FL-FAITH: A Hybrid MCDM-DDPG Framework for Multi-Objective Client Selection in Federated Learning

Client selection significantly influences the overall performance, convergence behavior, and resource utilization of Federated Learning (FL), particularly under circumstances typified by device heterogeneity and resource limitations. Conventional selection techniques often depend either on rule-based Multi-Criteria Decision Making (MCDM) or on adaptive Reinforcement Learning (RL), each exhibiting inherent drawbacks: MCDM offers explainability but lacks flexibility , whereas RL enables dynamic decision-making but is often opaque and computationally intensive. In order to mitigate these issues, we introduce FL-FAITH, a hybrid client selection framework that integrates the interpretability of MCDM with the adaptive optimization power of Deep Deterministic Policy Gradient (DDPG)-based RL. FL-FAITH dynamically adjusts the importance of client-side features through reward-driven learning, guided by a multi-objective reward function. This function encompasses not only global accuracy and training convergence but also operational constraints such as bandwidth usage, energy efficiency, processing capability, and data quality. The framework favors clients with reliable and high-bandwidth connectivity to reduce communication cost, selects energy-efficient nodes for sustained participation, and leverages computa-tionally powerful devices to accelerate local training. Additionally, data quality metrics—like label distribution and sample volume—are integrated into the 1 scoring mechanism to ensure clients meaningfully support global model generalization. By balancing these factors, FL-FAITH enables robust and context-aware client selection across diverse environments. Experimental results on the MIT-BIH Arrhythmia dataset, using varied client hardware profiles, demonstrate that FL-FAITH consistently surpasses standalone MCDM and DDPG baselines in accuracy, convergence, and resource-awareness. This framework effectively bridges the gap between interpretability and adaptivity, delivering a feasible and extensible solution for practical federated learning deployments.