Dynamic Routing in SDN Using Multi-Agent Deep Deterministic Policy Gradients

J. Delshi Julie, Dr. R. Beulah

Abstract: The separation of the control plane (CP) from the data plane (DP) facilitates in Software Defined Networking (SDN) centralizing the network management. SDN is an innovative network design. This separation will help in the configuration of the network, management of the network and optimization of the network. This separation also contributes more programmable and adaptable control over the network activities. Here, adjusting to high-dimensional (HD) state-action (S-A) spaces and quickly shifting network conditions are not contributed by the conventional Routing Protocols (RP). For the purpose of overcoming those limitations in SDN network routing, the (MA) Multi-Agent (DDPG) Deep Deterministic Policy Gradient (MADDPG) algorithm was suggested in this study. Decentralized agents were enabled by this MADDPG, and network nodes are stimulated by these agents for learning the Optimal Routing Policies (ORP) collaboratively. During training, the global network state is considered. Then, a scalable, and adaptive Decision Making (DM) was facilitated by this approach, as it integrates centralized training and decentralized execution. Important objectives are optimized by this model, as it attains minimizing latency, balancing network loads, and maximizing throughput. Adapting to dynamic traffic patterns and faults were facilitated by the MADDPG-based routing with the Reinforcement Learning (RL). This integration will also support in ensuring the robust and Real-Time (RT) operation. Simulation was conducted, and the outcomes of the simulation indicates that the suggested MADDPG performs better than the current RP by delay, Packet Loss (PL), and throughput (T). MADDPG became an effective method for future SDN settings.

Keywords: Software Defined Networking (SDN), Network routing, Optimal Routing Policies (ORP), Multi-Agent Deep Deterministic Policy Gradient (MADDPG) procedure