Highly flexible Automated Guided Vehicles (AGVs) are interconnected via Industrial Wireless Control Networks (IWCNs) in Multi-access Edge Computing (MEC)-assisted smart factories. The MEC alleviates the lack of computational resources in AGV systems through task offloading. However, IWCNs with limited communication resources struggle to support the highly concurrent offloading of AGVs. In the distributed AGV systems with multi-MEC servers, AGV mobility leads to uneven distribution across MEC server areas, potentially resulting in severe competition for communication resources. Therefore, in this paper, we design a Transferable joint Task Offloading and Multi-Channel Access (T2OMCA) algorithm based on multi-agent deep reinforcement learning. Specifically, AGV observations are modelled as graphs, in which edge relationships are learned through Transformer. This enables AGVs to utilize domain information to collaborate and alleviate concurrent offloading. Moreover, the T2OMCA algorithm converts network input into fixed embeddings to accommodate varying numbers of AGVs. Finally, to encourage exploration in the high-dimensional action space, the T2OMCA algorithm introduces a noisy network and a prioritized experience replay mechanism. Extensive simulations show that the T2OMCA algorithm outperforms existing algorithms in terms of average completion rate, processing delay, and access conflict rate under time-varying AGV topologies.

Federated learning (FL) is a distributed learning framework that allows clients to build models without disclosing local data. However, in resource-constrained scenarios, it is costly to participate in FL for all clients. Hence, selection strategy should be designed to select the most appropriate client groups. Current selection strategies are mainly cost and accuracy oriented, ignoring the autonomy of clients, which leads to the inability of clients to make autonomous decisions when participating in model training and updating. To realize autonomous selection of clients, we design a novel model accuracy-aware bidirectional client selection (MABCS) algorithm. The MABCS algorithm implements selection from both server and client dimensions. Specifically, the server evaluates the contributions of clients and design an accuracy-aware dynamic incentive mechanism. The client measures participation autonomy based on the reward and cost to decide whether or not to participate in FL. Thus, the client selection problem is modeled as a joint nonconvex optimization problem that maximizes the system revenue by optimizing the selection strategy and resource allocation strategy. The block coordinate descent algorithm is utilized to decouple the selection strategy and resource allocation strategy, and a linear approximation is employed to transform the selection strategy problem into a convex problem. An alternating optimization algorithm is used for the subproblems after the decomposition to obtain a near-optimal solution. Simulation results indicate that the MABCS algorithm exhibits superior convergence performance compared with other benchmark schemes.