Design and implementation of a lightweight autonomous AI agent for SRV6 multi-criteria optimization within frrouting

DOI: 10.31673/2412-9070.2025.050233

Abstract

Modern info-communication networks, increasingly leveraging Segment Routing over IPv6 (SRv6) for enhanced flexibility and programmability, face significant challenges in achieving dyna-mic and effective multi-criteria optimization (MCO). The relentless growth in network traffic volume, the escalating diversity of services driven by 5G, IoT, and edge computing, and the highly dynamic nature of performance demands necessitate a paradigm shift from traditional, often manual or statically configured, network management towards more autonomous and intelligent control systems. SRv6 provides a powerful architectural foundation for this evolution by encoding routing instructions directly within the IPv6 data plane, yet harnessing this programmability to simultaneously optimize multiple, often conflicting, criteria—such as latency, throughput, reliability, and resource utilization—remains a complex undertaking, particularly under fluctuating network conditions and diverse application requirements.
Traditional centralized network management approaches, including those based on Software-Defined Networking (SDN) controllers, often encounter limitations related to scalability, potential single points of failure, and the inherent latency involved in collecting global network state and distributing control commands. Conversely, deploying sophisticated decision-making intelligence directly onto network devices, while offering the promise of faster localized responses and enhanced resilience, is frequently hindered by the inherent constraints in computational resources (CPU, memory) typical of standard routing hardware. This paper specifically addresses the critical feasibility of de-signing and deploying a controller-less, autonomous Artificial Intelligence (AI) agent directly on a Linux-based routing platform to perform MCO for SRv6 traffic engineering. A core aspect of our investigation is the agent's seamless integration with the widely adopted FRRouting open-source routing suite, which serves as the operational platform for both monitoring network state and enacting SRv6 policy modifications.
We present the detailed design principles and a comprehensive implementation strategy for a lightweight AI agent. This agent is specifically architected to utilize resource-efficient Reinforcement Learning (RL) and/or Graph Neural Network (GNN) techniques, which are particularly well-suited for operation within such constrained environments. The proposed agent functions as a distinct software process, running independently yet interacting locally with the co-located FRRouting daemons (e.g., zebra, bgpd, ospfd/isisd, pathd). This interaction is facilitated through standard, well-defined Application Programming Interfaces (APIs), such as YANG/NETCONF, gRPC, or REST, operating over local Inter-Process Communication (IPC) mechanisms. This local API-driven approach enables the agent to continuously monitor relevant network state parameters derived from FRRouting and autonomously apply SRv6 Traffic Engineering (TE) policy modifications back to the FRRouting suite without reliance on any remote controller.
The paper meticulously details the overall system architecture, the conceptual components underpinning the agent's intelligent decision-making process (including state representation derived from network telemetry, the defined action space corresponding to SRv6 policy controls, and the formulation of multi-criteria objective functions), the specific local API-based integration mechanisms with FRRouting and the planned implementation methodology. This methodology leverages Python as the primary development language, augmented by standard AI libraries (e.g., TensorFlow, PyTorch), and employs a containerized environment (Docker) for consistent deployment and rigorous validation within the Mininet network emulator. The offline training of the AI models is envisioned to utilize scalable cloud platforms like Google Cloud Platform (GCP) to handle computational demands. The primary scientific and practical contribution of this work lies in the thorough demonstration of the feasibility of designing and implementing this novel autonomous, on-device agent architecture. By showcasing its capability to effectively interact with a standard, production-grade routing platform like FRRouting for the purpose of SRv6 MCO, this research paves the way for the development of more adaptive, resilient, and intelligent network control strategies deployed directly within the network infrastructure, thereby fostering a new generation of decentralized and autonomous network management solutions.

Keywords: autonomous network management; artificial intelligence; reinforcement learning; segment routing; SRv6; multi-criteria optimization; FRRouting; model; local API; inter-process communication (IPC).