Using artificial intelligence at MTS - page 7
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
"...But the culprit is not the giraffe,
"but the one who shouted from the branches
"The giraffe is big, he knows best..."
(c)
To eugenk1 (about the super-index) "classification by means of Kohonen Maps" (for example). But optimizing (collecting statistics) on every tick is a little bit boner...... not smart :)
To Demax ("... it seems that you can't write it in mql...") Well, why http://www.fxexpert.ru/forum/index.php?topic=656.0. Especially if you don't "reinvent the wheel", but at least understand the terms (at least "read" the pictures). But more correctly the last posts in the link about connection of "neural networks in dlls".
to eugenk1 ("... it doesn't matter what colour the cat is, as long as it catches mice...") IMHO, that's the problem. IMHO it would be nice at least read the definitions of terms that you throw yourself and your opponent wrote. For example for this subject here http://www.neuroset.ru/index.php?option=com_docman&Itemid=79
And the application of filters, better start with filtering incoming and outgoing information by a "personal neural network" AKA brain. :) Especially when "Academically/Pedagogically Intolerant of others' positions" are involved in the discussion. :)
to Yurixx (the rest of participants about the 0-bar from the "I hear a bell, but I do not know where it is" series). The essence of "training" a neural network is to pick up the coefficients of the neural network, but not the parameters of a super-duper-granular inductor. You should see how neural networks in "academically rigorous black boxes" are "trained". They either take inputs shifted to the past and calculate an error between the "predicted" current bar and the actual current bar, or based on inputs from the current bar and calculate an error between the future bar and the "predicted" future bar. But for the "suckers" (those for whom "did the drums") there is a "visual shift of the prediction by the length of the predicted bars to the right" in these programs.
I'm against the 0 bar as well. And also because, scared of MT3, not because I understand what I want, but for this Expert Advisor there is no point in looking at "exorbitant profits".
And what I liked in Reshetov's approach to MT4 and NEUROSETS, is his connection to the chain (which has yet to be implemented!!!) Learning->Training->Walk MT4 and using the built-in GA.
For the highbrow snobs out there :) no one is stopping the input vectors (including the 0-bar) and the network architecture (Hopfield, Grossberg and a bunch more "Angola-letter abbreviations") from being replaced as well. The single layer perceptron, on the other hand, is simply the "lightest network possible" plus its "simplified implementation" in terms of "neuron activation function".
I myself am expelled from 3rd grade parochial school, so your highly scientific mate towards me will be useless. :) I'm trained to filter inputs. ;)
Yeah. I'm afraid the "smart guys" have once again "stabbed the topic" :(. I'm afraid the "smart guys" have once again "cut off" the thread :(. The "enemies" have even killed the "bumsy" forum http://bamsi.e7e.info/index.php, the "market formulators" http://fx2you.com/ are also amateurs, while the "specially trained people" http://www.neuroproject.ru/forum/ or http://www.basegroup.r u/ are either up to forex or free communication .... far away. :(
You've got the man all beaten up, evil people! Pay more attention to the rationality in all this! And it's there. And it's not insignificant...
I'm afraid it's not people like you who are being slaughtered, but ideas. Reshetov doesn't put his finger in his mouth either. :) Such is the essence of your Genetic Algorithms. :(
High and Low are used to calculate AC, and where do we get them at the bar opening
And then as the cards fall.
Ie.
Suppose the first tick is at the bar opening and the price is 1.9722
Here we have
Open = 1.9722
High = 1.9722
Low = 1.9722
Close = 1.9722
Next tick, the price is equal to 1.97721
Now we have
Open = 1.9722
High = 1.9722
Low = 1.9721
Close = 1.9721
And so on.......
What future are we talking about?
Why, as a matter of fact, did this question arise? Because AC (as well as others) is drawn on the opened bar, not on the already formed one, i.e. it has a definite value. It is clear that this value will be constantly changing until the bar is formed, so what? This is the problem?
About the Kohonen maps. This is probably the only neural network topic that I've more or less mastered, at the very least with a C-minus. Yeah, I agree. Classification is a very important step. But alas, I have no idea what to do with it all next... No matter how you classify, no one guarantees you similarity. The same turkeys of tehanalysis with the Potik collecting history, it's at least more or less clear what's what. What classification with Kohonen Maps is much less clear... I read fxexpert.ru of course. It's a great site! It's a jewel of forex resources. I am the Black Cat there.
