Applying generative AI to revolutionize telco network operations

Generative AI is shaping the future of telecommunications network operations. The potential applications for enhancing network operations include predicting the values of key performance indicators (KPIs), forecasting traffic congestion, enabling the move to prescriptive analytics, providing design advisory services and acting as network operations center (NOC) assistants.

In addition to these capabilities, generative AI can revolutionize drive tests, optimize network resource allocation, automate fault detection, optimize truck rolls and enhance customer experience through personalized services. Operators and suppliers are already identifying and capitalizing on these opportunities.

Nevertheless, challenges persist in the speed of implementing generative AI-supported use cases, as well as avoiding siloed implementations that impede comprehensive scaling and hinder the optimization of return on investment.

In a previous blog, we presented the three-layered model for efficient network operations. The main challenges in the context of applying generative AI across these layers are:

• Data layer: Generative AI initiatives are data projects at their core, with inadequate data comprehension being one of the primary complexities. In telco, network data is often vendor-specific, which makes it hard to understand and consume efficiently. It is also scattered across multiple operational support system (OSS) tools, complicating efforts to obtain a unified view of the network.
• Analytics layer: Foundation models have different capabilities and applications for different use cases. The perfect foundation model does not exist because a single model cannot uniformly address identical use cases across different operators. This complexity arises from the diverse requirements and unique challenges that each network presents, including variations in network architecture, operational priorities and data landscapes. This layer hosts a variety of analytics, including traditional AI and machine learning models, large language models and highly customized foundation models tailored for the operator.
• Automation layer: Foundation models excel at tasks such as summarization, regression and classification, but they are not stand-alone solutions for optimization. While foundation models can suggest various strategies to proactively address predicted issues, they cannot identify the absolute best strategy. Toevaluate the correctness and impact of each strategy and to recommend the optimal one, we require advanced simulation frameworks. Foundation models can support this process but cannot replace it.

Essential generative AI considerations across the 3 layers

Instead of providing an exhaustive list of use cases or detailed framework specifics, we will highlight key principles and strategies. These focus on effectively integrating generative AI into telco network operations across the three layers, as illustrated in Figure 1.

We aim to emphasize the importance of robust data management, tailored analytics and advanced automation techniques that collectively enhance network operations, performance and reliability.

1. Data layer: optimizing telco network data using generative AI

Understanding network data is the starting point for any generative AI solution in telco. However, each vendor in the telecom environment has unique counters, with specific names and value ranges, which makes it difficult to understand data. Moreover, the telco landscape often features multiple vendors, adding to the complexity. Gaining expertise in these vendor-specific details requires specialized knowledge, which is not always readily available. Without a clear understanding of the data they possess, telecom companies cannot effectively build and deploy generative AI use cases.

We have seen that retrieval-augmented generation (RAG)-based architectures can be highly effective in addressing this challenge. Based on our experience from proof-of-concept (PoC) projects with clients, here are the best ways to leverage generative AI in the data layer:

• Understanding vendor data: Generative AI can process extensive vendor documentation to extract critical information about individual parameters. Engineers can interact with the AI using natural language queries, receiving instant, precise responses. This eliminates the need to manually browse through complex and voluminous vendor documentation, saving significant time and effort.
• Building knowledge graphs: Generative AI can automatically build comprehensive knowledge graphs by understanding the intricate data models of different vendors. These knowledge graphs represent data entities and their relationships, providing a structured and interconnected view of the vendor ecosystem. This aids in better data integration and utilization in the upper layers.
• Data model translation: With an in-depth understanding of different vendors' data models, generative AI can translate data from one vendor's model to another. This capability is crucial for telecom companies that need to harmonize data across diverse systems and vendors, ensuring consistency and compatibility.

Automating the understanding of vendor-specific data, generating metadata, constructing detailed knowledge graphs and facilitating seamless data model translation are key processes. Together, these processes, supported by a data layer with RAG-based architecture, enables telecom companies harness the full potential of their data.

2. Analytics layer: harnessing diverse models for network insights

On a high level, we can split the use cases of network analytics into two categories: use cases that revolve around understanding the past and current network state and use cases thatpredict future network state.

For the first category, which involves advanced data correlations and creating insights about the past and current network state, operators can leverage large language models (LLMs) such as Granite™, Llama, GPT, Mistral and others. Although the training of these LLMs did not particularly include structured operator data, we can effectively use them in combination with multi-shot prompting. This approach helps in bringing additional knowledge and context to operator data interpretation.

For the second category, which focuses on predicting the future network state, such as anticipating network failures and forecasting traffic loads, operators cannot rely on generic LLMs. This is because these models lack the necessary training to work with network-specific structured and semi-structured data. Instead, operators need foundation models specifically tailored to their unique data and operational characteristics. To accurately forecast future network behavior, we must train these models on the specific patterns and trends unique to the operator, such as historical performance data, incident reports and configuration changes.

To implement specialized foundation models, network operators should collaborate closely with AI technology providers. Establishing a continuous feedback loop is essential, wherein you regularly monitor model performance and use the data to iteratively improve the model. Additionally, hybrid approaches that combine multiple models, each specializing in different aspects of network analytics, can enhance overall performance and reliability. Finally, incorporating human expertise to validate and fine-tune the model's outputs can further improve accuracy and build trust in the system.

3. Automation layer: integrating generative AI and network simulations for optimal solutions

This layer is responsible for determining and enforcing optimal actions based on insights from the analytics layer, such as future network state predictions, as well as network operational instructions or intents from the operations team.

There is a common misconception that generative AI handles optimization tasks and can determine the optimal response to predicted network states. However, for use cases of optimal action determination, the automation layer must integrate network simulation tools. This integration enables detailed simulations of all potential optimization actions using a digital network twin (a virtual replica of the network). These simulations create a controlled environment for testing different scenarios without affecting the live network.

By leveraging these simulations, operators can compare and analyze outcomes to identify the actions that best meet optimization goals. It is worth highlighting that simulations often leverage specialized foundation models from the analytics layer, like masked language models. These models allow manipulating parameters and evaluating their impact on specific masked parameters within the network context.

The automation layer leverages another set of use cases for generative AI, namely the automated generation of scripts for action execution. These actions, triggered by network insights or human-provided intents, require tailored scripts to update network elements accordingly. Traditionally, this process has been manual within telcos, but with advancements in generative AI, there's potential for automatic script generation. Architectures with generic LLMs augmented with retrieval-augmented generation (RAG) show good performance in this context, provided operators ensure access to vendor documentation and suitable methods of procedure (MOP).

Generative AI plays a significant role in future telco operations, from predicting KPIs to responding to network insights and user intents. However, addressing challenges such as efficient data comprehension, specialized predictive analytics and automated network optimization is crucial. IBM has hands-on experience in each of these areas, offering solutions for efficient data integration, specialized foundation models and automated network optimization tools.

Interested in implementing generative AI use cases in your network? Bring us your use case and let us unlock its full potential. Contact us at [email protected]  and  [email protected].