Resource slicing and customization in RAN with dueling deep Q-Network

Abstract

The emerging future generation 5G technology is expected to support service-oriented virtualized networks where different network applications provide unique services. 5G networks have the potential to allow completely different slices to co-exist in a substrate network and satisfy the differentiated requirements of various users. In networks with heterogeneous traffics, operators are required to provide services in isolation since each operator has its own defined performance requirements. However, achieving an efficient resource provisioning mechanism for such traffics is very challenging. This paper proposes a coarse resource provisioning scheme and a dynamic resource slicing refinement scheme based on dueling deep reinforcement learning for virtualized radio access network. Firstly, coarse resource provisioning scheme provisions and allocates radio resource to slices based on preferences and weights at different base stations. Secondly, reinforcement learning based slicing refinement adjusts the resource allocated to slices autonomously in order to balance satisfaction and resource utilization. The proposed dueling DQN algorithm unifies two objectives (QoS satisfaction and resource utilization) by weights to indicate the importance of each factor in the reward function. After the dueling DQN algorithm has output actions to provision resource at slice level, BS-level resource update is performed. Also, a common learning agent is used to control the activities of all the slices in the network. Then, a shape-based resource allocation algorithm is proposed to customize the diverse requirements of users to improve user satisfaction and resource utilization. Finally, a comprehensive performance evaluation is conducted against state-of-the-art solutions based on OFDMA air-interface design. The results reveal that the proposed algorithm balances satisfaction and resource utilization with 80% of the available resources. The algorithm also provides performance isolation such that, a sudden change in user population in one slice does not affect the others.