Adaptive high-efficiency RWA algorithm for optical networks based on reinforcement learning

In optical networks, most of the current routing and wavelength assignment (RWA) algorithms based on deep reinforcement learning (DRL) rely on the K shortest path (KSP) routing algorithm to calculate alternative paths in advance. This method lacks flexibility and dynamics, and KSP needs to be recalculated for all node pairs when the network topology changes. To solve the above problems, this paper proposes an adaptive high efficiency (ADE)-RWA framework based on DRL. The key point and innovation of the ADE-RWA framework proposed in this paper is that during the training process, the DRL agent selects the optimal single-step connection from the current node when selecting actions, rather than the pre-calculated complete path, so that the routing strategy can be changed dynamically and in real time according to the changes in the network status to adapt to the changes in network topology such as link failures. Another key point of the ADE-RWA framework is that the DRL training process not only trains its decision neural network, but also stores it in the form of a routing table (LUT) every time a path is successfully assigned. After the DRL training converges, the LUT information of the network path will also be complete. At this time, using the LUT to directly search for available paths can effectively reduce the computational cost and improve the efficiency of the RWA algorithm. In addition, LUT search and DRL training can be dynamically switched. The proposed algorithm dynamically senses the changes in network status. When network failures cause topology changes, it switches to the DRL stage to continue training, ensuring the dynamic update of LUT. Experimental results show that compared with the KSP-First Fit (FF) and Deep Learning for Routing Modulation and Spectrum Allocation (DeepRMSA) algorithms, the blocking rate of the proposed ADE-RWA framework is reduced by 36% and 30% respectively; and when a link failure occurs, the ADE-RWA framework can quickly adapt to changes in network topology. This paper proposes a DRL-based ADE-RWA framework, which realizes adaptive RWA under dynamic optical network topology and reduces computational costs.