MACHINE LEARNING-BASED TECHNIQUE FOR SPECTRUM ALLOCATION IN 5G NETWORKS

Abstract

The exploitative increase in mobile data traffic and the spread of connected devices has increased the pressure on effective spectrum allocation in Fifth-Generation (5G) networks. Conventional, fixed spectrum management strategies are becoming less effective in response to the very dynamic and heterogenous nature of 5G conditions, which commonly results in under-utilization of spectrum, enhanced levels of interference, and diminished quality of service. This study proposes a machine learning-based spectrum allocation technique using a Deep Q-Network (DQN) reinforcement learning model to address these challenges. A publicly available dataset of 5G spectrum usage on the Kaggle platform consisting of 50, 000 records was used and includes such key parameters as Signal-to-Noise Ratio (SNR), Channel State Information (CSI), User Equipment (UE) density, bandwidth usage, traffic demand, and level of interference. The DQN agent was trained to make real-time, context aware spectrum allocation decisions, depending on the current network states, where spectral efficiency, throughput, latency, interference as well as fairness index were used as a metric to assess performance. The experimental findings showed that the proposed model attained spectral efficiency of 4.85bits/s/Hz, 980 Mbps throughput on average, 11ms latency, 8.2dB interference, and a fairness index of 0.93, which was superior in all metrics compared to the traditional methods of statical allocation. These results indicate that reinforcement learning can be effectively used in dynamic spectrum management in 5G networks, which provides a scalable and adaptable solution to the improvement of network functionality and the utilization of resources in next-generation wireless networks.