Neural network in the form of a script - page 8

 
Andy_Kon писал (а) >>

Consider the following situation:

The NS works, works, studies, studies, and then bang - someone Chubais (with a small letter) cuts off the electricity we need.

And all the work and learning goes down the drain (to Chubais).


The next introduction:

1. Periodically dropping (saving) the data of "learning".

2. In the case, as mentioned above, read this data during initialization of the Expert Advisor.


In this way we won't need to teach NS all over again.


of course! immediately after training

weights and other parameters will be exported to the file

---

when electricity is switched on, EA will load the file at startup



---

it is wise to retrain the network if the parameters are outdated - lagging behind the market situation

this is CALLED Adaptivity

or training

---

if you approach it differently (i.e. train once and go to market) then the network is unnecessary!

 
A quick question, at what frequency do you think the data should be dumped?
 
Andy_Kon писал (а) >>
Quick question, at what frequency do you think the data should be dumped?

after retraining

 

NS on CUDA:

http://quixotism.org/nnproj/index.php?title=CUDA_Code_Base

 
YuraZ писал (а) >>

---

it is reasonable to retrain the network if the parameters are outdated - lagged behind the market situation

it's called adaptability

or learning

---

Eh. What remains to be understood is when "the parameters have become obsolete" :( (The option of "slowing profit growth" and other "violations of key transaction parameters" is, IMHO, not the case, and, sometimes, it is too late).

It is approximately like with Trend/Flat - if you understand when one is over and when the other is over you do not need anything else (no indicators, no Expert Advisors ...). If we understand that "parameters are out of date", we just "tweak the parameters" ... For example MACD Sample.mq4.

:(

 

Found out where the legs are growing from.

http://richardbowles.tripod.com/neural/neural.htm

 
SergNF писал (а) >>

Eh. It remains to understand when "parameters are out of date" :( (The option with "profit growth slowdown" and other "violations of key trade parameters", IMHO, is not the same, and, sometimes, it is too late).

It is approximately like with Trend/Flat - if you understand when one is over and when the other is over you do not need anything else (no indicators, no Expert Advisors ...). If we understand that "parameters are out of date", we just "tweak the parameters" ... For example MACD Sample.mq4.

:(

maybe this is one of the complicated issues

actually there are a lot of questions

1-what do we want in the output

2-which circuit to choose

PNN MLP BPN etc.

( number of neurons in each layer number of hidden layers )

what kind of connections between them

feedback loops

( the circuit can be quite complex )

The networks that recognize printed letters are trained once and the circuit is not very complicated

3 - what to give as an input (it is clear how many inputs there are)

4 - when to retrain

5-which learning algorithm to choose

GA - genetic

BP -


---

some questions

1-output I want 3 values

1-sell 2-don't know 3-buy

2-PNN

scheme in process of experimentation

3 ?

as an acquaintance said - go INTIM

e.g. difference between EMA at reaching a threshold by some indicators

4 ?


come to some values that were not predicted ( the network has exceeded the allowable number of errors )

got a series of LOS or TAKE triggered ! that also went outside the grid forecast

got some thresholds - so we start looking for a wave in the opposite direction

I think to run on schedule - not reasonable - like after a new bar

For example, networks that recognize printed letters are usually trained once - for specific fonts - ours is not like that


5 ?

I am leaning towards GA scale.


---

 
Where can I read about PNN, MLP, BPN i.e. neural network diagrams?
 
ShestkoFF писал (а) >>
And where can I read about PNN, MLP, BPN i.e. neural network circuits?

I can tell you that there is everything on the internet!


Of course, it's not very nice to send a search engine

---

look here-- I've posted a lot of links

 

How important is accuracy for sigmoid?

After 20, the accuracy is in the 9th digit...