Deep Reinforcement Learning for Client Selection and Resource Allocation in Federated Learning: A Comprehensive Survey

Federated Learning (FL) enables decentralized model training
across distributed devices while preserving data privacy. However,
client heterogeneity, communication bottlenecks, and dynamic resource
availability pose substantial challenges. Deep Reinforcement Learning
(DRL) offers a powerful framework for adaptive client selection and
resource allocation in such environments. This survey examines DRLdriven
strategies across centralized, hierarchical (HFL), cross-silo (CSFL),
and edge-based (FEEL) FL architectures. We categorize and analyze
both single-agent and multi-agent DRL approaches, evaluating them on
complexity, scalability, fairness, and communication efficiency. Our study
aims to guide the design of robust, scalable FL systems optimized via
DRL under heterogeneous real-world conditions.