Related Experiment Videos
Sensitivity analysis applied to the construction of radial basis function networks.
1School of Computer Engineering, Nanyang Technological University, Singapore 639798, Singapore. asdmshi@ntu.edu.sg
Summary
This study introduces a new method for building radial basis function (RBF) networks using sensitivity analysis. This approach improves classification accuracy, matching support vector machine (SVM) performance.
Area of Science:
- Artificial Intelligence
- Machine Learning
- Computational Science
Background:
- Traditional radial basis function (RBF) network construction relies on maximum likelihood learning for cluster centers.
- This method can be suboptimal in determining the optimal number and placement of hidden neurons.
Purpose of the Study:
- To propose a novel learning algorithm for RBF network construction utilizing sensitivity analysis.
- To enhance the generalizability and accuracy of RBF classifiers for unknown data.
Main Methods:
- A new algorithm determines hidden neuron count and centers by maximizing output sensitivity to training data.
- Sensitivity analysis is employed during the training phase for RBF network construction.
- For classification, minimal hidden neurons with maximal sensitivity are selected for optimal generalization.
Main Results:
- The proposed sensitivity-based RBF classifier demonstrates superior performance compared to conventional RBFs.
- The accuracy of the sensitivity-based RBF classifier is comparable to that of support vector machines (SVMs).
Conclusions:
- Sensitivity analysis offers a viable and effective alternative for constructing RBF networks.
- This method enhances RBF network performance in classification tasks.