Optimization using GA tool from ANN output

Hi all, I am aiming at optimizing a problems which has 4 inputs and one output. I did ANN with 3input neurons, 9 hidden and one ouput neurons. The transfer functions are tansig for hidden and purelin for output layers respectively. I got the weights and biases as follows from ANN: w1 = [-3.8 -3.16 -3.88 -3.31;0.93 -2.24 -0.904 0.086;7.64 2.06 -10 4.845; -3.9 2.483 2.823 2.461;-6 -3.84 -6.821 0.497;-6.7 8.261 4.412 -7.66;-0.1 -2.93 0.48 -5.81;-1.7 -5.62 -4.59 5.304;9.18 -0.14 -3.943 -4.597] w2 = [-5.27;-3.93;-0.62;-0.65;2.056;0.671;-1.56;0.252;1.353] b1 = [-0.5;-2.15;-0.045;3.132;3.137;-5.93;-1.97;-4.24;2.036]' b2 = 16

and the follwoing equation is the one which I have to optimize by ga tool: z = (w2*((2/(1+(exp(-2*(((w1*x')+b1))))))-1))+b2;

lower bounds: [-1 -1 -1 -1] and upper bounds [1 1 1 1]

when I run the ga tool I get the output like this: Optimization running. Error running optimization. Your fitness function must return a scalar value.

can anyone help?

Answer 1

[Greg Heath]

1. Although the hidden and output layers contain neurons (tansig & purelin), the input layer nodes are fan-in units, NOT neurons.

2. You said 3 inputs instead of 4

3. Cutting and pasting your assignment statements yields

whos

Name      Size            Bytes  Class    

b1        1x9                72  double              
b2        1x1                 8  double              
w1        9x4               288  double              
w2        9x1                72  double

if size(x) = [ 4 N ] and size(t) = [ 1 N ], the correct dimensions should be

 [ 9 1 ] = size(b1)

 [ 1 9 ] = size(w2)

4. Then the output and error 1XN vectors are

 z = b + w2 * tansig( repmat( b1, 1, N ) + w1 * x) ;

 e = t - z;

5. The optimization function is the scalar mean-squared-error

 MSE = mse(t-z) = mean( (t-z).^2 )

Hope this help

Thank you for formally accepting my answer

Greg

Answer 2

[Alan Weiss]

I don't know anything about neural networks, but I think I can tell you how to diagnose this problem: give an input, say x0 = rand(4,1), and see what your fitness function returns

 fun(x0)

If I am correct, it will return a vector or matrix, not a scalar.

Optimization functions generally require scalar output from the fitness function.

Alan Weiss

MATLAB mathematical toolbox documentation

[Terry Chen]

Hi,

From above Raja's question and other discuss , can we optimize by ga tool: z = (w2*((2/(1+(exp(-2*(((w1*x')+b1))))))-1))+b2;lower bounds: [-1 -1 -1 -1] and upper bounds [1 1 1 1] ?

If possible , How to optimize it?Have anyone explain more ,Thanks

Terry