| 摘要: | 隨著物聯網（IoT）的快速發展，對計算敏感的終端設備數量顯著增加。透過將這
些設備的計算任務卸載到邊緣伺服器，邊緣計算在減少任務延遲和減輕雲端伺服器計
算負擔方面顯示出其效益。然而，無策略地卸載計算任務可能導致邊緣伺服器的資源
低效利用，引發延遲增加和計算成本上升。因此，設計有效的任務卸載和資源配置策
略，以優化延時和能耗，是目前研究的重點和挑戰。減少雲伺服器的任務延遲和計算
負擔已成為重要議題。邊緣伺服器在缺乏適當激勵機制下可能不願分享資源，因此提
供適當的獎勵至關重要。考慮到隱私洩露風險，移動用戶可能不願意透露私有資訊，
這導致雲平台和邊緣伺服器間的資訊不對稱。先前研究通常假設雲平台能完全掌握邊
緣伺服器資訊，但在實際中並非如此。本文提出了一種基於深度強化學習（DRL）的
契約激勵機制。與傳統方法不同，DRL 能在不需要預先了解環境詳情的情況下運作。
DRL 通過學習和適應設計激勵機制，使得在動態和不確定環境中有效地激勵參與者完
成任務，達到雲平台的最大效用。本文的貢獻包括在資訊不對稱情景下提出聯合資源
分配和計算卸載激勵問題，系統分析最優契約的充要條件，並將契約激勵問題制定為
不完全資訊情景下的馬可夫決策過程，設計深度確定性策略梯度（DDPG）方法，以在
高維度的動作和狀態空間下獲得計算資源和激勵報酬策略。;With the rapid development of the Internet of Things (IoT), the number of computation-
sensitive end devices has significantly increased. By offloading the computational tasks
of these devices to edge servers, edge computing has demonstrated its benefits in reduc-
ing task latency and alleviating the computational burden on cloud servers. However,
indiscriminately offloading computational tasks may lead to inefficient use of edge server
resources, resulting in increased latency and higher computational costs. Therefore, de-
signing effective task offloading and resource allocation strategies to optimize latency and
energy consumption is currently a key research focus and challenge. Reducing the
task latency and computational burden on cloud servers has become an important is-
sue. Without appropriate incentive mechanisms, edge servers may be unwilling to share
resources, making the provision of suitable rewards crucial. Traditional incentive mech-
anisms, such as auction theory and Stackelberg games, rely on frequent information ex-
change, leading to high signaling costs. Considering the risk of privacy leaks, mobile
users may be reluctant to disclose private information, resulting in information asymme-
try between cloud platforms and edge servers. Previous research often assumed that cloud
platforms have complete information about edge servers, which is not the case in prac-
tice. This paper proposes a contract incentive mechanism based on deep reinforcement
learning (DRL). Unlike traditional methods, DRL can operate without prior knowledge
of the environment’s details. DRL learns and adapts to design incentive mechanisms,
effectively motivating participants to complete tasks in dynamic and uncertain environ-
ments, achieving the maximum utility of the cloud platform. The contributions of this
paper include proposing the joint resource allocation and computation offloading incen-
tive problem under information asymmetry, systematically analyzing the necessary and
sufficient conditions for optimal contracts, formulating the contract incentive problem
as a Markov decision process under incomplete information, and designing a deep deter-
ministic policy gradient (DDPG) method to obtain computation resource and incentive
reward strategies in high-dimensional action and state spaces. |
