User-Centric Bandwidth Allocation in Network Slicing via Dueling DRL

Network slicing plays a pivotal role in 5 G and next-generation communication systems to support diversified service requirements. This paper investigates bandwidth allocation in a multi-base station, multi-user network where users probabilistically request network slicing services. We formulate a dynamic resource optimization problem to maximize user satisfaction, accounting for stochastic service demands and heterogeneous quality-of-service requirements. An Adaptive Hierarchical Dueling Deep Q-Network (AH-Dueling DQN) algorithm is proposed to address the NP-hard mixed-integer programming nature of this problem, leveraging deep reinforcement learning to handle high-dimensional state-action spaces. Simulation results demonstrate the algorithm’s effectiveness in reducing users’ latency while improving the throughput.