Multi-Agent Transformer-Based Workload Allocation and Worker Selection in Distributed Coded Machine Learning

Machine learning (ML) has been used in a wide range of applications. In recent years, the growing complexity of ML models has led to substantial computational demands. Distributed Machine Learning (DML) frameworks address this challenge by parallelizing computation across multiple devices. Yet, DML systems often suffer from the straggler effect, where the slowest devices delay task completion. Distributed Coded Machine Learning (DCML) mitigates this issue by injecting redundancy into subtasks, allowing partial results to suffice for completing the overall task. To fully exploit DCML, two critical problems, worker selection and workload allocation, must be addressed effectively, especially in large-scale edge networks characterized by unstable connectivity, dynamic device availability, and high-dimensional decision spaces. In this work, we propose a Multi-Agent Transformer-based workload Allocation and worker Selection (MAT-AS) scheme for DCML in large-scale edge networks. Technically, MAT-AS converts the multi-agent joint policy optimization problem into a sequential process to make robust decisions on worker selection and workload allocation. Experimental results demonstrate that MAT-AS significantly outperforms baseline approaches in task completion time and cost efficiency, making it a promising solution for efficient and adaptive DCML. Specifically, MAT-AS reduces task completion time by approximately 12% and computation cost by approximately 15% compared to the best baseline method.