Hyper-parameter settings.

Federated Learning (FL), as a distributed computing framework for training machine learning (ML) models, has garnered significant attention for its superior privacy protection. In typical FL, a subset of client models is randomly selected for aggregation in each iteration, which performs well when the data is independent and identically distributed (IID). However, in real-world scenarios, data is often non-independent and identically distributed (Non-IID). Random selection cannot capture knowledge from different data distributions, resulting in a global model with lower accuracy and slower convergence. To address this challenge, we propose base models, which are models with diverse data distributions on clients. By combining the parameters of these base models, we can approximate all client models. Meanwhile, we sufficiently demonstrate the existence of base models. Then we employ the evolutionary algorithm (EA) to identify base models on distributed clients by encoding client IDs and optimizing client selection through crossover, mutation, and other evolutionary operations. Our method addresses the issue of low efficiency in random selection. We conduct experiments on the FashionMNIST, MNIST, and TodayNews datasets, applying the proposed method to FL frameworks such as FedAvg, FedProx, and SCAFFOLD, all of which show superior performance and faster convergence.