Reinforcement Learning Based Routing in Software Defined Network

Abstract

The traditional routing algorithm employs distributed approach for path updates, where the convergence time is higher. The networking systems complexity is growing due to increasing cloud services, Internet of Things (IoT) applications, video streaming, and mobility enabled to the internet users. The drawback is the varying QoS requirements by the internet users are not meet. Machine learning enables the embedding of artificial intelligence into the system to handle complex networks. The traffic dynamics, complex network stacks, and mobility demand network programmability to be augmented into the traditional approach. In this paper, reinforcement learning (RL) algorithm for routing in Software Defined Networks (SDN) is proposed. The link-state parameters are used for making decision in routing by capturing network environment updates dynamically. The SDN platform is used to obtain the global view of the network and install the optimal paths in the switch/routing devices. The challenge is collecting the parameters for link-state changes in the network dynamically and calculating the reward value for decision making. The emulation is performed using real-time network topologies. The results obtained are promising compared to the traditional link-state algorithm. The RL algorithm to determine path enables the learning agent to learn the route over time (training time). In the testing phase time taken to find the path is n-1 comparisons for single source and destination pair where ‘n’ is the number of forwarding devices in the topology. The Q-learning algorithm is applied to select the best policy for finding the path using the learning agent and network environment specific Q-values.