Optimizing training delay of federated learning with reinforcement learning-based pruning
Yuan‐Cheng Lai, Chih-Wen Huang, Yen-Hung Chen, Pei-Fei Chen, Liang‐Chun Chen · Journal of the Chinese Institute of Engineers · 2025
In traditional federated learning frameworks, it is typically assumed that bandwidth is sufficiently robust, enabling artificial intelligence systems to exchange model weights freely. However, as models grow in complexity and size, real-world network conditions, which are often variable, may lead to insufficient bandwidth for the timely exchange of weights. Conventional methodologies, such as model pruning, frequently alter the original model structure and prove inadequate under fluctuating bandwidth. To address these challenges, this study proposes a novel approach known as Federated Learning-based Differential Weight Selection and Exchange (FedDW). FedDW mitigates constraints by selectively exchanging model weights that exhibit significant changes, optimizing transmission while preserving the integrity of the model. FedDW employs reinforcement learning as the mechanism for the selection and exchange of weights. Experimental results show that FedDW surpasses traditional full-model upload techniques and outperforms pruning strategies such as SynFlow in both efficiency and accuracy. Specifically, it achieves a time reduction of 8.49% to 15.39% in scenarios characterized by low bandwidth and non-independent and identically distributed (Non-IID) data. These findings demonstrate the effectiveness of FedDW in enhancing training time and model accuracy for federated learning systems facing latency and bandwidth constraints.