Promoting a Highly-Stable Core Network with Intelligent O&M Powered by Foundation Models

Host: The emergence of ChatGPT in 2023 has made large models a huge trend in all industries across the globe. And how to quickly build large models and achieve deep integration between AI and cloud core scenarios is a clear focus of attention. So Eric, what news can you tell us about large models in relation to the Huawei Cloud Core Network's ADN solution?

Eric: This year, we launched the ICN Master solution powered by the Large Language Model. In the era of Level-3 autonomous networks, the network O&M, basically rely on persons, tools, and the small models. However, it is a challenge to develop generalization and end-to-end closed-loop capabilities for the specific scenarios and use cases. Large models, in contrast, offer extensive knowledge, natural language understanding, task planning, reasoning, and scenario generalization capabilities. It can effectively overcome the O&M bottleneck of Level 3, and enable the transition to Level-4 autonomous networks, from automation to intelligence. Our pioneering applications of large models in core network O&M address critical issues such as high network risks, low operational efficiency, high skill requirements, and the lack of real-time service experience assurance. By focusing on creating a core network with high reliability, extreme simplicity, and optimized experience, we aim to empower operators with enhanced user services.

Host: So, Eric, in terms of intelligent networks. Most of what we hear in this regard relates to efficiency improvements with O&M automation. Can you explain why deploying intelligence in the core network emphasizes high network stability?

Eric: Global operators and industry organizations recognize the importance of stability in core networks. However, reaching a consensus is just the beginning. It is essential to establish network standards and define the attributes of a highly reliable core network. This forms the foundation for advancing automation and intelligence capabilities. Additionally, a reliable evaluation tool is crucial to assist operators in network evaluation. This tool should offer both static and dynamic evaluations of network reliability through fault injection and automated simulation. Our active involvement in the ANL evaluation and contributions to the TM Forum standards enable operators to proactively identify potential network risks based on measurable high stability standards.

Host: Robert, Appledore has been studying the self-intelligent network field, and the reports you've released have received frequent attention in the industry. So, what's your take on the relationship between self-intelligent networks and high-stability networks?

Robert: Yeah, it's very clear there's a close relationship between these two things. We're moving from a world where changes to the network used to be complicated, undesirable, high risk. We're moving to a world where that changes need to be made very frequently and very intelligently. For that, we have to rely on network data, and that ability to make change is only really safe if it comes with keeping the network stable.

Autonomous networks are capable of self-planning, self-deployment, self-optimization, and self-evolution. These capabilities enable networks to cope with various challenges and changes intelligently, reduce the probability of faults, and improve network stability and reliability.

With automatic and intelligent O&M methods, the autonomous network can monitor the network status in real time, predict and prevent potential problems, and quickly respond to and automatically rectify faults, ensuring high network stability. It can be said that an important goal of the autonomous network is to improve the stability and reliability of the network through automation and intelligence, and to realize the state of the high stability network.

Host: So you mentioned the benefits of intelligent networks, but what difficulties do you think exist in applying large models to the core network O&M field?

Robert: Yeah, it's clear we're still at early days, but I'm going to mention a couple of challenges in particular.

The first is to do with the complexity of the data. Core network O&M involves a large amount of real-time data, multi-modal data (data, images, pictures, and charts), and complex data relationships. These data is often characterized by strong domain knowledge, and the data varies greatly in different scenarios. Large models need to process and understand the complex data for effective O&M decision-making.

The difference between generalized language models, large models, and more context-specific models, and how we apply them in real-world situations, is also a challenge.

Large models need to solve the problem of how to maintain strong generalization capabilities in different scenarios, so as to quickly promote reuse and process tasks in unknown scenarios. At the same time, such as network devices, protocols, and technical details, a large model needs to effectively integrate knowledge of these fields to accurately diagnose faults and optimize performance.

Host: Eric, one thing we'd all like to know is how does our ICN Master solve these technical challenges?

Eric: Yeah, of course, we're also facing these problems, but we try to do some new ways to solve these problems. One, we need to improve our data quality, because we think it's very important to train the model to understand our intention.

And another one is that we combine different ways to improve the model's capabilities.

ICN Master integrates agents with long-term and short-term memory, RAG, and small models to facilitate semantic understanding, tool invoking, and logical reasoning. This approach enhances the intelligence of core network troubleshooting and complaint handling.

With over 30 years' experience in core network O&M and accumulated cases, we have developed a dedicated multi-modal analysis model for O&M data, covering signaling, time sequences, and network topology. Moreover, we have created standard training sets derived from this knowledge corpus for pre-training.

We are continuously refining our methods to enhance the accuracy and efficiency of foundation model training. We hope that more partners can join us in addressing technical challenges collaboratively.