DE-KAN: A Differential Evolution-Based Optimization Framework for Enhancing Kolmogorov-Arnold Networks in Complex Nonlinear Modeling

Kolmogorov-Arnold Networks (KANs) offer superior fitting ability and interpretability compared to traditional Multi-Layer Perceptrons (MLPs). However, their complex structure and lengthy training process pose significant challenges in practical applications. Existing optimization methods, primarily focused on replacing activation functions, fail to address the core issue of model complexity. In this paper, we introduce DE-KAN, a novel framework that leverages Differential Evolution (DE) to optimize the network structure of KANs. By modeling KAN as a discrete optimization problem, DE-KAN automates the selection of optimal connection patterns between network layers, significantly enhancing both efficiency and performance. Our experimental results on classification and regression datasets demonstrate that DE-KAN not only simplifies model complexity but also achieves notable improvements in accuracy and interpretability compared to traditional KANs and other baseline models.