Igaa Achieves 3.81% Performance Gain in Agentic AI Edge Service Scheduling

Researchers are tackling the challenge of reliably scheduling edge services in rapidly changing mobile networks. Yan Sun, Yinqiu Liu, and Shaoyong Guo, from Beijing University of Posts and Telecommunications and Nanyang Technological University, alongside et al, introduce a novel Intent-Driven General Agentic AI (IGAA) framework to overcome limitations in existing systems’ ability to generalise to new situations.This work is significant because it allows agentic AI to learn continuously from past experiences, improving scheduling capabilities and preventing performance drops when faced with unfamiliar demands. By combining a Network-Service-Intent matrix with innovative transfer learning algorithms and a generative replay mechanism, IGAA demonstrably adapts to new scenarios , even successfully applying lessons learned from real-time computing to optimise emerging Internet of Vehicles services with minimal impact on user satisfaction.

The team achieved this by designing a system that simulates novel scenarios and generates training datasets, allowing agents to proactively adapt to unforeseen circumstances.

Specifically, IGAA incorporates three core mechanisms to facilitate this learning process. First, a Network-Service-Intent matrix mapping method allows agents to simulate new scenarios and create training data, bridging the gap between high-level user requests and precise resource demands. Second, an easy-to-hard generalisation scheme, utilising Resource Causal Effect-aware Transfer Learning (RCETL) and Action Potential Optimality-aware Transfer Learning (APOTL) algorithms, helps IGAA adapt to new scenarios efficiently. These algorithms are designed to transfer learned policies to analogous new ones, such as optimising Internet of Vehicles (IoV) services using patterns derived from real-time computing.

Furthermore, to combat the issue of catastrophic forgetting during continual learning, the researchers propose a Generative Intent Replay (GIR) mechanism, synthesising historical service data to consolidate prior knowledge. Experiments demonstrate IGAA’s strong generalisation and scalability, showcasing its ability to rapidly adapt to new situations. The study reveals that IGAA maintains the intent-satisfaction rate gap within 3.81% compared to scenario-specific methods, a significant improvement in performance consistency. In unseen scenarios, IGAA outperforms the best existing method by 19.19%, highlighting its superior adaptability and decision-making capabilities.

To mitigate potential inaccuracies stemming from Large Language Model (LLM) hallucinations during scenario simulation, the team integrated a scenario evaluation and correction model, ensuring the generation of rational and reliable datasets. This breakthrough establishes a robust framework for general edge service scheduling, moving beyond the limitations of specialised agents. The research opens avenues for more intelligent and responsive edge networks, capable of seamlessly adapting to evolving user needs and resource availability. Researchers designed a Network-Service-Intent matrix mapping method to facilitate scenario simulation and generate training datasets for the agent. This innovative approach allows the agent to proactively anticipate and prepare for novel service demands. The team then presented an easy-to-hard generalization learning scheme, incorporating two customized algorithms: Resource Causal Effect-aware Transfer Learning (RCETL) and Action Potential Optimality-aware Transfer Learning (APOTL).

RCETL meticulously analyzes the causal effects of resource allocation on service performance, enabling the agent to transfer knowledge about resource utilization from familiar scenarios to new ones. Simultaneously, APOTL optimizes action selection by evaluating the potential optimality of different actions, ensuring efficient adaptation to changing conditions. These algorithms work in concert to facilitate rapid and effective learning in unseen environments. To combat catastrophic forgetting during continual learning, the research pioneered a Generative Intent Replay (GIR) mechanism. GIR synthesizes historical service data, effectively consolidating prior capabilities and preventing the agent from losing valuable knowledge as it learns new skills.

This ensures long-term performance stability and adaptability. Furthermore, to mitigate the impact of Large Language Model (LLM) hallucinations during scenario simulation, scientists integrated a scenario evaluation and correction model. This model rigorously assesses the rationality of generated scenarios and datasets, guiding the agent towards creating realistic and reliable training data. Extensive experiments demonstrated IGAA’s strong generalization and scalability, with the system achieving rapid adaptation by transferring policies to analogous new scenarios. For instance, latency-sensitive patterns learned from real-time computing were successfully applied to optimize novel Internet of Vehicles (IoV) services. Compared to scenario-specific methods, IGAA maintained the intent-satisfaction rate gap within 3.81%, and outperformed existing methods by 19.19% in unseen scenarios. This work highlights a significant advancement in agentic AI for edge computing, paving the way for more intelligent and adaptable service scheduling systems.