低軌道衛星網絡中基於深度強化學習的衛星輔助群眾外包直播

Abstract

在眾包直播中，觀眾端的觀看環境和對使用者體驗和偏好有異質性，導致如何符合成本效益地優化觀眾使用者體驗成了前所未有的挑戰。透過低軌衛星進行傳輸有望成為一個解決方案。然而，衛星的移動性、計算資源以及能源限制也成為了一大考驗。因此，本文提出資源分配的解決方案。基於直播主、衛星網路和觀眾的大量實時信息做出智能決策。而考慮到在這種背景下的過高計算複雜性，我們提出搭配深度強化學習的解決方案，能夠自動學習最合適的觀眾調度和轉碼選擇策略。

In crowdsourced live streaming, the diverse viewing environments and user experience preferences present significant challenges in cost-effectively optimizing audience experience. Transmission via low Earth orbit (LEO) satellites shows promise but also faces challenges due to mobility, computing resources, and energy constraints. This paper proposes a resource allocation solution. By leveraging real-time information from broadcasters, satellite networks, and viewers, we aim to meet user experience demands while minimizing system costs. Given the high computational complexity, we incorporate deep reinforcement learning (DRL) to automatically learn optimal strategies for viewer scheduling and transcoding selection.