Dear Sukhesh Reddy ,

ANN are used as surrogate models to teduce the computational cost when the objection function is expensive to evaluate. the first step is to properlt train an ANN model. After that this model can be used to calculate the value of objective function in your program. In other words, you use ANN as an approximate or surrogate model to avoid costly calculations of an expensive objective function.

I suggest to read the following articles

Jin, Y. (2005). A comprehensive survey of fitness approximation in evolutionary computation. Soft computing, 9(1), 3-12.

Zhou, Z., Ong, Y. S., Nair, P. B., Keane, A. J., & Lum, K. Y. (2007). Combining global and local surrogate models to accelerate evolutionary optimization. Systems, Man, and Cybernetics, Part C: Applications and Reviews, IEEE Transactions on, 37(1), 66-76.

Shen, C., Wang, L., & Li, Q. (2007). Optimization of injection molding process parameters using combination of artificial neural network and genetic algorithm method. Journal of Materials Processing Technology, 183(2), 412-418.

http://epubs.surrey.ac.uk/7610/2/SC2005.pdf

http://eprints.soton.ac.uk/43820/01/zhou_07.pdf

http://www.academia.edu/download/30525252/201031111520597.pdf

Dear Suresh,

I guess you meant to say how can use use GA to as a fitness function to re-direct ANN toward a solution in the search space. There are lots of works that have been done using hybrid ANN-GA algorithms. NEAT is an improved approach toward optimization since it allows for speciation. I am really interested to know the details of your approach.. please do share.

Hi Mr. Reddy,

I agree with Dr. Behrouz. However, I used GA to optimize no. of hidden neurons, learning rate and momentum rate while implementing Multi-layer perceptron Neural network in a bearing diagnostic problem. I defined an ojective function to maximize the training and test accuracy as well as minimize the training goal (mse). 

In another work i used the RBF NN, wherein i optimized the no. of hidden rbf units (hidden neurons), width (spread) of rbf units, learing rate and momentum rate. Here also the ojective was to maximize train & test accuracies and at the same time minimize training goal.

In both cases i achieved good NN performance but the cost of computation was high. You may pls refer to my papers that i have uploaded in my RG account.

I suggest that u use the mutiobjectve optimization and extreme learning machine (instead of ANN), which could improve overall performance of the system.

With best wishes.

Hi,

As I understand you have the training set representing some function and you want to find the maximum of this function?

So first you learn NN using the training set (be careful with overfitting). Then you use GA with vector of your 4 inputs as a chromosome. To evaluate chromosomes you use learned NN (you present for NN input variables coded in each chromosome and record NN output, which is a fitness).

Have fun

Hi Sukhesh Reddy you can use GA to select several aspects of the ANN: preprocessing technique, features and to optimize parameters like number of hidden layers, number of neurons, and others. I've tested GA with ANN in order to select best features in large datasets with more than 30 features. The value of your parameters comes in the chromosomes, you test different individuals with your ANN fitness function, R-square and RMSE if the nature of problem is regression, or accuracy specificity and sensitivity if it is a classification problem. Each individual will interact with others like in Darwin 's natural selection, but you can keep the best individuals for next generations; after a while you have a individual whose characteristics perform very well in the desired function.

You can export your trained ANN model to the directory and then create a function file calling your network. 

function y = network(x) saveVarsMat = load('NNet.mat'); net = saveVarsMat.net; y = (net(x'));

Then, call this function as fitness function for you Genetic algorithm optimization.

Well i am working on a similar project with 6 input parameter to optimize the output. If i get an optimum of the 6 parameters, my output will be optimize. The answer to Suresh question will help me alots

Thanks

Hi Peter Tekum

You have experience in this method. Could you please help me write a fitness function of the ANN model for the GA tool? Thanks u

Here is myNeuralNetworkFunction from ANN Fitting Tool:

function [y1] = myNeuralNetworkFunction(x1)

%MYNEURALNETWORKFUNCTION neural network simulation function.

%

% Generated by Neural Network Toolbox function genFunction, 28-Oct-2020 21:51:31.

%

% [y1] = myNeuralNetworkFunction(x1) takes these arguments:

% x = 6xQ matrix, input #1

% and returns:

% y = 2xQ matrix, output #1

% where Q is the number of samples.

