The exponential growth of data generated at the network edge has driven a paradigm shift from centralized cloud computing to local edge processing, accelerating the widespread adoption of edge computing across diverse applications. In this context, continual test-time adaptation (CTTA) on edge devices, which enables models to adapt to evolving target domains without access to source data or labeled samples, has become an emerging research focus due to its practical importance in dynamic environments with changing data distributions. However, limited computational and memory resources severely restrict CTTA on edge devices. Moreover, since adaptation relies on noisy unsupervised losses without access to labels, prolonged CTTA can lead to error accumulation. Additionally, the model is susceptible to catastrophic forgetting, an intrinsic challenge in continual adaptation. In this paper, we propose MccTTA, a memory efficient collaborative continual test-time adaptation framework for edge devices. Specifically, MccTTA incorporates a generative model to synthesize images as a replacement for replay data on the cloud, and a lightweight side network attached to the frozen original network to reduce memory consumption during edge adaptation. We further introduce 2SR (Two-Stage Rehearsal), which decouples active forgetting and knowledge integration into two separate stages to address the plasticity–stability dilemma caused by distributional discrepancies between synthetic and real task data during continual adaptation. Finally, extensive experiments are conducted to evaluate the effectiveness of MccTTA. The results show that, compared with conventional TTA methods, MccTTA achieves superior accuracy and mitigates forgetting while requiring less memory.

Deep neural networks for SAR image recognition often require compression for deployment on remote sensing platforms with limited computational and storage resources. Knowledge distillation (KD) is a key approach to improving the accuracy of lightweight networks. However, existing KD methods face challenges when applied to SAR images due to the small dataset size and the high noise in SAR images. To address this, this paper proposes a novel knowledge distillation method that relaxes the requirement for a strict linear relationship between the outputs of lightweight and large models, focusing instead on maintaining a Monotonic Relationship (MRKD). This reduces the difficulty of the KD task. Experiments on various SAR image classification and object detection datasets demonstrate that MRKD achieves state-of-the-art performance improvements for lightweight networks.

With the widespread adoption of the Internet of Things (IoT) and various smart medical devices, the volume of medical data has dramatically increased, making the processing of medical Internet of Things (IoMT) data increasingly challenging. Due to the integration of edge computing and cloud computing, IoMT can allocate increased computing and storage resources in proximity to the terminal, addressing the low-latency requirements of computationally intensive tasks. While existing initiatives have shifted services to edge servers, they have not taken into account the joint impact of task priorities and mobile computing services on Mobile Edge Computing (MEC) networks. Fortunately, the rapidly advancing field of Artificial Intelligence (AI) has proven effective in some resource allocation applications in recent years. In this article, we propose a mobile edge computing-based intelligent healthcare multitasking processing system aimed at addressing the issue of service prioritization in medical scenarios. Considering energy consumption and latency, we present a multi-objective task-aware service offloading algorithm under the framework of end-edge-cloud collaborative IoMT systems, employing deep deterministic policy gradients (DDPG). Adaptability to the diversity of different services is achieved through dynamic adjustments based on various business types and system requirements. Finally, the effectiveness of DDPG for IoMT is validated using real-world data.

Recently, computation offloading methods have greatly improved the Quality of Experience (QoE) in Multi-access Edge Computing (MEC) by offloading tasks to the edge servers. Since well-coordinated actions of Terminal Devices (TDs) are critical to improving the performance of the entire individual system, many practical MEC-based applications, i.e., firefighting robots and unmanned aerial vehicles, require great teamwork among TDs. However, real-world scenarios are usually bound by resource conditions. For instance, network connectivity may weaken or experience interruptions during emergency situations. In cases where the communication medium is utilized by multiple TDs, achieving effective coordination poses a significant challenge. In this paper, we propose a computation offloading scheme based on Scheduled Multi-agent Deep Reinforcement Learning (SMDRL) to make the most efficient decision in a resource-constrained scenario. First, we design a virtual energy queue based on the MEC system and maximize the QoE (related to service delay and energy consumption) in a real-time manner. Subsequently, we propose a scheduled multi-agent deep reinforcement learning algorithm to support each TD in learning how to encode messages, select actions, and schedule itself based on the received messages. Furthermore, a TopK mechanism is introduced. This mechanism chooses the most crucial TDs to broadcast their messages, and then the computation offloading problem in a communication-constrained MEC environment can be solved in a low-communication manner. Also, we prove that even under limited communication conditions, our proposed methods can still lead to the close-to-optimal performance. The final performance analysis shows that the developed scheme has significant advantages over other representative schemes.

