Discussion of article "Third Generation Neural Networks: Deep Networks" - page 15

 
Carl Schreiber:

Hi,

please help me to clarify some my negative prejudgements about neural networks (NN).

  1. Is it correct that you should firstly optimize the indicators to be in putted into the NN?
  2. Then you optimize the parameters of the NN?
  3. Or do you optimize the parameters of the NN and the indicators at the same time?
  4. Isn't it true that the more variables you have to optimize the greater is the danger of over adapting?
  5. If the data-sets for 1. and 2. are the same wouldn't that lead me to a kind of over adapting to the data set?
  6. Isn't exactly that indicated by "Again we have a profitable phase of about 5 weeks until the model deteriorates."
  7. a) Lets assume we have a bunch of indicators all together optimized by the tester and now
    b) we run a second optimization by the tester only to check which of the optimized indicators we need(*)
    c) so that we have a smaller bunch of our optimized indicators
    d) for what do I need the NN?
  8. Do you know an estimation about how big the data set has to be for a NN due to the number of inputs, layers and perceptrons?


(*) Unfortunately if you run mt4' optimizer in genetic mode and you want to try to bypass certain parameter sets (e.g. don't test if "indicator-A" is 'on') by returning from OnInit() with "INIT_PARAMETERS_INCORRECT" the genetic algorithm still counts this as a valid pass and that reduces the number of actually executed passes before this algorithm stops due to the number passes which is one of termination criteria.


1,2,3 and 4 , i believe whatever indicators and settings are passed in , inherently adjust to the underlying asset.

For example , lets say we create a simple optimization using the RSI and ZigZag Highs , ZigZag Lows .
We produce an average oversold on highs by summing the RSI value at the ZigZag Highs , and an average overbought
on lows by summing the RSI value at the ZigZag Lows . Our averages will essentially be the adjustment of RSI regardless
of settings to that asset .
The question is not if indicators should be optimized in my humble opinion ,but whether or not the indicator is utilizable 
fundamentally.
In the above example you can grasp my point by viewing the Averages for a RSI(3) versus an RSI(16) . 
The RSI(3) will constantly trigger our optimized levels versus the RSI(16).
 
Lorentzos Roussos:
1,2,3 and 4 , i believe whatever indicators and settings are passed in , inherently adjust to the underlying asset.
...
Our averages will essentially be the adjustment of RSI regardless of settings to that asset .
The question is not if indicators should be optimized in my humble opinion ,but whether or not the indicator is utilizable 
fundamentally.

In the above example you can grasp my point by viewing the Averages for a RSI(3) versus an RSI(16) . 
The RSI(3) will constantly trigger our optimized levels versus the RSI(16).

As your example - if I understand it correctly - tells me RSI(3) is of no help as it will not distinct between 'good' (potential profit > ??) and 'bad' (potential profit < ??) but RSI(16) does.

But if so there has been an optimization as after that we know 16 is better than 3 - or from where do you know that?

Now do you train the NN with RSI(3)? It will be deleted probably. Or are you trying RSI(3) (NN-input 1) and RSI(16) (NN-input 2) and if RSI(3) will be deleted (NN-input 1 is set to 0 e.g.)  RSI(x) has been optimized to 16 - even in very simple way. Do we need an NN for this having the MT-optimizer?

Or am I missing something in your example?

 
Carl Schreiber:

As your example - if I understand it correctly - tells me RSI(3) is of no help as it will not distinct between 'good' (potential profit > ??) and 'bad' (potential profit < ??) but RSI(16) does.

But if so there has been an optimization as after that we know 16 is better than 3 - or from where do you know that?

Now do you train the NN with RSI(3)? It will be deleted probably. Or are you trying RSI(3) (NN-input 1) and RSI(16) (NN-input 2) and if RSI(3) will be deleted (NN-input 1 is set to 0 e.g.)  RSI(x) has been optimized to 16 - even in very simple way. Do we need an NN for this having the MT-optimizer?

Or am I missing something in your example?

Reffering to RSI(3) and RSI(16) as an example of possible fundamental utilization gaps in real time.
The ideal would be a variable period RSI in this example 
 
Lorentzos Roussos:
Reffering to RSI(3) and RSI(16) as an example of possible fundamental utilization gaps in real time.
The ideal would be a variable period RSI in this example 

ok - so what is sent to the NN?

RSI(..) with a fix value (how did get it) with a variable value - can one optimize the calculation or not?

This all influences the danger of over adapting - therefore sorry being so nasty.

 
very good! but for me, it's very difficult!
 
Is there an English version from the resources you attached ?
 

Ошибка

When running in RStudio:

 >dt.b<-Balancing(dt)Error in Balancing(dt) : could not find function "upSample"

What is this function? Which package is it from and where is it defined?
Thanks!
 

I have R x64 3.3.1. After installation the following libraries were missing - svMisc, svSocket, TTR, xts, zoo. But Rstudio did not complain about the last three, I managed to find out only thanks to DebugView.

The indicator is installed, thinks for a long time and produces zigzags. When trying to set serv to true, it crashes:

[8904] <-1> TPlotEventLoop: terminating
[8904] <-1> TRConsole: destroying

The same thing happens when installing an Expert Advisor:

[10964] <-1> TPlotEventLoop: terminating

The terminal says "Rterm crashed".

I have not found anything clear about this error in Google. Where to dig?

 
Konstantin Kopylov:

When running in RStudio:

I apologise for the late reply.

The function is defined in the package "caret::upSample //downSample will randomly sample a data set so that all classes have the same frequency as the minority class. upSample samples with replacement to make the class distributions equal//

Good luck

 
m0rtal:

I have R x64 3.3.1. After installation the following libraries were missing - svMisc, svSocket, TTR, xts, zoo. But Rstudio did not complain about the last three, I managed to find out only thanks to DebugView.

The indicator is installed, thinks for a long time and produces zigzags. When trying to set serv to true, it crashes:

The same thing happens when installing an Expert Advisor:

The terminal says "Rterm crashed".

I have not found anything clear about this error in Google. Where to dig?

In the appendix to the article I have posted a revised e_DNSAE Expert Advisor without using the server.

Please take a look at it.

Good luck