AI/ML-driven RIC deployment models.

As shown in the figure, depending on the timescale of specific application use cases, an AI/ML training model can be hosted at non-RT-RIC or near-RT-RIC, and the model learning can be offline or online. For example, the model on the left-hand side of the figure is generally trained offline, and the main application use cases are for non-real-time, such as system-level network slicing and QoS assurance. The model in the middle of the figure is also trained offline, and both non-RT-RIC and near-RT-RIC jointly work to enable applications like traffic steering and QoE assurance. The online training model is shown on the right-hand side of the figure. In this case, the near-RT-RIC platform should host the AI/ML model and keep it updating periodically for applications implemented in dynamically changing environments, e.g., AI/ML-assisted beam tracking and channel estimation.

Many use cases for RIC are not new and 3GPP has already specified Self-Organizing Network (SON) functions in the official 4G standard. This type of automation solution, including QoS-based traffic steering and load balancing, was designed to make network installation, configuration, and optimization fast and straightforward. However, due to network architecture limitations and complexity concerns, such a solution mainly resides in a core network without the capability for time-sensitive radio control with granular user-level service assurance. 3GPP in its 5G standard decouples user and control plane functions and allows them to be dynamically deployed in either core network, edge network, or even at a cell site. This action provides network operators with a great opportunity to embrace more flexible network deployment strategies. Combined with the modular-based AI/ML-driven RIC framework, the new network architecture can support diversified network services with agility and flexibility.   

The mass deployment of RIC platform and corresponding applications, especially with AI/ML technology support, is arguably a long-term initiative. According to ABI Research’s market study, the main barriers are summarized in three folds:

• Unlike open fronthaul interface standardization, the development of a standardized RIC platform and the corresponding APIs is far more complex. Apart from handling multi-vendor interoperable complexity, developers also need to innovate platform-agnostic applications for large-scale deployment and complexity management. This requires the vendor ecosystem, together with standardization bodies, to clarify different implementation environments and take more proactive roles for intensive tests and trials.
• The implementation of AI/ML-based applications requires intensive computing resource support. Such support is even critical to deal with the dynamically changing environment. In this case, network operators should evaluate their existing computing assets and may need to collaborate with cloud computing service providers to extend their global footprints. In this case, a comprehensive and standard security framework is essential to help alleviate the perception of network deployment risks.
• Many incumbent vendors are developing their own intelligent RAN control platform, which can support third-party applications via open APIs but with limited access. Since a platform is a foundation for supporting various application use cases, it is essential to develop a unified standard platform framework to avoid a new type of vendor lock-in.

The development of a standard platform, potential applications, and open interface specifications is still ongoing. Although the timeline for large-scale deployment is still unclear, key stakeholders are proactively conducting implementations and trials for RIC and trying to pick up the speed of 5G deployment.