Scholarlink

<p>Distributed decision-making systems increasingly rely on reinforcement learning agents operating under privacy, scalability, and communication constraints. Although federated deep reinforcement learning (FDRL) enables collaborative policy learning without sharing raw data, most existing approaches adopt static or uniform aggregation strategies that fail to capture agent heterogeneity and coordination dynamics. To address these limitations, this paper proposes a novel <i>Federated Swarm-Coordinated Deep Deterministic Policy Gradient (FSC-DDPG)</i> framework that tightly integrates swarm intelligence with federated learning for continuous-control multi-agent systems. In FSC-DDPG, each agent learns locally using the Deep Deterministic Policy Gradient algorithm, while a swarm-inspired feedback mechanism continuously evaluates agent reliability, behavioral consistency, and learning contribution. These swarm signals are exploited to adaptively regulate aggregation weights and participation intensity during federated updates, enabling robust coordination without exposing private data. The resulting bidirectional interaction between swarm intelligence and federated optimization promotes stable learning and mitigates the impact of unreliable or noisy agents. Extensive experimental results on distributed decision-making benchmarks demonstrate that FSC-DDPG achieves faster convergence, improved stability, and superior performance compared with conventional federated and multi-agent reinforcement learning baselines. The proposed framework provides a scalable and extensible solution for privacy-preserving and adaptive coordination in heterogeneous decentralized systems.</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

Federated swarm-coordinated deep reinforcement learning for distributed decision-making systems

  • 期刊论文

  • 发表日期：

    2026年6月17日

查看全文
  • Tarek Amine Haddad

摘要

Distributed decision-making systems increasingly rely on reinforcement learning agents operating under privacy, scalability, and communication constraints. Although federated deep reinforcement learning (FDRL) enables collaborative policy learning without sharing raw data, most existing approaches adopt static or uniform aggregation strategies that fail to capture agent heterogeneity and coordination dynamics. To address these limitations, this paper proposes a novel Federated Swarm-Coordinated Deep Deterministic Policy Gradient (FSC-DDPG) framework that tightly integrates swarm intelligence with federated learning for continuous-control multi-agent systems. In FSC-DDPG, each agent learns locally using the Deep Deterministic Policy Gradient algorithm, while a swarm-inspired feedback mechanism continuously evaluates agent reliability, behavioral consistency, and learning contribution. These swarm signals are exploited to adaptively regulate aggregation weights and participation intensity during federated updates, enabling robust coordination without exposing private data. The resulting bidirectional interaction between swarm intelligence and federated optimization promotes stable learning and mitigates the impact of unreliable or noisy agents. Extensive experimental results on distributed decision-making benchmarks demonstrate that FSC-DDPG achieves faster convergence, improved stability, and superior performance compared with conventional federated and multi-agent reinforcement learning baselines. The proposed framework provides a scalable and extensible solution for privacy-preserving and adaptive coordination in heterogeneous decentralized systems.