Intelligent congestion control in 5G URLLC Software-Defined Networks using adaptive resource management via Reinforced Dueling Deep Q-Networks

Abstract

Centralized control of Software Defined Networking (SDN) yields efficient management of network resources and offers a global perspective. However, centralized controllers have many performance and scalability issues, particularly given the rapid expansion of 5G connectivity. The latest demands on the transport network come from areas such as increasing RAN and mobile broadband service capacity, new 5G-enabled services and the dynamic deployment flexibility of the 5G Radio Access Network (RAN) split architecture, with its tight transport characteristics. These characteristics are particularly evident in the fronthaul segment of RAN, where latency and synchronization requirements pose significant challenges. Enhanced automation capabilities in the operations and management domain represent a key requirement to meet these challenges. Traditional machine learning (ML) techniques, which concentrate the training data and carry out sequential model learning over a sizable data set, are the main emphasis of current wireless network learning approaches. However, using a huge dataset for training is inefficient since it takes a lot of time and does not use resources or energy efficiently. Hence, this work focuses on Reinforced Dueling Deep Q-Network (RDDQN), a revolutionary approach to network slicing design for load prediction and resource management in data-driven workflows. Moreover, it can reduce congestion by adopting an Ultimatum queuing game theory-based scheduling mechanism in the controller. The proposed RDDQN achieves an average throughput of 579.34 kbps, an execution time of 12.57 s, goodput fairness of 94.56%, and delay fairness of 10.37 s across various parameters.