On Improving Radial Basis Function Neural Networks for Regression

Radial Basis Functional Neural Networks (RBFNNs) are powerful neural network architectures known for their unique approach to learning and pattern recognition. RBFNNs can approximate complex functions efficiently, especially when the data distribution is not well-suited for traditional feedforward neural networks. Compared to fully connected feedforward neural networks, RBFNNs can have fewer parameters, making them potentially easier to train with less data. In this work, we first propose a modified RBFNN model to be used for regression tasks by comparing it with the models proposed in the literature. The effectiveness of four distinct RBFNN architectures for a regression problem is compared. We analyze the performance by changing the architecture and the activation functions in terms of the R2 score, mean absolute error and mean squared error. The performance of the proposed work is analyzed and useful inferences are drawn out.