부산대학교 도서관

Recently, federated learning (FL) has become a promising distributed learning technology by collaboratively training shared learning models on clients. However, due to the statistical heterogeneity of clients and differences in computing and communication resources, the convergence speed and accuracy of FL may decrease. The energy consumption and latency performance of clients may also be affected. To achieve a flexible balance between FL's model performance, overall energy consumption, and latency, thereby meeting customized requirements, we propose a deep reinforcement learning based FL framework called FedAEB. It adopts a dynamic optimization method based on the Soft Actor-Critic network for client selection and resource allocation, which can effectively adapt to complex and time-varying systems. The weight factors that balance optimization variables can be flexibly adjusted according to different application needs. Many experiments have been conducted on well-known and state-of-the-art datasets, demonstrating that our FedAEB outperforms the benchmark method in reward values, learning accuracy, energy consumption, and latency performance.