Yes absolutely possible - even you can optimize number of neurons and and other structure related parameters through GA

Depends on application and problem - you can try yourself to get answer for second part of your question by solving the same problem through BP and GA both ..

Hello, Sure you can you GA to find the weights, check the following paper " http://www.google.com/url?sa=t&rct=j&q=&esrc=s&source=web&cd=1&ved=0CDIQFjAA&url=http%3A%2F%2Fijcai.org%2FPast%2520Proceedings%2FIJCAI-89-VOL1%2FPDF%2F122.pdf&ei=hNAoUcLlK8K-0AGng4Fo&usg=AFQjCNGvrOeEx-khFR2BOdmVrN7HlmyJ8A&bvm=bv.42768644,d.dmQ&cad=rja"

Also you can check my work on optimizing Hidden markov models using GA and AIS. Both can applied to neural network

For the second part, Simply GA is much slower than GA but Theoretically it should give better results or the same results as BP. My suggestion from my experience with NN and GA to start with BP. It should give you almost the same results in much less time

btw, GA is much slower than BP

You can see the AutoMLP algorithm. It is a simple algorithm for both learning rate and size adjustment of neural networks during training. The algorithm combines ideas from genetic algorithms and stochastic optimization. It maintains a small ensemble of networks that are trained in parallel with different rates and different numbers of hidden units. After a small, fixed number of epochs, the error rate is determined on a validation set and the worst performers are replaced with copies of the best networks, modified to have different numbers of hidden units and learning rates. Hidden unit numbers and learning rates are drawn according to probability distributions derived from successful rates and sizes. For more information please be kindly referred to the following paper: AutoMLP: Simple, Effective, Fully Automated Learning Rate and Size Adjustment by Thomas Breuel, Faisal Shafait. The Learning Workshop, Cliff Lodge, Snowbird, Utah, United States, Online, 4/2010.

Thanks for your responses.

I got the idea of setting weights using Genetic algorithm for FFN but i am not getting how to do it for RNN(recurrent neural network) because we have a context layer in RNN which stores previous output of neuron and supplied for next iteration.

I don't understand what to store in case of GA because in back propagation we can store previous output but in GA every time we apply new set of weights from population so it will be totally new set, then how there can be a relation between previous output and this new set?

I think you have answered your own question. Your chromosome will need to have the previous weights as well as the current weights. Also, you will have to use some clever way of updating the weights. In this case mutation will be more useful than crossover. Then your weight update formula should be easier to compute. Hope that helps!

W ← (a./λ)*R by this equation you can weight matrix , R eigenector , λ eigenvalue a sesgo.

@Tanvi:- sorry but didn't got your answer. I have never use genetic algorithm in neural network so please explain in detail if possible. If i am not wrong we use to store previous output of neuron in context layer and apply it in next iteration.Let me tell you what i have understood (suppose i have 5-4-1 recurrent NN )

1) i can store weights for context layer to in chromosome(i.e from context layer to hidden)

2) i will apply a set of weight to network along with inputs to calculate fitness of that set.

3) according to fitness value i will reproduce population followed by crossover then mutation.

4) Again apply new population to network & repeat procedure until desired fitness is achieved.

-------

so i didn't get whether i have to store context layer o/ps in different array but after crossover the we get new chromosomes now how those previous output are related to new chromosome? I didn't got about weight update formula also?

I could be completely off but here's what i would try: your chromosome should have your current weights as well as the output of the previous iteration in the chromosome. For the GA to have meaning in this case, I would not try crossover but in mutation I would then store the current output in the location of the previous output. This is how you can store the information from the previous iteration. You can then use elitist selection to decide which chromosome goes forward since this technique can results in worse results. I will be very interested to hear the results since I have never tried GA with RNN. The "memory" aspect of the RNN can make it a tricky and interesting problem :)

Could you also state your problem? I mean why are you using RNN instead of a simple feed forward NN?

@Tanvi: I am trying for stock market prediction with RNN since RNN performs better in time series forecasting thats what i have read somewhere.( and I also think it will perform better).

Now as you suggest if we store previous output in chromosome and replace them with current chromosome but i think we have to do crossover on other part of chromosome to find out optimal weights otherwise network will not converge and fitness will not be improved. If you have any other suggestion please tell me so that i can think on it

And Thanks for your valuable responses.

You can use but using ga slow down your execution.

The best way to use ensemble