Elastic Optical Network Resource Allocation Integrating Reinforcement Learning And Service Aware

The emergence of various video services,cloud computing,data centers,and other high-speed traffic services,has called for higher requirements on optical network that carry communication services.Conventional WDM network using fixed wavelength grid cannot flexibly allocate appropriate wavelength resources to smaller traffic services.At the same time,in order to avoid interference between adjacent wavelengths,WDM network set guard band between adjacent bandwidth,which occupy more wavelength sources.In order to solve the above-mentioned problems and improve the utilization of optical network resources,the concept of elastic optical network is proposed.Elastic optical network divides the fixed spectrum grid of WDM network into finer granularity,so as to meet different service requirements and realize more flexible resource allocation.In elastic optical network,finding appropriate routing and spectrum allocation strategies is one of the main study concerns,on which is this thesis focus.Firstly,this thesis proposes the DQN-RSA algorithm,applying the DQN framework in reinforcement learning to the RSA process of elastic optical network DQN-RSA algorithm can choose the best action as the output when routing and spectrum resources are allocated,using the perception capabilities of reinforcement learning and neural network.Simulation experiment results show that DQN-RSA algorithm performs better in terms of traffic blocking rate and spectrum fragmentation than traditional RSA algorithm based on the shortest path and first fit algorithm.Subsequently,this thesis proposes an RSA algorithm that combines service aware and reinforcement learning: SA-A3C-RSA.Using the A3 C algorithm in reinforcement learning and focusing on the various service demands and service types existing in the network,which are analyzed through service aware when the request arrives and the analyzed results are input into the A3 C algorithm neural network as state data.After extracting the feature of state data and performing the training of reinforcement learning by the neural network,the best path and spectrum resource allocation scheme of the current state is output,so as to establish the optical path for the current service request.The simulation experiment results show that the SA-A3C-RSA algorithm performs better than the SP-FF-RSA algorithm in terms of service blocking rate and spectrum utilization.Finally,the robustness of the SA-A3C-RSA algorithm is verified through experiments.Compared with the traditional SP-FF-RSA algorithm,the DQN-RSA and SA-A3C-RSA algorithms proposed in this thesis can effectively reduce the traffic blocking rate and spectrum fragmentation in optical network,and improve the utilization of spectrum resources.