Predict Critical Node Pairs Based on Reinforcement Learning in SDN

Recognizing the node pairs that contribute more packets and bytes compared to others in a Software-Defined Network is essential for many network management tasks, such as traffic engineering, flow-rule placement, anomaly detection. However, recognizing these critical node pairs poses significant challenges because of the enormous solution space for critical node pairs selection and the changing distribution of network traffic. Therefore, it is not reasonable to design heuristic algorithms based on fixed rules to solve this problem because they can not adapt to the changes of the network. In this paper, we propose a reinforcement learning-based framework that a) persistently collects raw network traffic and network configuration information, b) constructs a changing environment and generates reward signals according to the collected data, c) trains the agent periodically to predict critical node pairs dynamically. Our preliminary experiment results show that the proposed framework can predict critical node pairs effectively.