Research on Traffic-aware Intelligent Slicing Resource Allocation for Radio Access Network

‍ ‍Dynamic network resource allocation is crucial in a radio access network （RAN） to meet the requirements of service-level agreements （SLAs）， while simultaneously ensuring that the different service-quality requirements of different applications are effectively met. This paper proposes an intelligent bandwidth-allocation strategy that combines time-series prediction and deep reinforcement learning to address the scenarios of network traffic fluctuation and rapid network state change by reasonably using a long short-term memory （LSTM） network and dueling deep Q network （Dueling DQN） to maximize the spectral efficiency and satisfy the SLA of radio access network slices. By using LSTM networks to predict traffic in slices， the computational cycle of the deep reinforcement learning algorithm can effectively be decoupled from the actual slice configuration cycle. In order to reduce the computational complexity of the LSTM while maintaining its performance to fit the limited computational resources of a RAN， a randomly connected LSTM （RCLSTM） network with a controlled neuronal connectivity ratio is used. In addition， the Dueling DQN can improve the accuracy of the Q-value estimation for the slicing strategy learning process compared to a conventional DQN to enhance the convergence speed. Simulation results showed that the proposed RCLSTM-Dueling DQN scheme could effectively reduce the impact of network-environment fluctuations on wireless slicing resource management in dense traffic scenarios by sensing network performance changes in advance， compared to the original DQN， advantage actor-critic （A2C）， and hard slicing approaches. In RAN slicing scenarios with three different traffic fluctuation patterns and three different quality-of-service requirements， significant improvements in the convergence speed， spectral efficiency， and slice SLA satisfaction rates could be achieved. Also， a 10% connected RCLSTM network could reduce the computation time of the original LSTM by approximately 11% while maintaining a very low performance loss.