%#ok

% ===== NEURAL NETWORK CONSTANTS =====

% Input 1

x1_step1.xoffset = [37.7;33.711;108;13.8751;65;1.45];

x1_step1.gain = [0.065359477124183;0.0781121847196944;0.0253164556962025;0.0269451356620218;0.0588235294117647;0.44627715596494];

x1_step1.ymin = -1;

% Layer 1

b1 = [2.8879840368006131079;-1.8024459794372993393;2.2779515314980849361;5.971025628044506206;3.1355432111290468811;-2.7525929588071798015;1.297028337344750204;-0.45542871345809110561;-0.1968010285845217533;0.60210752365484743187];

IW1_1 = [3.5726681637389905433 0.29170483214456532917 -0.19841796131177430462 4.5325961843523847605 0.72031210145284341184 -14.244280611129411795;-6.6660793842584151747 -1.0042131984849427262 0.72201172298350146939 -2.7917053260678277304 -4.9965963929961283441 1.7563036791454456065;0.14628021585912026614 9.466978438946551222 -2.6374014533190375253 0.88672322101160971908 5.1710150426939973656 -2.4765738097126059891;-5.0017706812542259698 -0.84670364506682527228 -1.2544213522500204938 -0.27982310891090661498 14.850409207355353303 3.0715887820172027567;-0.12523375561854377391 -2.7722614212789693511 -3.5898161415394107543 2.5242482890865218792 -2.8210868593135089988 -2.372613636109296209;-11.371559126457448485 -3.8800729440535333126 3.5038981547199377076 1.8867150350227186451 -6.434963520515181834 5.0123150162527787543;-1.3582311663016251035 1.4656439082835599486 1.6543439036873219816 3.0352581252758037422 -4.8733143388829951448 -6.4817677339123953573;-2.3388062994211198387 1.2433093267909431923 1.017897428004527649 -4.4009642933127697262 -2.3535749743682856661 1.8028451958626590113;-2.0269895331748766232 -2.4285739970700830348 3.7829139882405939943 -4.1862233114642943832 1.59269051067088796 0.11283473238730951094;7.4248180249611834824 -4.3954574481695019017 -12.61349616626842618 3.5218184675398953587 -5.5036914450627607565 8.7026819576086094088];

% Layer 2

b2 = [0.36860620245894359925;-0.083084243673893803939];

LW2_1 = [0.16427317633998383961 -0.22262253756552802253 0.1775966877768416774 0.074896982333248277031 -0.4067737109597457934 0.31168085589022070936 -0.25428561131256771644 0.022479897418241995277 -0.044529198877253095556 0.10069680403761865906;-0.17115225681244555789 -0.40939284015394494798 -0.37169247075913564027 -0.10955747927970435618 -0.087935155842651935298 0.10368916139884838679 0.29015894778042738977 -0.16501713779514770586 0.16991526978770260525 -0.081198977020344620392];

% Output 1

y1_step1.ymin = -1;

y1_step1.gain = [0.25;0.0350877192982456];

y1_step1.xoffset = [0;15];

% ===== SIMULATION ========

% Dimensions

Q = size(x1,2); % samples

% Input 1

xp1 = mapminmax_apply(x1,x1_step1);

% Layer 1

a1 = tansig_apply(repmat(b1,1,Q) + IW1_1*xp1);

% Layer 2

a2 = repmat(b2,1,Q) + LW2_1*a1;

% Output 1

y1 = mapminmax_reverse(a2,y1_step1);

end

% ===== MODULE FUNCTIONS ========

% Map Minimum and Maximum Input Processing Function

function y = mapminmax_apply(x,settings)

y = bsxfun(@minus,x,settings.xoffset);

y = bsxfun(@times,y,settings.gain);

y = bsxfun(@plus,y,settings.ymin);

end

% Sigmoid Symmetric Transfer Function

function a = tansig_apply(n,~)

a = 2 ./ (1 + exp(-2*n)) - 1;

end

% Map Minimum and Maximum Output Reverse-Processing Function

function x = mapminmax_reverse(y,settings)

x = bsxfun(@minus,y,settings.ymin);

x = bsxfun(@rdivide,x,settings.gain);

x = bsxfun(@plus,x,settings.xoffset);

end

function MSE=ffann1(input)

global net expected;

y=net(input');

MSE=abs(y-expected);

Thank you Peter Tekum

https://www.sciencedirect.com/science/article/pii/S2212017312006056/pdf?md5=5e9a5fd79c582df5c81faf159411a4f5&pid=1-s2.0-S2212017312006056-main.pdf