QSplitFL: Capability Aware Deep Q-Learning for Optimal Split Point Selection in Split Federated Learning

The paper introduces QSplitFL, a capability-aware Deep Q-Network (DQN) framework designed for optimal split point selection in Split Federated Learning (SFL) environments. By utilizing lightweight state representations based on client hardware metrics and implementing a decayed loss-drop reward function, QSplitFL improves convergence and accuracy across various datasets (MNIST, CIFAR-10, etc.) using architectures like CNN, ResNet50, and MobileNetV4. This advancement is significant for practitioners as it addresses the challenges of resource heterogeneity in federated learning, optimizing training efficiency and stability.