Decentralization of multi-criteria optimization of SRv6 routes based on implicit stigmergy: theoretical foundation and positioning among modern DRL approaches

Abstract

The article addresses the multi-criteria optimization (MCO) problem of routing in programmable SRv6 networks across four QoS criteria (latency, throughput, packet loss, jitter). Existing deep reinforcement learning (DRL) methods - AROM-DRL (Tayeh, 2022), MDQ (Zhang et al., 2024), QR-MO (arXiv:2503.18122, 2025), WA-SRTE (IEEE ICC, 2024) - solve MCO in a centralized SDN controller, where a single arbiter determines the weighted convolution of criteria or computes the Pareto set. A fundamentally different architecture is proposed: MCO is solved independently by each autonomous AI agent on edge routers, while the consistency of their decisions across the network is achieved through implicit stigmergy - shared reading of the IS-IS LSDB synchronized by LSP flooding as the stigmergic medium. The article formalizes Heylighen's coordination model (5-tuple S = <A, T, M, Fa, Ft>) applied to IGP LSDB, justifies the methodological role of open-stack constraints (FRRouting + IS-IS flat area + SRv6 BE + Linux LWT) as a worst-case validation strategy, and provides a five-axis comparison with the main modern works. The proposed architecture is shown to be the only one satisfying simultaneously all five criteria: >=3 QoS metrics, edge inference, GNN generalization, controller-less operation, eBPF-based reward telemetry. Empirical verification is carried out by a 15-minute five-phase experimental procedure that confirms the admissibility of the stigmergic approach as a single switch in each direction (1 TRIGGER, 1 WITHDRAW), zero false positives across three baseline phases, and reaction times of 409 ms (TRIGGER) and 825 ms (WITHDRAW). The 50 179-parameter GNN-PPO model (GCNConv 4->64->128 + global mean pool + MultiInputPolicy[64,64]) is deployed as an INT8-quantized ONNX artefact with sub-millisecond inference; wall-clock reaction is structurally bounded by K-of-N hysteresis and the eBPF probe interval, not by AI cost. The symmetric WITHDRAW barrier is implemented without a hard severity threshold via training-distribution-matched feature wiring - the same neural network at the same learned 40 ms SLA boundary drives switching in both directions.

Keywords: multi-criteria optimization, SRv6, IS-IS, implicit stigmergy, decentralization, GNN-PPO, edge inference, eBPF.