Reinforcement Learning-Based Resource Reservation for Mobile Edge Computing With Probable Failure

Mobile edge computing (MEC) has increasingly been propelled into the spotlight as a promising paradigm to address the growing demand for latency-sensitive mobile applications. The practical challenges associated with deploying and maintaining traditional edge service devices highlight the attraction of leveraging end devices as edge servers. Yet, the inherent mobility and power constraints of such servers make them prone to failure. As an effective strategy, some edge servers are earmarked to mitigate probable failure, which may cause resource redundancy. In this paper, a novel partial resource reservation scheme is proposed to diminish this waste, each edge server in system reserves a part of their resources to defend probable failure. The approach directs reservation actions towards the individual resources, rather than the entirety of edge servers. Given that edge servers serve as key participants within the system, it is imperative to optimize overall income by carefully balancing incurred costs and generated benefits. We investigate this partial resource reservation problem from the perspective of edge servers, and transform it into Markov decision process (MDP). A scheme that fomulates strategies for edge servers then been proposed, called Q-learning-based resource reservation (QLRR), which can accommodate the dynamic nature of edge server failure. Empirical evaluations demonstrate that our proposed scheme outperforms existing benchmark schemes, significantly enhancing the income of edge servers.