Research on Resource Allocation under Unlicensed Spectrum Using Q-Learning
Uyoy Ial
TL;DR
The paper tackles resource allocation in mixed licensed/unlicensed spectrum under LAA by comparing a traditional MID channel selection with a Q-learning-based channel selection that uses a spectrum map to maximize throughput. It surveys Q-learning variations (DQN, Double DQN, Dueling DQN) and embeds them in a framework where LAA UEs and WiFi nodes contend for channels, evaluating throughput through a defined expression $Throughput = \frac{T_{max} \cdot \log_{10}(1 + \frac{Num}{Deno})}{I_{cca} + T_{max}}$. Empirical results show a persistent performance gap favoring MID under the tested neural architectures and training settings, with minimal gains from the Q-learning variants, suggesting the need for more expressive models (e.g., CNN or Transformer-based Q-learning) and careful architectural choices for large-scale, interference-limited wireless environments. The work informs spectrum-sharing design for 5G NR-U and highlights directions for reinforcement-learning-driven resource allocation in future wireless networks.
Abstract
In response to the advent of the 5G era, enhancing throughput and increasing transmission efficiency within limited spectrum resources is an important research topic. In the LTE system, utilizing unlicensed spectrum to assist traditional mobile networks, known as License Assisted Access, has emerged as a viable solution to effectively improve transmission efficiency. However, as the unlicensed spectrum also accommodates other users, such as Wi-Fi for mobile communication, there is a need to address the issue of spectrum resource allocation, aiming to achieve fair transmission among different mobile communication users. This research project aims to explore and compare two approaches: traditional communication algorithms and reinforcement learning method Q-leaning, under the condition of achieving maximum system throughput, in order to determine the differences between the two methods.