COMA-FNN: Fuzzy Reinforcement Learning for DUDe-Aware Handover

The deployment of high-frequency carriers in 5G networks and beyond introduces significant challenges, particularly due to increased path loss and physical limitations of user equipment (UE). In particular, these challenges lead to uplink/downlink coverage imbalances, critically affecting applications that rely on robust uplink capacity such as autonomous driving, telemedicine, and live streaming. While downlink-uplink decoupling (DUDe) and supplementary uplink (SUL) have emerged as promising solutions for uplink enhancement, their flexible cell association mechanisms disrupt conventional handover strategies by introducing multidimensional, decoupled decision-making. To address this, we propose a heterogeneous network architecture that integrates DUDe into SUL, supporting non-co-located deployment of supplementary uplink carriers and enabling dual decoupling at both the base station and carrier levels. Building on this architecture, we introduce Counterfactual Multi-agent fuzzy neural network (COMA-FNN), a multi-agent reinforcement learning (MARL) algorithm based on centralized training with decentralized execution (CTDE). COMA-FNN incorporates fuzzy neural networks to model the complex, nonlinear relationships among multiple communication attributes, improving the precision and adaptability of handover decisions. The algorithm also employs a centralized critic with counterfactual baselines to effectively resolve credit assignment issues among competing agents. To accommodate varying service requirements, COMA-FNN incorporates service-specific reward functions aligned with four 3GPP-standardized service types. These rewards are weighted using the Analytic Hierarchy Process (AHP), enabling the algorithm to support different QoS policies. Simulation results demonstrate that COMA-FNN significantly improves handover efficiency, reduces latency, and enhances throughput, making it a robust solution for intelligent mobility management in decoupled uplink/downlink architectures.