Routing and spectrum assignment employing long short-term memory technique for elastic optical networks

With the prevalence of some high bandwidth-demanding applications, such as cloud computing, traditional wavelength-division-multiplexing passive optical networks have difficulties in satisfying such growing bandwidth demands due to its limited allocation-flexibility and utilization-efficiency. Therefore, elastic optical networks (EONs). In order to realize the flexibility in EONs, sophisticated routing and spectrum allocation (RSA) algorithms areone of the keyenabling technologies. However, most of the previous RSA algorithms were proposed with invariant routing and spectrum allocation strategies, which ignored considering the time-varying characteristics of EONs due to the variable network architecture and service provisioning. And such time-varying characteristics can deteriorate the spectrum fragmentation and the service blocking performances of EONs, which stimulates the application of various machine-learning technologies in EONs. In this paper, a long short-term memory based routing and spectrum assignment (LSTM-RSA) algorithm is proposed for EONs. By employing the long short-term memory technique to sense the complex status of EONs (e.g. spectral usage on the selected paths), the proposed LSTM-RSA algorithm gradually learns successful strategies through accumulating operation experience in the process of interaction and obtains higher returns through enhanced operation, which helps improve the spectrum fragmentation and the service blocking performances in EONs. Simulation results show that the spectrum fragmentation rate and the blocking rate of the proposed LSTM-RSA algorithm are reduced by about 6% and 8.9%, respectively, when compared to the traditional shortest-path-routing first-fitting RSA algorithm.