Scholarlink

<p>Kolmogorov–Arnold Networks (KANs) have recently emerged as an innovative neural architecture that replaces fixed linear weights with trainable univariate spline functions, providing an alternative representation of nonlinear relationships compared with multilayer perceptrons. Despite this architectural advantage, their performance strongly depends on appropriate hyperparameter tuning, particularly with respect to layer width, spline order, grid size, optimizer, learning rate, and weight decay. The interactions among these hyperparameters in a high-dimensional search space make it difficult to determine an effective configuration, making it an optimization problem. In this study, a binary cardiovascular disease classification task is used to assess the effectiveness of Bayesian optimization for hyperparameter tuning of KANs. Three cardiovascular datasets of varying complexity were chosen to examine how model performance changes with different architectures and learning parameters. The study is based on three complementary experiments. The first investigates the impact of the structural hyperparameters, and the second considers different optimizers. The second experiment integrates optimizer selection into the optimization process to optimize both the network architecture and the learning strategy. The third broadens the search space by optimizing the structural hyperparameters, the learning rate, and the weight decay simultaneously. Afterwards, a detailed comparison is carried out between KANs and multilayer perceptrons using the same experimental setup with Bayesian optimization, grid search, random search, and genetic algorithms. The results present an analysis of the impact of each strategy on hyperparameter selection, predictive performance, model interpretability, and computational efficiency. This study thus presents a systematic investigation of Bayesian optimization for KANs and its advantages over the most frequently used optimization methods for training neural networks for cardiovascular disease classification on different datasets of varying complexity.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Optimal tuning of Kolmogorov Arnold Network hyperparameters using Bayesian optimization

  • 期刊论文

  • 发表日期：

    2026年7月22日

查看全文
  • Fatima Fatih,
  • Hanae El Fahfouhi,
  • Zakariae En-Naimani,
  • Khalid Haddouch

摘要

Kolmogorov–Arnold Networks (KANs) have recently emerged as an innovative neural architecture that replaces fixed linear weights with trainable univariate spline functions, providing an alternative representation of nonlinear relationships compared with multilayer perceptrons. Despite this architectural advantage, their performance strongly depends on appropriate hyperparameter tuning, particularly with respect to layer width, spline order, grid size, optimizer, learning rate, and weight decay. The interactions among these hyperparameters in a high-dimensional search space make it difficult to determine an effective configuration, making it an optimization problem. In this study, a binary cardiovascular disease classification task is used to assess the effectiveness of Bayesian optimization for hyperparameter tuning of KANs. Three cardiovascular datasets of varying complexity were chosen to examine how model performance changes with different architectures and learning parameters. The study is based on three complementary experiments. The first investigates the impact of the structural hyperparameters, and the second considers different optimizers. The second experiment integrates optimizer selection into the optimization process to optimize both the network architecture and the learning strategy. The third broadens the search space by optimizing the structural hyperparameters, the learning rate, and the weight decay simultaneously. Afterwards, a detailed comparison is carried out between KANs and multilayer perceptrons using the same experimental setup with Bayesian optimization, grid search, random search, and genetic algorithms. The results present an analysis of the impact of each strategy on hyperparameter selection, predictive performance, model interpretability, and computational efficiency. This study thus presents a systematic investigation of Bayesian optimization for KANs and its advantages over the most frequently used optimization methods for training neural networks for cardiovascular disease classification on different datasets of varying complexity.