how to use weights and thresholds from Neural Network toolbox

Establish a very simple feed forward network to approximate sin wave:

x=[0:0.1:10];

y=sin(x);

net=newff(x,y,30); %30 neurons

net=train(net,x,y); %train

y=sim(net,x);
plot(x,sin(x),x,y);

The result is good. However, if I take the weight and threshold matrix from this network:

w1=net.IW{1,1};

w2=net.LW{2,1};

b1=net.b{1};

b2=net.b{2};

and use a loop to simulate it:

xx=[0:0.5:5];

for i=1:size(x,2)

    Y(i)=w2*tansig(w1*xx(i)+b1)+b2;
end

plot(xx,sin(xx),xx,Y);

the result is totally different. Is there any problem?

Answer 1

[Andres Zarate de Landa]

I figured out the problem! The thing here is that newff by default normalices the inputs and outputs using the mapminmax function. You can see that by typing net.inputs{1}.processFcns 'fixunknowns' 'removeconstantrows' 'mapminmax'

net.outputs{2}.processFcns 'removeconstantrows' 'mapminmax'

So, if you don't want normalized inputs and outputs as in my case just remove the mapminmax function: net.inputs{1}.processFcns = {'fixunknowns', 'removeconstantrows'}; net.outputs{2}.processFcns = {'removeconstantrows'};

And that's all! the weights and biases won't be normalized and you can use these parameters in your own ANN using any software. I hope this helps

Answer 2

[Mark Hudson Beale]

You need to first preprocess inputs, then post process outputs as follows:

xx = [0:0.5:5]
for i=1:length(net.inputs{1}.processFcns)
  xx = feval(net.inputs{1}.processFcns{i},...
      'apply',xx,net.inputs{1}.processSettings{i});
end
for i=1:size(x,2)
    Y(i)=w2*tansig(w1*xx(i)+b1)+b2;
end
for i=1:length(net.outputs{2}.processFcns)
  Y = feval(net.outputs{2}.processFcns{i},...
      'reverse',Y,net.outputs{2}.processSettings{i});
end

You can replace the FEVAL calls with direct calls to the functions in "net.inputs{1}.processFcns" and "net.outputs{2}.processFcns" if you like.

Answer 3

[Andres Zarate de Landa]

I have the exact same problem, I need the Weights and biases to simulate the ANN using equations on ADS. The function aproximation results using the "sim" function are ok, however when I try to simulate it it doesn't work at all. I don't know what's wrong can anybody help please?

Thanx