Resource allocation policies for next generation networks using reinforcement learning and large language models

Abstract

The rapid growth of next generation wireless communication systems has significantly increased the complexity of resource management in Radio Access Networks (RANs). Modern 5G and future 6G networks require intelligent and adaptive resource allocation mechanisms capable of handling dynamic traffic conditions while satisfying diverse Quality of Service (QoS) requirements. Conventional static and heuristic based approaches often fail to efficiently manage varying network demands, motivating the use of Artificial Intelligence and Reinforcement Learning (RL) based techniques for autonomous wireless network optimization. This thesis presents an intelligent resource allocation framework for next generation networks using Reinforcement Learning and Large Language Models (LLMs). Initially, a Knowledge Based Reinforcement Learning (KBRL) framework is studied for SLA aware network slicing and Physical Resource Block (PRB) allocation under multiple traffic scenarios. The KBRL framework enables adaptive allocation decisions while balancing resource utilization and service quality constraints. To improve scalability and decision interpretability, the work further extends the conventional KBRL framework using Large Language Models for policy learning and reasoning aware resource allocation. Runtime trajectories and allocation decisions generated by the KBRL framework are utilized to construct a structured dataset for supervised learning and LLM fine tuning. The developed LLM based framework learns state to allocation mappings and generates allocation decisions across different network conditions while approximating the behaviour of the expert KBRL policy. The implementation includes scenario generation, KPI based state representation, supervised policy learning, and comparative performance evaluation using metrics such as PRB utilization, SLA violation behaviour, and prediction consistency.