基于预测与强化学习的5G混合切片资源优化分配

PDF下载HTML

Title:

Prediction and reinforcement learning-based optimized resource allocation for 5G hybrid network slicing

作者:

陈智雄^1,2,3, 谢宇鹏¹, 郭以贺¹

1. 华北电力大学 电子与通信工程系, 河北 保定 071003;
2. 华北电力大学 河北省电力物联网技术重点实验室, 河北 保定 071003;
3. 华北电力大学 电力物联智慧化技术河北省工程研究中心, 河北 保定 071003

Author(s):

CHEN Zhixiong^1,2,3, XIE Yupeng¹, GUO Yihe¹

1. Department of Electronic and Communication Engineering, North China Electric Power University, Baoding 071003, China;
2. Hebei Key Laboratory of Power Internet of Things Technology, North China Electric Power University, Baoding 071003, China;
3. Hebei Engineering Research Center of Intelligent Technology for Power Internet of Things, North China Electric Power University, Baoding 071003, China

关键词:

5G混合场景; 切片; 空口资源分配; 服务水平协议; 强化学习; 预测; 随机启动概率门限; 学习频次

Keywords:

5G hybrid scenarios; network slice; radio access resource allocation; service-level agreement; reinforcement learning; predictive; randomized initialization probabilistic threshold; learning frequency

分类号:

TP393；TN929.5

DOI:

10.11992/tis.202506012

文献标志码:

2026-3-5

摘要:

针对现有算法在切片资源合理分配和服务水平协议(service-level agreement, SLA)保障中的学习频次过高、计算复杂度较大等问题，提出了一种融合时间序列预测与强化学习的智能时频资源分配策略。考虑SLA违反概率约束和5G时频资源等约束，以时频资源使用数量最小化为目标，建立了5G混合场景的切片资源优化分配模型。首先利用堆叠长短期记忆网络(stacked long short-term memory, Stacked-LSTM)对eMBB、URLLC、mMTC场景的信噪比、排队缓冲区容量、时延、设备连接数量的SLA关键指标进行预测；定义了高斯核强化学习的状态、动作和奖励函数，并利用SLA指标预测值和随机启动概率门限执行学习算法获得准最佳分配策略。相比于现有算法，仿真结果表明所提算法可以在保障SLA违背概率约束和时频资源数量等基本性能的条件下，在预测精度、收敛速度和资源分配等综合性能上表现较好，有效地减少机器学习的频次和复杂度。

Abstract:

To address the high learning frequency and high computational complexity of existing algorithms for rational slice resource allocation and service-level agreement (SLA) assurance, this paper proposes an intelligent time-frequency resource allocation strategy that integrates time series prediction with reinforcement learning. Under constraints on SLA violation probability and 5G resources, an optimization model for slice resource allocation in hybrid 5G scenarios is established, with the objective of minimizing the time-frequency resource usage. First, a stacked long short-term memory network is employed to predict key SLA indicators such as signal-to-noise ratio, queue buffer occupancy, latency, and the number of connected devices for enhanced mobile broadband, ultra-reliable low-latency communications, and massive machine-type communications scenarios. Next, the states, actions, and reward function of the Gaussian-kernel-based reinforcement learning algorithm are defined. The learning process is guided by the predicted SLA indicators and a randomly initialized probabilistic gating threshold to derive a near-optimal resource allocation policy. Simulation results demonstrate that, compared with existing algorithms, the proposed approach achieves better overall performance in prediction accuracy, convergence speed, and resource allocation efficiency while ensuring basic performance metrics such as SLA violation probability and resource usage. It also reduces the frequency and complexity of machine learning operations.