Absolute error instead of Mean Square Error - NN

It is possible to change the MSE performance to a simple Error performance? So, considering the minimization of the (maximum) error in a neural network?

Thanks in advance!

Answer 1

[Amith]

Hi Francesco,

It is possible to change the performance function in MATLAB to minimize the maximum error instead of the mean squared error (MSE). You can achieve this by creating a custom performance function. Here’s a general approach to do this:

1. Define the Custom Performance Function: Create a new function that calculates the maximum error. Save this function as a .m file.

function perf = max_error(t, y)    
% t: targets    
% y: outputs    
% Calculate the maximum absolute error    
perf = max(abs(t - y)); 
end

2. Set the Custom Performance Function in Your Neural Network: Assign this custom performance function to your neural network.

% Create a feedforward neural network
net = feedforwardnet(10); 
% Set the custom performance function
net.performFcn = 'max_error'; 
% Train the network
[x, t] = bodyfat_dataset; 
% Example dataset
net = train(net, x, t); 
% Evaluate the performance
y = net(x); 
perf = perform(net, t, y); 

This approach allows you to customize the performance evaluation to focus on minimizing the maximum error rather than the mean squared error.

Hope this helps!