抄録
In recent years, federated learning (FL), a decentralized machine learning approach, has garnered significant attention. FL enables multiple devices to collaboratively train a model without sharing their data. However, when the data across devices are non- independent and identically distributed (non-IID), performance degradation issues such as reduced accuracy, slower convergence speed, and decreased performance fairness are known to occur. Under non-IID data environments, the trained model tends to exhibit varying accuracies across different devices, often overfitting on some devices while achieving lower accuracy on others. To address these challenges, this study proposes a novel approach that integrates reinforcement learning into FL under Non-IID conditions. By employing a reinforcement learning agent to select the optimal devices in each round, the proposed method effectively suppresses the emergence of low-accuracy devices compared to existing methods. Specifically, the proposed method improved the average accuracy of the bottom 10% devices by up to 4%, without compromising the overall average accuracy. Furthermore, the device selection patterns revealed that devices with more diverse local data tend to be chosen more frequently.