Hybrid Framework for Hyperparameter Optimization in Federated Learning.

Federated Learning (FL) enables collaborative model training across decentralized clients while preserving data privacy, yet optimizing its hyperparameters remains a critical challenge due to data heterogeneity, dynamic resource constraints, and heavy communication overheads. Existing hyperparameter optimization techniques often fail to jointly tune system-level variables, execution dynamics, and internal algorithm parameters. In this paper, we propose PSO-REALF, a novel three-tier hybrid optimization framework designed to maximize efficiency and model accuracy in non-IID federated environments. The architecture leverages a Particle Swarm Optimization (PSO) at the intermediate layer to dynamically adjust core FL hyperparameters, and Resource-Efficient Accelerating Federated Learning (REAFL) at the execution layer to optimize local iteration assignments ($t_n$) under fixed time budgets. Evaluated on non-IID partitions of the MNIST and CIFAR-10 datasets using Convolutional Neural Networks, PSO-REALF demonstrates superior convergence speed and accuracy compared to state-of-the-art baselines like FedEx and Auto-FedRL.