Multiaccess edge computing (MEC) provides task offloading services to facilitate the integration of idle resources with the network and bring cloud services closer to the end user. By selecting suitable servers and properly managing resources, task offloading can reduce task completion latency while maintaining the Quality of Service (QoS). Prior research, however, has primarily focused on tasks with strict time constraints, ignoring the possibility that tasks with soft constraints may exceed the bound limits and failing to analyze this complex task constraint issue. Furthermore, considering additional constraint features makes convergent optimization algorithms challenging when dealing with such complex and high-dimensional situations. In this article, we propose a new computational offloading decision framework by minimizing the long-term payment of computational tasks with mixed bound constraints. In addition, redundant experiences are gotten rid of before the training of the algorithm. The most advantageous transitions in the experience pool are used for training in order to improve the learning efficiency and convergence speed of the algorithm as well as increase the accuracy of offloading decisions. The findings of our experiments indicate that the method we have presented is capable of achieving fast convergence rates while also reducing sample redundancy.

Cross-silo federated learning (FL) is a privacy-preserving distributed machine learning where organizations acting as clients cooperatively train a global model without uploading their raw local data. Recently, the cross-silo FL in multiaccess edge computing (MEC) is used in increasing industrial applications. Most existing research on cross-silo FL pays attention to the performance aspect, ignoring the incentive mechanism for high-quality client selection and long participation in model training for efficient and stable FL, which has prevented the widespread adoption of cross-silo FL in MEC. In this article, we propose an incentive mechanism with quality-Aware and reputation-Aware based on the infinitely repeated game for cross-silo FL named VARF. VARF selects high-quality and high-reputation edge nodes (ENs) as candidates for model training in the cross-silo FL by a heuristic algorithm and then motivates the selected ENs to actively contribute their resources. VARF also models the long-term behavior of ENs in cross-silo FL as an infinitely repeated game and derives a stable and long-term cooperative strategy for clients while maximizing the amount of local data for model learning in cross-silo FL. Extensive simulations with real-world data sets demonstrate that the performance of VARF is more beneficial than other benchmarks. Meanwhile, experimental results show that cloud platforms (CPs) and ENs eventually form a long and stable cooperative relationship under the trigger strategy.

This article investigates how to enhance the Multi-access Edge Computing (MEC) systems performance with the aid of device-to-device (D2D) communication computation offloading. By adequately exploiting a novel computation offloading mechanism based on D2D collaboration, users can efficiently share computational resources with each other. However, it is challenging to distinguish valuable information that truly promotes a collaborative decision, as worthless information can hinder collaboration among users. In addition, the transmission of large volumes of information requires high bandwidth and incurs significant latency and computational complexity, resulting in unacceptable costs. In this article, we propose an efficient D2D-assisted MEC computation offloading framework based on Attention Communication Deep Reinforcement Learning (ACDRL), which simulates the interactions between related entities, including device-to-device collaboration in the horizontal and device-to-edge offloading in the vertical. Second, we developed a distributed cooperative reinforcement learning algorithm that includes an attention mechanism that skews computational resources towards active users to avoid unnecessary resource wastage in large-scale MEC systems. Finally, to improve the effectiveness and rationality of cooperation among users, we introduce a communication channel to integrate information from all users in a communication group, thus facilitating cooperative decision-making. The proposed framework is benchmarked, and the experimental results show that the proposed framework can effectively reduce latency and provide valuable insights for practical design compared to other baseline approaches.

Computation offloading decisions play a crucial role in implementing mobile-edge computing (MEC) technology in the Internet of Things (IoT) services. Mobile network operators (MNOs) can employ computation offloading techniques to reduce task completion delay and improve the Quality of Service (QoS) for users by optimizing the system’s processing delay and energy consumption. However, different IoT applications (e.g., entertainment and autonomous driving) generate different delay tolerances and benefits for computational tasks from the MNO perspective. Therefore, simply minimizing the delay of all tasks does not satisfy the QoS of each user. The system architecture design should consider the significance of users and the heterogeneity of tasks. Unfortunately, rare work has been done to discuss this practical issue. In this article, from the perspective of MNO, we investigate the computation offloading optimization problem of multiuser delay-sensitive tasks. First, we propose a new optimization model, which designs different optimization objectives for the cost and revenue of tasks. Then, we transform the problem into a Markov decision processes problem, which leads to designing a multiagent iterative optimization framework. For the strategic optimization of each agent, we further propose a cooperative multiagent deep reinforcement learning (CMDRL) algorithm to optimize two different objectives at the same time. Two agents are integrated into the CMDRL framework to enable agents to collaborate and converge to the global optimum in a distributed manner. At the same time, the priority experience replay method is introduced to improve the utilization rate of effective samples and the learning efficiency of the algorithm. The experimental results show that our proposed method can effectively achieve a significantly higher profit than the alternative state-of-the-art method and exhibit a more favorable computational performance than benchmark deep reinforcement learning methods.