.png){kind=icon}
Return
Transformer-pointer DRL model for static resource allocation problems in SDM-EONs
.png){kind=icon}
PRE
AI
DOI：10.1364/JOCN.580228.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The static resource allocation problem in space-division multiplexing elastic optical networks (SDM-EONs) requires joint optimization of routing, modulation, space, and spectrum assignment (RMSSA) for efficient resource use. Because integer-programming models and sophisticated heuristics are computationally expensive, the First-Fit algorithm is often used for fast feasible solutions, yet its quality is usually poor and highly order-dependent. To address this, we propose a deep reinforcement learning method for static resource allocation. We reformulate the combinatorial problem as sequence optimization by pairing with a fixed First-Fit allocator and prove that, for the space-spectrum assignment (SSA) subproblem, First-Fit is order-expressive and can achieve an optimal solution under a suitable ordering. A Transformer encoder extracts features of the request set, and a pointer-network decoder optimizes the output order. The trained network can be used as a black-box heuristic. Compared with hand-crafted orderings, it delivers higher solution quality with rapid inference speed, making it highly suitable for time-sensitive reconfiguration and large-scale planning tasks, while maintaining a runtime comparable to First-Fit. Finally, ablation studies verify the contributions of the encoder and decoder, and we attempt to interpret the trained network.
Keywords:
Resource management
Decoding
Dynamic scheduling
Transformers
Routing
Deep reinforcement learning
Training
Space division multiplexing
Reviews
Optical switches
Journal
IF:
4.3
Papers:
2.2K
Citations:
3.8K