Autonomous Networks and ADN, a Discussion with Huawei’s Stephen Shao

During the Ultra-Broadband Forum 2024 Mark Mortensen had an opportunity to sit down with Stephen Shao, vice president of the General Development Department at Huawei.

During the Ultra-Broadband Forum 2024 on 30 October through 1 November in Istanbul, Türkiye, I had an opportunity to sit down with Stephen Shao, vice president of the General Development Department at Huawei. We discussed Huawei’s views on how an operator can accelerate its evolution to autonomous networks, through Level 3 to early phases of Level 4. 

One Map, One Master

Key to achieving AN Level 4 operations autonomy where the network manages itself is to have two main capabilities in place. The first is a comprehensive digital twin of the network, including equipment, services, and their status, which are typically displayed as a digital map, One Map. The second is a set of intelligent applications based on the Ultra-Broadband telecom foundation model, One Master. This capability is incorporated into the Huawei iMaster NCE system. The collaboration of One Map and One Master enables intelligent network decision-making by itself, just like a self-driving car in the road, for automatic service configuration, early risk detection, real-time traffic optimization, and fast service recovery, helping operators quickly evolve to AN Level 4.

MyTake: I have always been a strong advocate for a consolidated view of the network topology, equipment, services, and the status. We now call that a digital twin. It can be a federated system or a single system.

ADN Agents and AN L4

To achieve AN Level 4 operations autonomy takes a major new capability, an AI agent. These agents can assess the situation, decide if any actions should be taken, and take actions that are within their assigned scope of capabilities (or ask for permission, if not). Practically speaking, there needs to be many of them, each devoted to specific scenarios such as network maintenance, optimization, and operations.

There is still a major challenge, however, in the accuracy of the decisions made by the ADN agents. Both from the general issue of AI hallucinations as well as the differences in the way the operators like to plan, configure, and maintain their networks. Since different operators use different network layouts and management rules, it is not possible to have a single, generic AI agent for all operators. Rather, it needs to be layered in its training with knowledge of:

  • language (whether English, Chinese, etc.),
  • telecommunications terms and concepts,
  • the specific capabilities of equipment and systems that the operator has deployed, which are presented through a set of APIs,
  • the specific policies and procedures of the operator.

Huawei sees that AI agent as built into the iMaster NCE. It would work closely with the digital twin, gaining current knowledge of the network and services, determine if any network changes should be made, and making those changes, according to the rules set by the operator.

MyTake: The Huawei approach of having the digital twin and the agent be tightly coupled is probably necessary, given the level of available technology.

Retraining AI Copilots and Agents

When there are new specific capabilities of equipment or system, or new specific policies and procedures of the operator, beyond parameter value changes or adding new network elements, then retraining of the AI agent will be necessary. With the layered approach, retraining should be accomplished reasonably easily. But further research is necessary to determine the triggers for retraining and methodology at the lowest cost. In particular, the Huawei approach is to provide tools to modify the chain-of-thought via natural language or updated inference model for operators to use to add their specific capabilities, policies and procedures to the AI agent.

MyTake: The issue of when and how to retrain any AI, whether a CoPilot or an agent, is an important one. I see the need for “AI wranglers” in the operators’ workforces that make these kinds of decisions. The marrying of AI and operations process management tools is also a good approach. I expect that vendors will work to differentiate their solutions in this area of AI training.

Huawei Already Has Commercial Deployments

Huawei has implemented AI CoPilots and agents already in the field, pushing towards AN L4. Of note are:

  • China Mobile (2023) the first copilot-based SPN, FMEmate for the IP transport network. It helps onsite engineers troubleshoot issues without NOC support.
  • AssurSpirit (2024), an AI agent for troubleshooting, based on chain-of-thought technology. It can learn through natural language input and adjust or generate new fault diagnosis logic accordingly. This allows its knowledge to evolve under the direction of the operator. So far, this system provides the capabilities of 120 digital employees in the Guangzhou SPN network. The level of expertise is equivalent of a human employee with eight years of experience. The system has reduced the number of trouble tickets by 15% and brought the fault location duration to five minutes. In addition, on-site maintenance personnel can execute self-troubleshooting in real time without help from NOC personnel.
  • In the future, AI agents will be built for other scenarios such as network optimization.

MyTake: Huawei has a practical way to reach L3 and push toward L4: There is a lot of hype around AI agents and CoPilots, with every vendor and systems integrator developing neural net AI CoPilots and with autonomous AI agents being the next step. The Huawei layered approach, including chain-of-thought procedural programming, is very practical as it will allow its customers to implement their own processes and policies without requiring detailed changes or major retraining of the AIs in most cases.

The most impressive part of the Huawei push into AN L3 and L4 is how it is working so closely with leading operators around the world. These AI system are already working.