Multi-access edge computing provides proximate intelligent services for distributed users. Due to the user’s mobility and highly dynamic network, edge servers with limited coverage cannot ensure continuity of running services and maintain high-level Quality of Service. To tackle this issue, an effective service migration strategy is of paramount importance. However, the current approach ignores the cooperation between multiple edge servers and independent users. In this article, we study service migration with edge collaboration to realize lightweight migration by layer-sharing framework of containers, saving redundant transmissions of migration. Then, we formalize the migration decision problem as maximizing the migration utility problem. To obtain efficient online decisions, we proposed a dynamic service migration strategy (MA-DSM) based on multi-agent proximal policy optimization (MAPPO) algorithm, which leverages a flexible multi-policy framework to achieve user preference adaptation. Specifically, we improve the basic MAPPO by devising a context-aware grouping method to cluster agents with user’s mobility patterns and service preferences. Parameter sharing is introduced into the actor–critic network to learn customized policies for different clusters, facilitating cooperation among users in the same cluster. Extensive experiments demonstrate that our proposed approach outperforms baselines in terms of convergence, latency and migration utility.

中文翻译：

---

边缘协同服务迁移：结合用户偏好适应的多智能体深度强化学习方法


多接入边缘计算为分布式用户提供邻近的智能服务。由于用户的移动性和高度动态的网络，覆盖范围有限的边缘服务器无法确保运行服务的连续性和保持高水平的服务质量。为了解决这个问题，一个有效的服务迁移策略至关重要。然而，目前的方法忽略了多个边缘服务器和独立用户之间的合作。在本文中，我们研究了边缘协同的服务迁移，通过容器的分层共享框架实现轻量级迁移，节省了迁移的冗余传输。然后，我们将迁移决策问题正式化为最大化迁移实用程序问题。为了获得高效的在线决策，我们提出了一种基于多智能体近端策略优化 （MAPPO） 算法的动态服务迁移策略 （MA-DSM），该算法利用灵活的多策略框架实现用户偏好适应。具体来说，我们通过设计一种上下文感知分组方法来将代理与用户的移动模式和服务偏好进行集群，从而改进基本的 MAPPO。在 Actor-critic 网络中引入参数共享，学习不同集群的自定义策略，促进同一集群内用户之间的合作。广泛的实验表明，我们提出的方法在收敛性、延迟和迁移效用方面优于基线。