Federated learning on non-IID data with reinforcement learning and quantifying historical contribution

In recent years, machine learning related techniques are widely used in edge computing scenarios. This has led to an increase in privacy concern due to the huge amount of data in end devices. To resolve this issue, federated learning (FL) has been proposed. Multiple participating devices perform distributed training locally, keeping training data local to train global neural network model. However, Limited network connectivity restricts parallel model updates and aggregation in federated learning across all devices. Additionally, the presence of non-Independent and Identically Distributed (non-IID) on diverse devices further slow the convergence of FL due to increased local and global model discrepancies. In this paper, we propose Federated learning on non-IID data with Reinforcement learning and Quantifying historical contribution (FedRQ), a non-IID FL framework. The framework aims to intelligently select clients to participate in each round in order to balance the bias introduced by non-IID data and promote model convergence. We evaluated this FedRQ under various FL tasks and verified that this framework outperforms other benchmark methods.