The Challenges of Artificial Intelligence in the RAN

By Iain Gillott, Senior Research and Technical Advisor, WIA

The first two blogs in this series described what artificial intelligence applied to the Radio Access Network (RAN AI) is, the areas impacted and the potential benefits provided. The summary from these two pieces was basically that as mobile networks have become more complex, the RAN has become both more important and harder to deploy, tune and manage. AI helps address these challenges and ultimately enables improved spectral efficiency, improved user experience and increased operational flexibility for the mobile network operators (MNOs).

But as with any new technology, there are issues and challenges with deploying AI in the RAN.  As we will discuss in this report, these issues are not insurmountable but they do require examination and careful consideration.  Deploying AI across diverse RAN environments — crossing traditional, virtualized, Open RAN, and private networks — presents technical, operational, regulatory and strategic challenges. Understanding these limitations is essential for MNOs, vendors and enterprises to set realistic expectations and plan for effective implementation.

The challenges and issues can be broken down into the following areas, each of which we will expand on below:

• Data collection
• Model applicability
• Compute constraints
• Operational trust
• Security and privacy
• Business considerations.

Data Collection

AI models depend heavily on large volumes of high-quality data that has been labelled and sourced; in other words, the origin is known and the data can be identified. For the mobile RAN, data that can be used in an AI model includes:

• Network KPIs (latency, throughput, handover success);
• Radio measurements and data: Signal to Interference plus Noise Ratio (SINR), Reference Signal Received Power (RSRP) and Channel Quality Indicator (CQI);
• Hardware and environmental telemetry.

The challenges with data collection include:

• Data fragmentation across multiple vendor platforms and network domains. Since data comes from multiple places within and out of the RAN, the data can be fragmented and disjointed, leading to an incomplete picture of what is happening in the RAN.
• Incomplete or noisy data, especially in legacy infrastructure. Simply, if required data is not available or is corrupted in some way, the model will be inaccurate.
• Data labeling and contextualization, which is time-consuming and often requires domain expertise. If the data is incorrectly labelled, someone is going to have to correct the issues.

The old saying “junk in, junk out” applies here for AI models. If the data used to build and train the model is incomplete or inaccurate, the resulting model will be flawed.

Model Applicability

AI models that are trained in one network environment, such as an urban macrocell, may not be effective in other areas, such as indoor small cells or rural areas. This can limit the scalability and effectiveness of pre-trained models and creates the need for continuous model retraining, environment-specific models and learning techniques that are applied across the entire network.

In Open RAN (ORAN) networks, there can be added problems arising from multi-vendor components that may have different performance characteristics.  Industry initiatives, such as the O-RAN Alliance and ETSI Zero Touch Network & Service Management (ZSM), aim to standardize interfaces and frameworks for AI-driven RAN control. But widespread interoperability is still a work in progress.

Careful planning of the AI model build and training is required to ensure that all environments, vendor characteristics and variations in the entire network are considered. Otherwise, the result could be that multiple AI models are needed to manage the complete network, adding cost and complexity.

Compute Constraints

Certain mobile RAN functions, such as beam selection, scheduling and interference management, are latency-sensitive, which requires decisions in milliseconds. Running AI inference in real-time therefore needs:

• Low-latency edge computing infrastructure.
• Lightweight AI models that can operate with limited Central Processing Unit and Graphics Processing Unit (CPU/GPU) resources.
• Robust fallback mechanisms so that if the AI model fails or underperforms, the RAN can still function and provide service.

Ensuring the AI model has sufficient computing power available with suitable execution speed is a key design consideration, especially in edge-based deployments.

Operational Trust

In highly regulated or mission-critical networks, it is essential that AI decisions are:

• Transparent (the network operators understand the rationale for the decisions being made);
• Explainable (the AI model’s internal logic can be interpreted and explained); and
• Auditable (AI decisions can be traced and justified).

Explainable AI frameworks are still evolving, and black-box models (those where AI decisions are output but the mechanisms used to come to the decisions are opaque to the user) may be unacceptable for certain use cases such as public-safety networks or critical infrastructure.

Security and Privacy

AI can expand the attack vulnerabilities in the network by introducing:

• New software components;
• Real-time data pipelines that may expose sensitive information; and
• AI-specific threats, such as data poisoning or adversarial model attacks.

Mitigating these risks requires careful planning and execution and initiatives such as secure model training and inference environments and privacy-preserving analytics. Understanding the security and privacy risks introduced by AI at the outset is an important step toward building an AI-enhanced network.

Business Considerations

As with any industry, the overall impact of AI on the RAN and the MNO’s operations should not be underestimated.  When designing and implementing RAN-AI models, MNOs should consider:

• Return on Investment and cost justification, particularly for large-scale AI infrastructure investments. In a large, nationwide RAN with tens of thousands of cell sites, the cost of AI can be considerable.
• Change management and workforce retraining as operational processes shift toward automation. Simply, AI changes how the work is done and the workforce must be able to take advantage of the changes in order to increase efficiency and productivity.
• Governance models, defining who has oversight over AI behavior, especially in networks with shared or outsourced infrastructure.

It is important to understand that these challenges and issues are not insurmountable and that many organizations, MNOs, vendors and industry groups are working on solutions.  Careful planning is required to deploy an AI-RAN – after all, the RAN is critical to the revenue stream for the MNOs and the entire industry.  Get AI right and the efficiency improvements can be significant.  But if the AI-RAN introduces new problems, the business risk can be significant.