Scholarlink

<p>The rapid of heterogeneous, data-generating devices, such as wireless network and Internet of Things (IoT) devices, necessitates scalable and robust distributed learning frameworks capable of handling non-independent and identically distributed (non-IID), heterogeneous, and dynamic data. In this work, we propose a unified adaptive federated learning (UA-FL) framework, which integrates an adaptive federated dual-level momentum with variance control (AFDM-VC) optimizer at the client level and a multi-agent reinforcement learning (MARL)-driven participant selection mechanism at the aggregator level. AFDM-VC mitigates gradient variance arising from heterogeneous local data while leveraging momentum for accelerated convergence. The unified aggregator layer performs variance-aware weighted aggregation and asynchronous meta-aggregation with the global controller, ensuring stability and robustness in dynamic, large-scale networks. We provide formal convergence analysis, establishing linear convergence under strongly convex objectives and sublinear convergence for general non-convex settings. Extensive theoretical insights and algorithmic design demonstrate that UA-FL achieves enhanced communication efficiency, convergence stability, and resilience to unreliable clients, making it suitable for real-world deployments across heterogeneous environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Unified adaptive federated learning for heterogeneous environments via AFDM-VC and MARL

  • 期刊论文

  • 发表日期：

    2026年1月13日

查看全文
  • Majid Mohammadpour,
  • Seyedakbar Mostafavi,
  • Jamshid Abouei

摘要

The rapid of heterogeneous, data-generating devices, such as wireless network and Internet of Things (IoT) devices, necessitates scalable and robust distributed learning frameworks capable of handling non-independent and identically distributed (non-IID), heterogeneous, and dynamic data. In this work, we propose a unified adaptive federated learning (UA-FL) framework, which integrates an adaptive federated dual-level momentum with variance control (AFDM-VC) optimizer at the client level and a multi-agent reinforcement learning (MARL)-driven participant selection mechanism at the aggregator level. AFDM-VC mitigates gradient variance arising from heterogeneous local data while leveraging momentum for accelerated convergence. The unified aggregator layer performs variance-aware weighted aggregation and asynchronous meta-aggregation with the global controller, ensuring stability and robustness in dynamic, large-scale networks. We provide formal convergence analysis, establishing linear convergence under strongly convex objectives and sublinear convergence for general non-convex settings. Extensive theoretical insights and algorithmic design demonstrate that UA-FL achieves enhanced communication efficiency, convergence stability, and resilience to unreliable clients, making it suitable for real-world deployments across heterogeneous environments.