Good references:

Leung, F. H., Lam, H. K., Ling, S. H., & Tam, P. K. (2003). Tuning of the structure and parameters of a neural network using an improved genetic algorithm. Neural Networks, IEEE Transactions on, 14(1), 79-88.

Huang, H. X., Li, J. C., & Xiao, C. L. (2015). A proposed iteration optimization approach integrating backpropagation neural network with genetic algorithm. Expert Systems with Applications, 42(1), 146-155.

Ahmadizar, F., Soltanian, K., AkhlaghianTab, F., & Tsoulos, I. (2015). Artificial neural network development by means of a novel combination of grammatical evolution and genetic algorithm. Engineering Applications of Artificial Intelligence, 39, 1-13.

https://opus.lib.uts.edu.au/bitstream/10453/15148/1/2010004238.pdf

Article Artificial neural network development by means of a novel co...

Article A proposed iteration optimization approach integrating backp...

Read this:

Using Genetic Algorithm to Optimize Artificial Neural Network: A Case Study on Earthquake Prediction--Cheng Wang, Qiuwen Zhang

FITNESS FUNCTION CAN BE MSE OR RMSE. THE RMSE IS THE ROOT MEAN SQUARED ERROR BETWEEN THE ACTUAL AND TARGET VALUES. THE OBJECTIVE OF THE GA WILL BE TO REDUCE THIS FITNESS FUNCTION. A SUITABLE STOPPING CRITERION CAN BE CHOSEN FOR STOPPING GA. ONCE GA IS STOPPED THE OPTIMAL WEIGHTS AND BIASES CAN BE SAVED. THE BACK-PROPGATION ALGORITHM IS INITIALIZED WITH THESE OPTIMIZED WEIGHTS AND BIASES TO FURTHER REDUCE THE ERROR TO AN ACCEPTABLE MINIMUM.

Hi Beverly!

Is there a reason you want to use a GA to optimize your NN? Normally, NNs are trained only with backpropagation and this is sufficient for most problems.

To that end you use the equation you mentioned and calculate e.g. the MSE or RMSE from actual output to target outputs. This is the cost function you want to minimize. From that function you calculate the first derivative with respect to each single parameter (weights and biases) which gives you the gradient for each parameter. At each timestep, you subtract the gradients from the respective, current parameters. This will lead to a cost minimization.

Of course, if you have a good reason for using GAs, it is totally fine, however, that is not the usual way to training NNs.

You might be interested in the NEAT/HyperNEAT work from the EPlex group at UCF lead by Ken Stanley.

http://eplex.cs.ucf.edu/

http://eplex.cs.ucf.edu/hyperNEATpage/

If you have ANN model with discused strucure above and you want to use Ga as learning mechnism to calculate the fit function is e means difference between actual and desired value and e by ga is minimized.  

Hi, annex article using industry by  optimize RN using genetic algorithm

You could use Cartesian Genetic Programming. It has been proven in the past to optimise well neural networks.

you could find useful our work, for optimization of the number of inputs of a NN using GA

Dear Adam,

interesting point of view.

Coul you please explain better the method you use for optimization of inputs?

thanks.

I am teaching standard two ,what can I use as a set induction for sources of water?They are 8 years old.

Good references:

Leung, F. H., Lam, H. K., Ling, S. H., & Tam, P. K. (2003). Tuning of the structure and parameters of a neural network using an improved genetic algorithm. Neural Networks, IEEE Transactions on, 14(1), 79-88.

Huang, H. X., Li, J. C., & Xiao, C. L. (2015). A proposed iteration optimization approach integrating backpropagation neural network with genetic algorithm. Expert Systems with Applications, 42(1), 146-155.

Ahmadizar, F., Soltanian, K., AkhlaghianTab, F., & Tsoulos, I. (2015). Artificial neural network development by means of a novel combination of grammatical evolution and genetic algorithm. Engineering Applications of Artificial Intelligence, 39, 1-13.

https://opus.lib.uts.edu.au/bitstream/10453/15148/1/2010004238.pdf

Article Artificial neural network development by means of a novel co...

Article A proposed iteration optimization approach integrating backp...

You can use parameters such as the MSE, Current epoch, maximum epochs, Number of neuron in the hidden layer, and Maximum neuron to write a suitable fitness function.

You can set the fitness function based on average RMS error (not absolute relative error-because of zero data). I have firstly hypothesized a topography. Then select random weight and bias matrices as well as the steepness matrix for the slope of logistic function. Then by GA, these matrices have been modified to provide least RMS error. The fittings are very better in compare to MATLAB toolbox.

If you can withe your neural network problem as a mfile function in MATLAB you can use MATLAB GA toolbox easily.