Examples: Recipes for Neuronets - page 2

New comment
 

All Great stuff! Compliments for article.


But something still unclear.  Wikipedia sure give you a basic and very understendable concept

of neural, and the codes in the article are very clean and plain.


But i think you should give us an help about the files attachment to download.


!CreatePatternNeyro.mq4 what do exactly? It seems to not be an indicator nor an ea; it is a script?

What it do? It runs in conjunction with files !NeuroInd.mq4 and ClassesC++ stuff?


!NeuroInd.mq4 is clear is the indicator. But some trouble and strange thing happens

for me about this indicator. First of all, if i compile it from Opening the file with MetaEditor,

after compiling, i cannot found the indicator in the MetaTrader Indicators, but i think

because the compiled file is on the folder where i downloaded it (in example on a Desktop folder).

Where to copy the compiled !NeuroInd.ex4 file in order to have it into the MetaTrader Indicators?

This is off-topic a bit: i cannot find on the hard disk where my custom created indicators are;

if i create an indicator directly from MetaEditor, in example name IndicatorTor.mq4 IndicatorTor.mq4;

with Windows search files function, i cannot be able to find it. Where ALL the stuff compiled from MetaEditor

has been stored in the hard disk? Which path? Which folder?


Since this "trouble" i can solve the problem of !NeuroInd indicator by just opening the downloaed

file with MetaEditor, than create another new indicator, copy the code from the !NeuroInd.mq4 and paste

in the indicator new, i.e. NeuroIndcus; just compile and i found it into the list of indicator of MT4.


But this indicator seems be a little different from the indicator in the article.


My indicator does not have "arrows" on candles, and it oscillate only upper and lower, hardly, and not

in the article figure, with tiny oscillations upper, than a "fall" of indicator (indicating price move bearish),

or the tiny bottom obscillations and then a raise of indicator (indicating the price will move bullish).

Here an attachment of what is my indicator:


As you can see it differs from the indicator example of the article(below), and is impossible to forecast prices move with

this graph.

---


Image of the article:




Thanks for any help or clearings.



 

FileName= "C:\\Program Files\\MT4\\experts\\files\\iles\\MA25_15.bar";

FileName= "C:\\Program Files\\MT4\\experts\\files\\iles\\MA25_15.bar";

 
FileName= "C:\\Program Files\\MT4\\experts\\files\\MA25_15.in";
file.Open(FileName, CFile::modeRead);
// 麒蜞屐 玎泐腩忸?羿殡?
file.Read(&nBAR, sizeof(int));// 麒耠?犭铐钼
file.Read(&nIN, sizeof(int)); // 疣珈屦眍耱?怩钿?
file.Read(&nOUT, sizeof(int));// 疣珈屦眍耱?恹躅溧
// 徨疱?桤 羿殡?怩钿?
for (i=0; i<nBAR; i++) for (j=0; j<nIN; j++) file.Read(&NN.in[j][i], sizeof(double)); 
// 徨疱?桤 羿殡?恹躅潲
for (i=0; i<nBAR; i++) for (j=0; j<nOUT; j++) file.Read(&NN.out[j][i], sizeof(double));
file.Close(); // 玎牮?羿殡 溧眄

FileName= "C:\\Program Files\\MT4\\experts\\files\\iles\\MA25_15.bar";


 

Hi,


Very insightful article! Many questions arise (many basic for users like me :D ), however:


1. How does the Kohonen relate to financial time series?


2. What is the relationship between input values, output values, and input vectors - how are input vectors made? This is a little confusing.


3.

Is the purpose of the Kohonen layer’s classifying input vectors help find patterns better? But how does this work if only one kohonen neuron takes the inputs and the remaining kohonen neurons give output of 0 (winner take all method)?


4. How to normalize input vectors if the input values are constantly changing in a time series (because the max and min values are always unique)? Even doing this with the initial training of NN would be useless because the continual, real-time learning’s max and min values are not the same. Wouldn't normalizing input vectors be harmful since the min and max values are always different in live, continuous learning?


5. How do you find the optimal number of Kohonen neurons? Experimentally – does this mean we try trial and error to minimize error and minimize epoch (training time)?


6. Which is better – the random initial kohonen weight assignment, or the method of convex combination? In the eqn, what are the super and subscripts? How do the eqns relate to actual shifting the kohonen weights?


7. You mentioned this was the basic Kohonen version: what are more advanced versions and does it help with financial time series? If so, in what ways?


8. In your listing 2, what are the output values?


9. Do the kohonen weights shift with the actual network weights during training (in both real time and initial training)?



Thanks, I hope my questions are clear!

 

Hi


I am trying to make a compile but I have


error C2668: 'sqrt' : ambiguous call to overloaded function
d:\1\vc\include\math.h(581): could be 'long double sqrt(long double)'
d:\1\vc\include\math.h(533): or 'float sqrt(float)'
d:\1\vc\include\math.h(128): or 'double sqrt(double)'

while trying to match the argument list '(int)'


Can someone help me?

Thanks

 
I think I found a small error in Fig.1. The output of K2 neuron should be defined as follow: K2=X1*W21+X2*W22.
12
New comment
Reason: