Decentralized multi-criteria optimization of SRv6 traffic by an autonomous AI agent on frrouting: implementation and experimental validation

Abstract

A previous publication by the authors presented the architectural concept of an autonomous AI agent for multi-criteria optimization (MCO) of SRv6 traffic without an external controller, limited to a mathematical model without empirical realization. The present work moves for the first time from architectural design to full experimental validation on a working testbed (FRRouting 10.x, Containerlab, Linux 6.17) and provides quantitative evidence for the proposed approach. A decentralized MCO routing method for SRv6 is proposed, in which coordination among autonomous agents on edge routers is achieved through implicit stigmergy (shared reading of the synchronized IS-IS LSDB as a stigmergic medium, formalized via Heylighen's 5-tuple S = <A,T,M,Fa,Ft>) without a central arbiter of MCO weights and without direct inter-agent messaging; stability of decisions is confirmed by a five-phase experimental procedure - a single switch in each direction with zero false positives across three baseline phases (p<10^-3 against H0). Four-criterion optimization (latency, throughput, packet loss, jitter) and topology generalization are combined in a single GNN-PPO model (GCNConv x2, 50,179 parameters); the symmetric TRIGGER/WITHDRAW gate is realized without a hand-coded threshold via training-distribution-matched feature wiring (TDFW): the same network with the learned 40 ms SLA decision boundary drives switching in both directions. Production deployment on edge routers is enabled by ONNX Runtime with INT8 quantization (approximately 8 MB dependencies, approximately 0.13 ms inference) paired with eBPF XDP/TC kernel telemetry (1-2 ms path-RTT histogram readout per cycle); the full PyTorch training stack (approximately 700 MB, approximately 50 ms per inference) remains in the training loop only and does not enter the router's production cycle. End-to-end reaction times - TRIGGER 409 ms, WITHDRAW 825 ms - are structurally bounded by K-of-N hysteresis and probe rate, not by AI cost: the model accounts for less than 0.1 % of total reaction.

Keywords: SRv6, GNN-PPO, ONNX INT8, eBPF XDP/TC, multi-criteria optimization, FRRouting, implicit stigmergy, edge AI.