5G 網路架構上，異質網路 (heterogeneous networks, HetNets)，或稱作小型基地台網路 (small cell networks) 是被廣泛討論的 5G 技術，而且如何分流 (offloading) 是一項重要議題。在支援雲端控制的架構下，未來的資源分配需要主動預測問題，採取先發製人的措施，在尚未造成效能的減損之前，做出決策。而隨著 5G 演進大量導入的新技術大大增加了資源分配的複雜度，卻也造成了以機器學習為基礎的功能展露頭角的機會。各種機器學習方法中，又以深度增強式學習 (deep reinforcement learning, DRL) 最具前瞻與網路控制問題上的實用性。本計畫將聚焦在未來 5G 技術下，延續執行中計劃對於流量預測的做法，深入研究 DRL 在 HetNets 高能源效益分流問題的應用效果，並開放原始碼以供驗證，經由提出領先且可行之無線資源管理方法，做為進一步研究 5G 網路問題之重要參考。 ;In 5G networks, heterogeneous networks (HetNets) with small cells is a promising architecture to be deployed. The traffic offloading among macro and small cells is inevitably a key issue. Based on the cloud controlling structure, it is possible to design proactive strategies, so operation issues can be predicted and treated before suffering performance degradation. At the same time, the much more complex nature of 5G resource management is happen to be a suitable target to apply advanced machine learning approaches. In the project, we propose to apply deep reinforcement learning (DRL) on energy-efficient mobile traffic offloading. Taking advantage of our traffic forecasting works, we will investigate the DRL model for 5G networking issues, and further provide suggestion for future 5G resource management works.
