How to build a NN-EA in MT4 - page 4

To add comments, please log in or register
Sergey Golubev
Moderator
114230
Sergey Golubev  

I don't know, sorry.

may be, train it for EURUSD and after that place the trained Vasya;64;64;7 from /tester/files to some place so place it to expert/file for forward testing if EURUSD. In this case we will have many files ...

I don't know sorry. As I said i am not very professional with NN.

Sergey Golubev
Moderator
114230
Sergey Golubev  
omargray:
What if you train it on more than one set of data at a time? Like training it on the EURUSD 5m and also the GOLD 5m since the movements of the two correspond somewhat? Will having a wider range of data and more patterns create a smarter NN EA in this case?

More data - more good

Really don't know. Even in the theory. I will need to read something about NN during the weekend.

Sergey Golubev
Moderator
114230
Sergey Golubev  

Because you know ... this December is not usual case for every month concerning volatility so I think may be we need to train it every month, or for many years (starting since 2004 for example)... and I don't know about 90% every tick: do we really need to bacltest/train this EA with 90% modelling quality if we are backtesting just for training only?

I need to read something.

changeea
9
changeea  

Hi this is my report.

experts\libraries\NS2-32.dll

KOHENENN v15.mq4

KOHENENN v15.ex4

c:\Koh_2007.def

BY

Files:
kohonenn.gif 5 kb
kohonenn.htm 12 kb
Sergey Golubev
Moderator
114230
Sergey Golubev  
changeea:
Hi this is my report.

experts\libraries\NS2-32.dll

KOHENENN v15.mq4

KOHENENN v15.ex4

c:\Koh_2007.def

BY

It is the other EA?

It should be trained I think.

What if you train it on more than one set of data at a time? Like training it on the EURUSD 5m and also the GOLD 5m since the movements of the two correspond somewhat? Will having a wider range of data and more patterns create a smarter NN EA in this case?

I think it's possible. Because this NN is trying to find same case which was some time ago. Thus it may be good to have as many cases as possible.

bolt
39
bolt  
newdigital:
As to this post.

Place NS2-32.dll to C:\Windows\system32\

As to Koh_2007.def file so I am not sure: may be c:\Koh_2007.def, or create folders and place: C:\NeuroShell 2\EXAMPLES\Test\Koh_2007.def

I have no idea, sorry.

yes stick in your c drive ie c:\Koh_2007.def as the EA refers to this location but as far as i can tell this is a wieghts file and assumes you have already have access to neuroshell where you perform your own optimization first then when you have a great forwards and backtest generated in neuroshell you copy the wieghts file and paste it into the koh-def file which then tells the EA what to do as i dont think it can do it itself. Im trying the neuroMTC ea but having trouble to get the tester file to update. I tried deleting the contents etc but it still wont write new wights to it so im stuck.

tiger_wong
150
tiger_wong  

Hello forum,

I am very interested with new idea of NN EA. I think this topic will be the next generation of automated trading. Even tough I am a novice in programming, I would like to suggest to make conditions for building a NN EA. If someone has some opinions, let write them below these.

The criteria for NN EA :

  1. Have one or more NN Function (of course, LOL). It can be Baysian Filter, MonteCarlo, Kalman Filter, or maybe others.
  2. Have a function to collect its own tick data. I think this one is important because if the EA uses data from the OCHL chart from the broker's server, it is useless since the data has already been interpolated.
  3. Have a function for automated optimization (this function could be found at Automated Optimization of a Trading Robot in Real Trading - MQL4 Articles). It is important to get this function as we know that price movement never always been same as history even it is similiar. So, this EA must be always optimized with the latest data that collected from criteria #2

PS : You can add your own opinion / criteria

bolt
39
bolt  

The best way to avoid all the issues regarding ticks, broker data and odd pips is simply to ignore them and take all signals to nearest bar close instead. So decisions are made on the close of 5 min or even 15 min bars even though NN is fed with larger timeframes. Clearly as been the case for sometime now the best performing EA's as far as MT contest anyway use longer timeframes and may only trade 3 or 4 times a week. Very few scalpers survive past the first fortnight so little point in taking decisions sub 1 minute when 5 mins is good enough. Also very important the fact that data intensive NN's with more than 3 neurons and backtesting is so CPU intensive to the nearest 1 min tick will see stupid backtest delays of 100+hours.

ruparaw
117
ruparaw  
newdigital:
The first EA they created one year ago was #_e_NeuroMTC.

In attachments:

- place #_e_NeuroMTC EA to expert folder;

- place #_i_e_MTC_Etalon indicator to indicators folder;

- DT_ZZ indicator - to indicators folder.

- place Vasya;64;64;7 to experts/files and to ..../tester/files

- this file #_lib_NeuroFFT - to library folder.

This EA is for M5 timeframe.

How to train this EA.

Backtest this EA in visual mode (every tick), attaching 'etalon enter' indicator (#_i_e_MTC_Etalon) to visual chart. This indicator is a trainer. Later you can deattach this indicator if you are sure that EA was trained enough and Ea will continue backtesting without etalon.

Then you can forwatrd test EA on trained pair. M5 timeframe.

Sory, I am not a coder so it is very difficult for me to explain something which I don;t understand I just translater the meaning of the first 2 pages of this thread.

I am in the middle of to train #_e_NeuroMTC EA by backtesting using m5 tf visual mode from Nov 8th 2007. As so far about 1 hr already the testing just reached to Nov 23rd 2007....yet not completed. below is the progress so far.

Can i know what was the reason behind we have to train this ea using visual mode. It is possible if we just run through normal backtesting then paste Vasya;64;64;7 to experts/files from..../tester/files

Sergey Golubev
Moderator
114230
Sergey Golubev  

I don't know ...

It is written by author that training is doing by backtsting in visual mode by attaching etalon indicator to the visual chart.

I am think about how tom improve this EA. Because this Ea is based on zigzag indicator and some other indicators. But is it possible to use this EA as a template to create some other EAs? So, i am trying to understand how it works.

To add comments, please log in or register