How to create Neural Network with Mql5?

[Abhishek Yadav]

Hey, my specific question in creating a neural network is, how do you train a neural network in mt5? Is it by optimizing all the weights that goes into the neuron? If it the case we cannot say it as learning, instead it is a deep optimization. Also in mt4 the optimization cannot handle large number of optimization of weights. All the experts on neural network, please respond. I want to create an ea, in which when we first start the ea it will automatically learn for the period and give a learned string. Then when i put learned string in the inputs, it will trade like that. Do anyone have any idea? If this is not possible way how can i train the neural network? Thank You in advance for your responses.

[Lorentzos Roussos]

Abhishek Yadav:
Hey, my specific question in creating a neural network is, how do you train a neural network in mt5? Is it by optimizing all the weights that goes into the neuron? If it the case we cannot say it as learning, instead it is a deep optimization. Also in mt4 the optimization cannot handle large number of optimization of weights. All the experts on neural network, please respond. I want to create an ea, in which when we first start the ea it will automatically learn for the period and give a learned string. Then when i put learned string in the inputs, it will trade like that. Do anyone have any idea? If this is not possible way how can i train the neural network? Thank You in advance for your responses.

It is a function approximation , whether it is "learning" or not is equivalent to your experience in school as a child.

You had the kids that memorized and recited from memory and the kids that understood (not comparing with history class where you could only memorize)

I've tried a neural net with the weights exposed as inputs and let the mt5 genetic algorithm optimize the weights . (i'll link the thread when i find it) (edit : Link)

You could equate this process as spraying a "population" of networks on a surface and seeing which ones survive based on what is considered "survivability" (which you can define)

Then the tester ,based on your "survivability" , "breeds" the best fit networks and mutates some to create newer generations .

The conventional way is to "teach" it directly from what its supposed to do.

Think of a network as a function , the function outputs a forecast , you provide what the correct answer was , it calculates how wrong it was (and where in its connections) updates itself and tries again.

To train it using the first method (with the mt5 genetic algorithm) you get a limitation on your maximum number of connections simply because the tester cannot process that many possible combinations .

To train it using the second method you could pull the problem elsewhere or use a custom approach where you'd collect the data , split it in parts and what you'd be interested to see at first is how well the "educated" network up to a point performs to the following time slot (s).

I also have a silly barebone oop neural net that may help you understand what is going on , when i find that i'll link it too.

(edit : Link)