Edge Intelligent Multi-User Uplink Radio Access Methods for Beyond 6G Networks

Uplink radio access is a critical component of future 6G wireless networks, enabling simultaneous data transmissions from multiple users to access points (APs). Due to dynamic wireless environments, users must adaptively adjust modulation and coding schemes (MCS) and random access strategies. The proliferation of machine-type communication devices, characterized by sporadic traffic and limited resources, underscores the need for distributed, low-complexity algorithms to optimize uplink access.

This thesis develops distributed learning-driven solutions for these challenges in 6G uplink access, offering three main contributions. First, a federated learning method is proposed for adaptive orthogonal frequency division multiplexing with index modulation (OFDM-IM), utilizing k-means clustering. This approach aggregates the learning outcomes of local devices into a global model, reducing local training requirements and enhancing throughput while introducing reduced federation overhead. The effectiveness of this federated learning assisted adaptive modulation inherently relies on robust and efficient multiple access methods to manage the anticipated massive connectivity in future networks. Second, a learning-based two-step non-orthogonal random access (NORA) strategy is developed for massive connectivity, allowing users to independently select transmission slots and power levels without channel state information (CSI). The base station (BS) applies successive interference cancellation (SIC) to decode overlapping transmissions. This joint slot-power selection is modeled as a Markov decision process (MDP) and solved with tailored multi-state and confidence-aided Q-learning algorithms, significantly improving throughput and fairness, particularly under high congestion. Finally, the thesis presents a decentralized deep reinforcement learning-based NORA framework for multi-AP machine-type communications. Users autonomously select APs, transmission slots, and power levels without CSI, while APs decode packets using SIC. Custom deep $Q$-network algorithms address the AP-slot-power selection, showing substantial improvements in throughput, fairness, and scalability under diverse, high-traffic scenarios, emphasizing the framework’s potential for intelligent random access in 6G ecosystems.