Hierarchical Reinforcement Learning in Multi-Domain Elastic Optical Networks to Realize Joint RMSA

Optimizing performance in elastic optical networks using advanced reconfigurable optical add-drop multiplexers: A novel design approach and comprehensive analysis

“…Notably, achieving higher spectrum efficiency at information rates above 100 Gb/s becomes particularly concerning when comparable optical reaches are required [8]. Elastic optical networks (EONs) are a promising solution that provides configurable granularity aggregation, flexible spectrum grids, and improved spectral efficiency compared to traditional WDM systems [3][4][5][6][7][8][9]. The flexible grid concept was introduced in the 2012 International Telecommunication Union (ITU-GT) guideline (694.1), reinforcing the spectrum allocation approach of EON.…”

“…Furthermore, this core frequency is an integral multiple of 6.25 GHz and encompasses various frequency ranges comprising an optical channel. The slot width is also determined by an integral multiple of 12.5 GHz [7,9]. For example, a 32 GB symbol rate with zero roll-off Nyquist filtering presents a 56 % spectral efficiency enhancement [9].…”

“…The slot width is also determined by an integral multiple of 12.5 GHz [7,9]. For example, a 32 GB symbol rate with zero roll-off Nyquist filtering presents a 56 % spectral efficiency enhancement [9].…”

Busyness level-based deep reinforcement learning method for routing, modulation, and spectrum assignment of elastic optical networks

“…As a result, many heuristic RMSA schemes are proposed [3]. Recently, deep reinforcement learning (DRL) based approaches rise and show advantage over the conventional heuristic methods, thanks to the development of machine learning [4][5][6][7]. Under the DRL framework, the agent observes the EON state, emits a RMSA action and receives a reward from the EON which is to reflect the goodness of the action.…”

“…Since the agent's learning is based on the state it observes and the reward it receives, key information should be carried by the state and the reward. In previous studies [4][5][6][7], however, the observed and feedback information is limited.…”