基於機器學習的B5G無線資源管理之研究

The Study of Radio Resource Management Based on Machine Learning for B5G Mobile Networks

隨著第五代(5G)行動通訊技術的普及，無線通訊網路正在迅速發展，未來將會面對不斷增加的設備數量、更多元化的應用場景，因此B5G的無線通訊系統需要更靈活且智能的資源管理，以滿足不斷增長的多樣化應用需求。本研究提出一個融合機器學習（Machine Learning, ML）、網路切片（Network Slicing）與無線資源管理方法，稱為智能網路切片資源管理系統（Intelligent Network Slicing Resource Management System, INSRMS），以應對當今與未來動態且異構的無線網路環境。本研究旨在有效地優化無線資源的分配，並利用機器學習來更加精準的分類出切片需求的種類，並且依照切片種類給予優先級，針對切片優先級進行處理。資源管理架構的目標是讓優先級高的切片需求有更好的服務品質。
首先本研究將將針對三種機器學習演算法，分別為SVM（Support Vector Machine）、KNN（K Nearest Neighbor）、Random Forest進行網路切片之分類。研究結果顯示隨機森林在分類切片需求上有較好的準確度。接著針對不同佇列長度、不同資源預留區域與兩種不同切片需求到達率之場景探討無線資源配置的效能。研究結果顯示提出的INSRMS方法在不同切片種類的到達率增加時，最高優先級切片需求的平均等待時間趨近於0，且不會有丟失的狀況發生，可以保證最高優先級的切片需求相比其他種類的切片有更良好的服務品質。雖然第二優先級之切片平均等待時間最長，但不會有丟失的情況發生。第三優先級之切片的平均等待時間略低於第二優先級，但會因為最高優先級之切片需求增加導致丟失率也隨之升高。

This With the widespread adoption of fifth-generation (5G) mobile communication technology, wireless communication networks are rapidly evolving. In the future, these networks will face an increasing number of devices and more diverse application scenarios. Therefore, B5G wireless communication systems need more flexible and intelligent resource management to meet the growing and diverse application demands. This research proposes an Intelligent Network Slicing Resource Management System (INSRMS) that integrates Machine Learning (ML), Network Slicing, and radio resource management methods to address the dynamic and heterogeneous wireless network environments of today and the future. The aim of this research is to effectively optimize radio resource allocation and use machine learning to accurately classify slice demand types. Based on the type of slices, priorities will be assigned and managed accordingly. The goal of the resource management framework is to ensure that higher-priority slice demands will receive better quality of service.
We will first classify network slice types using three machine learning algorithms: Support Vector Machine (SVM), K Nearest Neighbor (KNN), and Random Forest. The results show that Random Forest has better accuracy in classifying slice demands. Then, we investigate the performance of radio resource allocation under different queue lengths, resource reservation sizes, and two scenarios with different slice demand arrival rates. The results show that the INSRMS method can ensures the average waiting time of the highest priority slice approaching zero as the arrival rate of different slice types increases. Besides, the highest priority slice has no loss occurred. Although the average waiting time of the second priority slice is the longest, it has also no loss occurred. The average waiting time for the third priority slice is slightly lower than that of the second priority, the loss rate increases as the arriving rate of highest priority slice increases.