Ai-enhanced dynamic preemptive resource allocation in next generation cellular networks

Abstract

The rapid expansion of diverse slices and applications in 5G and beyond (5GB) communication systems has necessitated the fulfillment of stringent key performance indicators (KPIs), including higher data rates, reduced latency, improved spectral efficiency, enhanced energy efficiency, and expanded network capacity. Each service category is defined to satisfy one or more different KPIs. These specific needs are addressed with network slices, which divide network resources into dedicated slices. Likewise, Radio Access Network (RAN) slicing plays a crucial role in managing resources at the radio access level, enabling the coexistence and efficient sharing of resources across multiple slices. The individual needs and demands of the two primary slices, namely, Enhanced Mobile Broadband (eMBB) and Ultra-Reliable Low-Latency Communication (URLLC), create a complex optimization problem where traditional scheduling approaches often fail to achieve satisfactory performance for both traffic types simultaneously. In response to this challenge, this thesis adopts a systematic, multi-phase approach to enhance scheduling performance incrementally. Initially, a heuristic scheduling approach is proposed to establish a baseline and facilitate easy integration with existing systems. This approach utilizes the allowed latency of URLLC slices to optimize concurrent scheduling by mitigating the adverse impact of preemption on the perceived quality of the eMBB slices. The Patient and Intelligent Preemption (PIPE) algorithm demonstrates how strategic delaying of URLLC scheduling decisions can minimize unnecessary preemption of eMBB resources while still meeting strict URLLC deadlines. The algorithm introduces a novel preemption allowance rate that defines how many times an eMBB codeblock can be preempted before requiring retransmission, effectively creating a buffer against throughput degradation. Evaluation through link-level simulations shows PIPE's ability to maintain eMBB throughput within 5\% degradation while fully satisfying URLLC latency constraints, outperforming traditional scheduling methods like aggressive and threshold-proportional approaches. Building upon the heuristic foundation, the research progresses to develop a 3GPP-compliant preemption mechanism that incorporates all aspects of the radio access network. This enhanced approach introduces selective preemption targeting low-priority eMBB resources, such as those using lower modulation orders or having higher error-correction redundancy. The mechanism includes comprehensive modeling of preemption indication signaling, which enables efficient buffer management at user equipment by providing timely notifications about punctured resources. The system-level implementation spans the entire protocol stack from physical layer processing to MAC scheduling and HARQ operations, ensuring realistic performance evaluation under diverse network conditions. Simulation results demonstrate the mechanism's ability to maintain URLLC latency below 1ms while minimizing eMBB throughput degradation, even under high URLLC traffic loads. The solution maintains low computational complexity and requires minimal control channel overhead, making it practical for real-world deployments in dense network scenarios. The final phase of the research introduces an artificial intelligence-driven approach using Deep Reinforcement Learning (DRL) to further optimize the preemption decisions. The Context-Aware Multi-DQN Scheduling (CA-M-DQN) framework represents a significant advancement over static scheduling algorithms by employing multiple Deep Q-Network models, each specialized for different Channel Quality Indicator (CQI) levels and URLLC packet sizes. This modular design allows the system to adapt dynamically to varying network conditions and traffic patterns. The framework incorporates a carefully designed hybrid action space with 28 discrete preemption strategies that comply with 3GPP mini-slot configurations, along with a multi-objective reward function that balances URLLC latency compliance against eMBB throughput preservation. System-level validation shows that CA-M-DQN achieves superior eMBB throughput compared to heuristic baselines while maintaining URLLC reliability with block error rates below 1\%, meeting the most stringent 3GPP requirements. The success of this approach lies in its ability to learn optimal policies through interactions with the network environment, continuously improving its decision-making capabilities without requiring explicit mathematical models of the underlying system dynamics. The practical implications of this research extend across various deployment scenarios in 5GB networks. The proposed solutions demonstrate scalability for high-density urban environments and industrial IoT applications where diverse traffic types must coexist. Energy efficiency improvements are achieved through reduced retransmissions and optimized resource allocation, contributing to more sustainable network operations. The strict adherence to 3GPP standards ensures compatibility with existing infrastructure, lowering barriers to adoption for network operators. Furthermore, the frameworks' modular designs provide flexibility to accommodate future network enhancements and new service categories beyond eMBB and URLLC, such as massive Machine-Type Communications (mMTC) and extreme URLLC (xURLLC). Looking ahead, several promising directions emerge for extending this research. Multi-Agent Reinforcement Learning (MARL) could address the challenges of inter-cell interference in multi-cell deployments by enabling coordinated resource allocation across multiple base stations. Hardware validation through software-defined radio (SDR) testbeds would provide valuable insights into real-world performance characteristics and implementation challenges. The integration of predictive capabilities using advanced machine learning techniques could further enhance the system's ability to anticipate traffic patterns and optimize resource allocation proactively. Additionally, the framework could be extended to support emerging use cases in 6G networks, such as holographic communications and digital twins, where even more demanding latency and reliability requirements are expected. This thesis makes significant contributions to the field of wireless communications by developing practical solutions to the complex problem of resource allocation in heterogeneous 5GB networks. The systematic progression from heuristic methods to AI-driven approaches demonstrates how incremental enhancements can lead to substantial performance improvements. By combining theoretical rigor with practical implementation considerations, the research bridges the gap between academic concepts and real-world deployment requirements. The proposed frameworks establish a solid foundation for future innovations in intelligent radio resource management, paving the way for self-optimizing networks that can adapt to the ever-evolving demands of next-generation wireless applications and services. The integration of multi-layer preemption mechanisms with advanced reinforcement learning techniques represents a paradigm shift in how network slicing can be managed, offering unprecedented flexibility and efficiency in resource utilization while maintaining strict quality-of-service guarantees for diverse traffic types.