次世代 6G 網路必須支援對低延遲和高傳輸速率的資料有著前所未有需 求的高階應用。然而，傳統基於人工智慧的資源管理難以適應動態流量 模式，尤其是在分佈外 (Out-of-Distribution, OOD) 環境中。本研究提出 了一個新穎的框架，該框架將元強化學習 (meta-RL) 與注意力機制相結 合，以實現自適應無線電資源管理 (RRM)。所提出的框架利用多頭自注 意力機制來增強在不同場景中的特徵辨識能力，並引入了觸發機制，以 便在遇到 OOD 條件時能及時進行微調。此模型在基於 5G QoS identifier (5QI) 規範的真實流量場景下進行了評估。結果顯示，與單純的元強化 學習和其他深度強化學習方法相比，結合自注意力的元強化學習提升在 不同任務下的推斷能力，而使用者滿意度提升了 3.8% 至 15.2%。此外， 觸發微調機制提供了高達 13.5% 的額外改進。這證明了該框架在確保跨 越不同和未見網路條件下穩健、服務品質 (QoS) 和資源管理方面的有效 性。 ;Next generation 6G networks must support advanced applications with unprecedented demands for low latency and high data rates. However, traditional radio resource management (RRM) methods based on artificial intelligence (AI) struggle to adapt to dynamic and unpredictable traffic patterns, especially in out-of-distribution (OOD) environments. This work presents a novel framework to address this by first introducing an attention-enhanced metareinforcement learning (meta-RL) approach. This method utilizes a multihead self-attention mechanism to improve feature identification and enhance the generalization of the model across diverse scenarios. Furthermore, to tackle OOD conditions, the framework incorporates a trigger mechanism that initiates timely fine-tuning, enabling the system to self-adapt and ensure robust performance when encountering diverse environments. The model is evaluated in realistic traffic scenarios based on the 5G QoS identifier (5QI) specifications. The results show that our attention-enhanced meta-RL improves user satisfaction by 3.8% over other meta-RL and 15.2% over other DRL methods, and the trigger mechanism provides a further improvement up to 13.5% after fine-tuning. This demonstrates the effectiveness of the framework in ensuring QoS-aware resource management across varied and unseen network conditions.
