DDPG-FL: A Reinforcement Learning Approach for Data Balancing in Federated Learning

Federated learning (FL) is a novel distributed machine learning framework aimed at preserving privacy. A significant challenge in federated learning is the presence of non-independent and non-identically distributed (Non-IID) data among clients. Since local data is generated in diverse environments, the data distribution across data partitions may differ considerably, leading to slower and less accurate model training and increased communication overhead. In this paper, we propose a reinforcement learning-based federated learning data balancing algorithm on Non-IID data, DDPG-FL, which does not require the collection or inspection of any private information and does not introduce additional communication overhead. Additionally, DDPG-FL can be combined with existing FL algorithms as a data balancing plug-in. Experimental results on the MNIST, FMNIST, and CIFAR-10 datasets demonstrate that DDPG-FL enhances model performance while significantly reducing the number of communication rounds required for model convergence.