This repository demonstrates the application of efficient Kolmogorov-Arnold Networks (KAN) in a curve fitting (regression) task. The original KAN can be found here, while the original efficient KAN can be found here. Another similar example of Lorentzian curve fitting using KAN can be found at here.
The curve function here is: y = a·sin(b·x)+c·cos(d·x), x = 0:0.2:10. [a, b, c, d] used here are randomly sampled from [0.1:0.1:2].
You may change it to whatever function you would like to fit.
The training dataset was created using the matlab code ‘create_dataset.m’ under 'Data' folder.
Network specifics: size(inputlayer, hiddenlayer, outputlayer) = [51, 100, 4].
The input is curve values y with a length of 51, and the output is the coefficients vector [a, b, c, d] with a length of 4, as shown below.
The loss curves of KAN and MLP after training for 30 epochs are as follows:
The predicted curves by MLP and KAN after training for 30 epochs are as follows:
