First Field Trial of LLM-Powered AI Agent for Lifecycle Management of Autonomous Driving Optical Networks

Xiaomin Liu,Qizhi Qiu,Yihao Zhang,Yuming Cheng,Lilin Yi,Weisheng Hu,Qunbi Zhuge

2024-09-24

Abstract:We design and demonstrate the first field trial of LLM-powered AI Agent for ADON. Three operation modes of the Agent are proposed for network lifecycle management. The Agent efficiently processes wavelength add/drop and soft/hard failures, and achieves comparable performance to human-designed algorithms for power optimization.

Systems and Control

What problem does this paper attempt to address?

This paper attempts to address the problem of automating full lifecycle management in Autonomous Driving Optical Networks (ADON). Specifically, the authors design and demonstrate the application of the first AI agent based on Large Language Models (LLM) in a real-world environment, aiming to handle various events in the network lifecycle, such as wavelength addition/deletion, soft/hard faults, etc., through three different operational modes (LLM native mode, LLM-centric mode, and rule-centric mode). The goal of the paper is to verify whether this AI agent can efficiently handle these events and achieve performance comparable to human-designed algorithms.

### Main Issues:
1. **Full Lifecycle Management**: How to achieve automated full lifecycle management from network establishment to retirement in ADON.
2. **Event Handling**: How to automatically handle typical events in the network lifecycle, such as wavelength addition/deletion, soft/hard faults, etc.
3. **Performance Optimization**: How to optimize Optical Amplifiers (OA) through the AI agent to achieve near-optimal Q-factor.
4. **Digital Twin**: How to build a high-precision digital twin model and synchronize network status in real-time.
5. **Fault Localization and Recovery**: How to quickly locate and recover from faults, including fiber breaks and link aging.

### Solutions:
- **LLM Native Mode**: Completely relies on LLM for planning, reasoning, and execution without using any human-designed algorithms.
- **LLM-Centric Mode**: Utilizes LLM for planning and reasoning but combines some human-designed algorithms.
- **Rule-Centric Mode**: LLM guides the planner to retrieve and execute predefined workflows and existing tools/algorithms.

### Experimental Results:
- **Wavelength Addition/Deletion**: In different modes, the AI agent can efficiently handle wavelength addition/deletion events and optimize the gain of optical amplifiers.
- **Fault Handling**: The AI agent can successfully locate and recover from faults such as fiber breaks and link aging within tens of seconds.
- **Performance Comparison**: Some LLMs achieved performance comparable to Bayesian optimization methods in optimizing optical amplifier gain.

Through these experiments, the paper demonstrates the great potential of LLM in handling complex tasks and optimizing ADON performance.