Efficient and intelligent resource allocation in optical networks

Doctoral thesis, 2025

The exponential growth in bandwidth demand, driven by emerging network services with diverse requirements, necessitates a cost-efficient design and effective resource management in optical networks. This requires approaches that enhance network architecture flexibility, spectrum usage efficiency, and automation while maintaining low operational costs. The thesis addresses these challenges through three key contributions: developing a cost-efficient network planning approach, introducing machine learning-based approaches for dynamic resource reallocation, and extending resource management strategies to multi-band elastic optical networks (EONs) to improve spectrum usage efficiency.

To optimize resource utilization while maintaining network flexibility, the thesis studies programmable filterless optical networks (PFONs), an architecture that replaces conventional reconfigurable optical add-drop multiplexers with programmable optical white box switches. The routing, modulation format, and spectrum assignment problem in PFONs is formulated as an integer linear program aimed at minimizing spectrum and passive optical component usage. Simulation results show a 54% reduction in spectrum dissipation compared to passive filterlessoptical networks, while also achieving greater cost efficiency over conventional wavelength-switched optical networks.

A significant obstacle to efficient resource usage in dynamic single- and multi-band EONs is spectrum fragmentation (SF), where arrivals and departures of service requests leave stranded, unusable gaps in the available spectrum. To alleviate SF, we introduce DeepDefrag, a novel deep reinforcement learning-based framework designed to address spectrum defragmentation (SD) challenges. DeepDefrag dynamically determines appropriate timing for SD, selects connections to be reconfigured, and identifies suitable parts of the spectrum for reallocation. Through intelligent decision-making, DeepDefragoutperforms traditional heuristic algorithms, such as the older first-fit algorithm, by achieving a lower service blocking ratio (SBR) andminimizing the control overhead associated with SD.

To further extend resource management to multi-band EONs, where different achievable quality of transmission (QoT) levels across different bands exacerbate fragmentation, the thesis proposes a fragmentation- and QoT-aware routing, band, modulation format, and spectrum assignment algorithm. The approach integrates proactive SD and traffic re-grooming to improve spectral efficiency. Simulations on three network topologies demonstrate a significant reduction in both SBR andspectrum fragmentation compared to QoT-only benchmarks, albeit with a slight increase in the average path length.