Using artificial intelligence at MTS - page 17

 

There's a lot written, but when I got to the end I forgot where I started :) Not many messages on the subject :(
Generally, it's the simplest one-layer neural network but one-layer network always gives a lot of errors, so the number of layers should be increased. For this task we need only two layers. And the first layer must have 4 neurons and the second one with one layer. A larger number of layers makes no sense and only loads the computer and complicates the task.
Of course, you can increase the number of input parameters and then increase the number of layers, but nothing good will come out :(.
In principle, if interested, I can write and lay it out. Test (I myself will not do it).
And for the future, I'll tell you right away - I do not intend to explain why there should be two layers!!!! And why the second layer should have one neuron and the first four. There are some clever books for that :). Read them and you will understand. I'm not going to enter into discussions either because it simply makes no sense.
The idea is certainly good, it is necessary to discuss it, but only write posts on a subject, and say something concrete, instead of philosophizing!

 
kazeev:

There's a lot written, but when I got to the end I forgot where I started :) Not many messages on the subject :(
Generally it's the simplest one-layer neural network but one-layer network always gives a lot of errors, so the number of layers should be increased. For our task we need only two layers. And the first layer must have 4 neurons and the second one with one layer. A larger number of layers makes no sense and only loads the computer and complicates the task.
Of course, you can increase the number of input parameters and then increase the number of layers, but nothing good will come out :(.
In principle, if interested, I can write and lay it out. Test (I myself will not do it).
And for the future, I'll tell you right away - I do not intend to explain why there should be two layers!!!! And why the second layer should have one neuron and the first four. There are some clever books for that :). Read them and you will understand. I'm not going to enter into discussions either because it simply makes no sense.
The idea is certainly good, it is necessary to discuss it, but only write posts on a subject, and say something concrete, instead of philosophizing!

Well, do it, I think it'll be interesting to read.
Just open a separate thread for it - there's too much unnecessary junk in this one.
 
If you're interested in playing around with neural networks, here's the second version of my expert...
There are comments on the code, I think you can understand it, although in some places the comments may not correspond.
I don't want to comb through it, because I am going to rewrite it very thoroughly, in order to develop it further.
Minimal number of layers in the networks is 2. The last layer has one neuron. The maximum number of layers is almost unlimited in the code, but as the number of layers increases, the learning rate falls catastrophically, up to a feeling that it is not learning at all. I don't think it's worth taking more than two layers. At least I haven't seen any difference in quality between 2 and 4 layers.
The current configuration consists of five independent networks with different inputs (indices+prices).
The nets are fully coupled, learning error propagation backwards.

In a pair of places there is a cunning code If anything, ask, but to very silly questions on the code I will answer only
if I have nothing to do :)

If anyone is interested where I will develop this further.
More clever data preparation, increase of network outputs, in general towards classification more than
rather than course prediction.


2Kazeev "And why in the second layer should be one neuron, and in the first four." The number of neurons in the output layer is far from obvious, what would say, that there should be only one :)
Files:
nn_expert.mq4  19 kb
 

2Kazeev "And why should the second layer have one neuron and the first four?" The number of neurons in the output layer is not so obvious, that it shouldn't be one :)

Because I'm based on that program, which was posted here. And the layers can be added as you want. But you need to look at the task you want to accomplish.

But with more layers, the learning rate drops catastrophically, to the point where it feels like it's not learning at all. It's hardly worth taking more than two layers. At least between 2 and 4 layers, I didn't see any difference in quality.

This is what I was talking about.

Your Expert Advisor is good, but you need to improve it! Well done!
 
maveric писал (а):
For those interested in dabbling with neural networks, here is the second version of my expert...

Thank you, it's not often that something interesting is posted. Not very profitable - but very informative.
 
kazeev:

There's a lot written, but when I got to the end I forgot where I started :) And there are few messages on the subject :(
Generally, it's the simplest one-layer neural network but one-layer network always gives a lot of errors, so the number of layers should be increased. For our task we need only two layers. And the first layer must have 4 neurons and the second one with one layer. A larger number of layers makes no sense and only loads the computer and complicates the task.
Of course, you can increase the number of input parameters and then increase the number of layers, but nothing good will come out :(.
In principle, if interested, I can write and lay out. Test (I myself will not do it).
And for the future, I'll tell you right away - I do not intend to explain why there should be two layers!!!! And why the second layer should have one neuron and the first four. There are some clever books for that :). Read them and you will understand. I'm not going to enter into discussions either because it simply makes no sense.
The idea is certainly good, it is necessary to discuss it, but only write posts on a subject, and say something concrete, instead of philosophizing!

Good afternoon. I would like to look at your modification of Reshetov's Expert Advisor (I think I'm not the only one). If you do not mind, please post it.
Sincerely, Pooh.
 
Ladies and gentlemen, what should we give for the entrance? A candle?
 
maveric:
For those interested in dabbling with neural networks, here is the second version of my expert...

When I ran the test, I got an error saying that the handle variable is null. Which is good because, judging by the code, the Expert Advisor should download the grid from a file but this file with the grid is not attached.

The Expert Advisor hasn't opened any deals and hasn't saved anything. Subsequent tests have yielded similar results, or rather the complete absence of results.

Do I know how to use it correctly or should I use it at all? Because understanding somebody else's code is a tedious and thankless task. And if you know the algorithm, it's much easier to write your own program from scratch than to understand someone else's.

Maybe it's just a bad joke of the author?
 
njel:
Ladies and gentlemen, what should we give for the entrance? A candle?
Respect to the author ...
Please formulate a network training rule in the forum.

--
rip
 
Pyh:
I would like to see your modification of Reshetov's expert (I think I'm not the only one). If you don't mind, please post it.

I don't think this code has anything in common with Reshetov Expert...