You can add a hidden layer between input and output using "feedforwardnet" function (https://www.mathworks.com/help/releases/R2020a/deeplearning/ref/feedforwardnet.html?s_tid=doc_ta) instead of initializing net with "linearlayer". This creates one hidden layer of neurons with the number of neurons you specify. You can then set the transfer function at each of the nodes to be linear by specifying "purelin" as the value of the attribute "transferFcn" (https://www.mathworks.com/help/releases/R2020a/deeplearning/ref/purelin.html?s_tid=doc_ta):
function trainLNN(P, T)
% Example Input:
% P = [1.7300 1.9500 2.3800; 1.4400 5.0000 4.7500; 3.4000 3.2000 3.3000];
% T = [0 1 0; 1 0 0; 0 0 1];
% Define the size of the hidden layer
hiddenLayerSize = 3;
% Create a feedforward network with one hidden layer
net = feedforwardnet(hiddenLayerSize);
% Set all layers' transfer functions to be linear
% For hidden layers
for i=1:(length(net.layers)-1)
net.layers{i}.transferFcn = 'purelin';
end
% For the output layer
net.layers{end}.transferFcn = 'purelin';
%
% net.trainParam.epochs = 1000;
% net.trainParam.goal = 0.1;
% net.trainParam.lr = maxlinlr(P, 'bias');
%
% net = train(net, P, T);
end
