MARL-Converge: Decentralized Multi-Agent Reinforcement Learning for Failure-Aware Routing Convergence
Routing failures demand rapid control-plane convergence, yet uniformly aggressive detection and flooding increases signaling load and can induce unstable parameter oscillations. This paper presents MARL-Converge, a decentralized multi-agent reinforcement learning framework in which every router selects a local conservative, balanced, or aggressive control profile from router-local and one-hop telemetry. A parameter-shared dueling Double Deep Q-Network is trained with shared experience and a global-dominant team reward, while execution remains decentralized. A failure-aware action mask prevents routers without local evidence from blindly escalating to the aggressive profile. We implement a reproducible discrete-time link-state convergence proxy with independent topology views, stochastic failure detection and link-state advertisement diffusion, shortest-path recomputation, and forwarding-loop/drop checks. Evaluation comprises 600 matched episodes on the real Abilene and held-out AARNet topologies from the Internet Topology Zoo. Relative to a static balanced profile, MARL-Converge reduces convergence steps by 19.0% and modeled packet loss by 83.6%; the 6.1% overhead reduction is directionally favorable but not significant at the 0.05 level. Relative to a local heuristic, it is 9.3% faster, uses 14.0% fewer control units, and makes 56.4% fewer profile switches. It is 7.6% slower than always-aggressive control but cuts overhead by 35.4%. These results establish a measured speed-stability-efficiency trade-off without claiming physical-router timing equivalence.