计算机科学
光谱(功能分析)
代表(政治)
物理
电子工程
拓扑(电路)
算法
光通信
信号处理
光学
数学
光谱密度
网络拓扑
扩频
编码(内存)
作者
Zhiwei Zhou,Danping Ren,Bowen Bao,Rentao Gu,Jijun Zhao
摘要
Routing, band, modulation-format, and spectrum assignment (RBMSA) under quality-of-transmission (QoT) constraints is a challenging problem in C+L-band elastic optical networks (EONs) because discrete allocation decisions are tightly coupled with dynamic spectrum occupancy and physical-layer impairments (PLIs). Many existing deep reinforcement learning (DRL)-based RBMSA methods handle QoT constraints through post hoc verification and rely on handcrafted, single-scale spectrum features, limiting their ability to learn QoT-feasible and fragmentation-aware allocation policies. To address these issues, we propose a deep Q-network (DQN)-based QoT-aware RBMSA framework with multi-scale spectrum representation for C+L-band EONs. The method incorporates generalized signal-to-noise ratio (GSNR)-related physical-layer information and Haar wavelet-based multi-scale spectrum features into the state space and integrates modulation-format selection into the action space, enabling QoT-aware joint path, band, and modulation-format decision-making. In addition, a dynamic reward-centering mechanism is incorporated to reduce reward-mean shift and improve DQN training stability under dynamic traffic. Simulation results show that, under high-load scenarios, the proposed method reduces the blocking probability by up to 86.27% and 43.99% compared with representative heuristic and DRL-based baselines, respectively, while alleviating spectrum fragmentation and improving the average GSNR of established lightpaths.
科研通智能强力驱动
Strongly Powered by AbleSci AI