The findings show that OBS-GraphSyn-2025 offers a scalable, interpretable, and reproducible complexity-aware metric for evaluating routing-state in OBS-inspired networks while abstracting from the constraints of the optical-layer.
Abstract
Optical Burst Switching (OBS) networks require adaptive and interpretable routing mechanisms to handle dynamic traffic variation and structural complexity. In this work, they aim to design a benchmark-oriented framework, called OBS-GraphSyn-2025, that combines graph-derived structural modeling, traffic temporal representation, and reinforcement learning based on PPO to evaluate the adaptivity of burst routing. It converts 3.57 M real traffic flows into 199,870 routing flows (with topology information), 79,133 burst-level routing states, and 15,000 PPO-compatible samples. A Traffic Burst Complexity Index-Graph (TBCI-G) is proposed, which combines the routing state complexity metrics of hop count, path complexity, edge load, and route uniqueness to quantify routing complexity. The PPO policy is based on fused states across time and space, and is able to adaptively route under controlled benchmark conditions using reward signals. The results of experimental evaluation over 250 episodes show stable learning performance, with a mean reward of 0.5610 ± 0.0068, throughput of 0.6717 ± 0.0709, and stability score of 0.5856 ± 0.0627. The scalability analysis showed that the runtime increases by a controlled amount (0.32-1.56 s) and the reward retention remains fairly unchanged (0.97-1.00). The consistency of the policy behavior was assessed by reproducibility, and resulted in a coefficient of variation of less than 1.20% in 30 independent runs. The findings show that OBS-GraphSyn-2025 offers a scalable, interpretable, and reproducible complexity-aware metric for evaluating routing-state in OBS-inspired networks while abstracting from the constraints of the optical-layer.
Load-aware routing in Software-Defined Networking (SDN) has two structural limitations: path selection is based on instantaneous link utilization, forcing oscillations in routing as the controller makes decisions that perturb the measurements driving them; and this instantaneous view cannot capture the time-scale over which link reliability accumulates. We present HAASP (History-Aware Adaptive Shortest-Path), a closed-loop SDN routing controller that composes static topology weights, decaying load memory, and decaying fault memory into a single bounded cost. A hysteretic state machine adapts the decay coefficient to network conditions without itself oscillating, and a per-flow migration hysteresis bounds route-change rates. We introduce the Path Stability Index (PSI), a byte-weighted per-flow stability metric, and prove closed-form bounds on penalty growth, path stretch, convergence, and PSI. HAASP is evaluated against five baselines (static Dijkstra, ECMP, EWMA-smoothed routing, CFR-RL, DRSIR) across four topologies, four failure scenarios, and a full hyperparameter sweep, with a scalability study to 200 nodes.
F. Nife, S. Hatem· Muthanna Journal of Pure Sci...· 0 citations
Fast and stable routing convergence is critical in large IP networks, and the interior-gateway-protocol (IGP) timers that govern failure detection (Hello/Dead intervals) expose a fundamental tension: aggressive timers detect failures quickly but inflate control overhead and trigger route flaps, whereas conservative timers are stable but slow. Recent work such as DRL-Adapt has shown that deep reinforcement learning can tune these timers better than static defaults, but it operates on a flat, globally-aggregated network state and emits a single network-wide timer, so it can neither exploit the spatial heterogeneity of real topologies nor transfer architecturally across networks of different size. We propose GraphRoute-Transfer, a graph-neural-network policy that assigns per-node timers from local structural features and is by construction permutation- and size-invariant. Because control-plane fragility and failure criticality are spatially heterogeneous, the cost-minimizing timer assignment varies across the graph; our policy learns this mapping and applies it zero-shot to unseen topologies of arbitrary size. Training is guided by a coordinate-descent search oracle on a convergence-cost objective, so the expensive per-topology optimization is amortized into a sub-millisecond inference. On 231 real topologies from the Internet Topology Zoo, GraphRoute-Transfer reduces mean convergence time by 37.3% relative to the OSPF default and to a flat DRL baseline, lowers the composite convergence-cost objective by 16.4% over the flat baseline, and attains 1.317 cost—within 0.3% of the search oracle—while running about 8,160× faster than the search. Crucially, a policy trained only on networks with ≤70 nodes maintains its gains on unseen networks up to 140 nodes, whereas the flat baseline degenerates to a global constant that cannot adapt.
Yuto Nakamura· Journal of Computing and Ele...· 0 citations
INTRODUCTION: The development of modern power systems imposes stringent requirements on communication networks, including highly dynamic loads, low latency, and high reliability. Although recent Software-Defined Networking routing, link-utilization-aware scheduling, and reinforcement learning-based methods improve path optimization, challenges remain in bottleneck-link perception, congestion feedback, and stable decision-making under dynamic traffic conditions.
OBJECTIVES: To counteract problems like delayed response times and inadequate congestion identification in conventional path selection algorithms owing to dynamic link modifications, this research paper presents a path selection model that combines bottleneck link usage and reinforcement learning.
METHODS: Under the Software-Defined Networking control architecture, the proposed model incorporates link utilization, Graph Convolutional Network structures, and Gated Recurrent Units, while introducing a deep reinforcement learning algorithm to optimize routing strategies.
RESULTS: Experimental results demonstrate that under a 60 Mbit/s load, the proposed method achieves a throughput of 57 Mbit/s, maintains the minimum transmission delay within 0.043 s, and yields a link utilization rate of 89%. In experiments carried out to evaluate dynamic decision-making, the approach for choosing paths adopted by the model averages convergence at the 15th round, resulting in an error rate of 4.2%, minimal load balancing at 0.23, and median latency time in real-time decisions of only 31 ms, which is better than other models.
CONCLUSION: These results demonstrate that the model achieves superior state awareness and adaptive routing performance in multi-source heterogeneous networks. The model also exhibits good performance in terms of congestion control and path optimization and hence can be used effectively for intelligent routing in next-generation power communication networks.
Ying Zeng, Xingnan Li, Yubeng Bao et al.· EAI Endorsed Transactions on...· 0 citations
This work proposes Double-Channel Graph Attention (DCGA), an end-to-end reinforcement learning framework that isolates network reachability and demand-service logic into separate graph channels and constructs valid routes using a simulator-coupled, constraint-informed decoder.