Machine learning in trading: theory, models, practice and algo-trading - page 396

 
elibrarius:
Is Reshetov's classifier and that program still a single neuron, not a network? Or are they combined into a network of Reshetov neurons?

Honestly, I don't even know how it works. I think it goes like this. The sample is divided into two parts - a test part and a training part. One grid learns on one part, and is tested on the second. Another on the contrary is trained on the second, and tested on the first, then the result is summed up and the overall result is calculated, so that is IMHO
 
Mihail Marchukajtes:

I think the author will not be able to get through, I wrote him and so and so. Silent. But as far as I know, he once wrote that it was paralleled in it all you can, these are his words. Yes indeed he is training two grids that work in the committee. I wrote that in my article. If both show yes, then yes, if not, then no, if they both show in a mismatch, then "I don't know". I do not know about the latest version, but the basic description is on google, the link that you gave me. Once ran version 3 on the VPN server and to my disappointment the optimizer loaded only one core, but recent versions load all the cores evenly, so I think there is still razparalelivanie. There is only one thing left to do. To increase the number of cores :-)

I have written to Yuri here at mql, maybe he will answer me :) I don't buy this approach with 2 nets, why it is so? On google site there is no description of innovations, only about Brown Robinson Reshetov Algorithm
 
Maxim Dmitrievsky:

I wrote to Yuri here at mql, maybe he will answer me :) In general, I do not buy this approach with 2 networks, why is it so, on his google site there is no description of innovations


Well, yes. There is only the old version and the basic approach. But as practice has shown, the 2-grid approach increases generalizability significantly. The result of the optimizer's work is the following file. Here you can see two grids and for each grid a different normalization, and then the results are combined at the very end.

So Reshetov made a good product, you shouldn't have criticized him if honestly......

Files:
HARD.mqh  7 kb
 
Mihail Marchukajtes:


Well, yes. There is only the old version and the basic approach. But as practice has shown, the approach with two grids increases the generalization ability significantly. The result of the optimizer's work is the following file. It shows two grids and a different normalization for each grid.

So Reshetov made a good product, you shouldn't have criticized him to be honest ......


Did I say that? There was no such thing... yes, I've seen this file
 
Maxim Dmitrievsky:

Was I praising? There was no such thing... yes, I saw this file

I'm not saying that it was you, I'm talking about the inhabitants of this thread in general......
 
Mihail Marchukajtes:

I'm not talking about you, I'm talking about the inhabitants of this branch......

there are a lot of trolls in all the threads :)
 
Anyway, the question now is this. If there is a lot of code and need to move in, and it is time. Maybe we will try to run the optimizer on GPU???? If performance will increase, the program is unparalleled and there is no need to change anything in it. Let's try to run the optimizer on the GPU????
 
Mihail Marchukajtes:
Anyway, the question now is this. If there is a lot of code and we need to move in, and it takes time. Maybe we will try to run the optimizer on GPU???? If performance will increase, the program is parallelized and there is no need to change anything in it, and to rewrite it, as they say, to taste and color. Let's try to run the Optimizer on the GPU????


It's not easy, for gpu it's necessary to write kernels separately, take out part of code. I was thinking about rewriting it in mql, it would be easier to implement, because I'm not so good at Java. But it's almost unreal to rewrite a new version, especially without knowing the algorithm.

And send me some simpler set for his predictor, which is quickly classified, I'll compare with other MO models the results. If it's really cooler on all fronts, I'll find ways to make it on gpu

 
Maxim Dmitrievsky:


It's not easy there, for gpu it's necessary to write kernels separately, to take out part of the code. I was thinking about rewriting to mql, it would be easier to implement there, since I'm not very good at Java. But it's almost unreal to rewrite a new version, especially without knowing the algorithm.

I'll try to use some easier set for Jpredictor, that quickly classifies, and I'll compare results with other MOE models.


I have a shortened version, this is a set the result of which I posted just above.
Files:
HARD.txt  49 kb
 
Mihail Marchukajtes:

I have a shortened version, this is the set result of which I posted just above.

Ok, good, later today I will send you comparative tests