What future are we talking about?
Why, as a matter of fact, did this question arise at all? Because AC (like others) is drawn on the open bar, not on the already formed one, i.e. it has a certain value. It is clear that this value will be constantly changing until the bar is formed, so what? This is the problem?
The problem is that we actually discussed testing Reshetov's Expert Advisor in a tester, not in real time.
In the tester, in "Open price only" mode, it does not work correctly. In this mode, ticks are not simulated and the Expert Advisor receives a complete set of OHLC immediately at transition to a new bar. As a result, it has a possibility to look 30 minutes into the future - this is exactly the AF Reshetov used to test his Expert Advisor on.
Thanks for the clarification, now I finally know where High and Low come from.
At the same time you finally learned something new about the tester and testing.
Very useful communication ! :-)))
PS
I suggest now, for the sake of full enjoyment, run this Expert Advisor in the tester (with the same settings) in the "All ticks" mode and compare the results with what Reshetov posted. I think you will like it.
2 SergNF
I'm against the 0 bar too. And also because I'm scared of MT3, not because I understand what I want, but for this Expert Advisor it makes no sense to look at "exorbitant profits".
That was a strong passage to me, but I don't understand what you want either. Reshetov used the results of the tester as an argument in favour of his "neural network". The results are flawed, as is his tone. For lack of other arguments the whole discussion revolved around this testing. And the size of these "profits" nobody here is interested in.If he had at least tried to raise the issue of testing the way it is done in "academically rigorous black boxes", the conversation would have been much more interesting even without the testing itself and would have gone in a different direction. As it is ..., nothing interesting.
But you are still dissatisfied, although you do not know why. My dear, it's the same ringing!
Yrixx, I've already raced on all the ticks. I've already raced all the ticks. Look it up in codebase if you're interested. Athlon64 3200, with a gigabyte of RAM. But I got noteworthy results there too. Alas, it is not very stable. I mentioned it in codebase too.
Yes I have seen it. Until May it drains, then it pours. :-)
Strange results, I must say.
It is possible to check in another way. Increase parameters of all 4 ACs by one and test them by opening prices. This will be correct and not so long as for all ticks. Try to do it. If you do, share the results here, pls.
SergNF, you've given them all away, no one is forgotten ! :))))))))))))))))
About the Kohonen maps. This is probably the only neural network topic that I've more or less mastered, at the very least with a C-minus. Yeah, I agree. Classification is a very important step. But alas, I have no idea what to do with it all next... No matter how you classify, no one guarantees you similarity. The same turkeys of tehanalysis with the Potik collecting history, it's at least more or less clear what's what. What classification with Kohonen Maps is much less clear... I read fxexpert.ru of course. It's a great site! It's a jewel of forex resources. I am the Black Cat there.
You take your favourite inductor (if external, through iCustom) and output to the file the value of the inductor for some amount (how much you want to predict in the future) BACK and, there are options, Close/High/Low "in the first bar" or Highest/Lowest for this interval. You can analyze and think how to apply it by reading articles from http://www.tora-centre.ru/library/ns/spekulant04.htm, http://www.tora-centre.ru/library/ns/spekulant03.htm in parallel.
My fxexpert.ru takes too long to load + it was down for 3 days. And the constructive mostly comes from klot.
To Yurixx Sorry, confused my nickname. Beads will be collected. :)
"...The problem is that there was actually a discussion about testing Reshetov's expert here... ". Here, in fact, the technology (albeit very simplified, as Keny/Uncle put it - "metatrading") inherent in his expert was discussed. In the title of the thread, by the way, there is neither expert, nor tester, and generally speaking, artificial intelligence and neural networks are not the same thing. Yes and it doesn't prevent the subroutine double perceptron(), "twist it to your liking.
"Toincrease all AC parameters by one". Listen, do you seriously believe that this expert is the quintessence of the entire universe. We need to do research - what depends on what, with what parameters, etc. etc. Why 0/7/14/21, why AC, why 4 input neurons, why such (no) activation function. In pure "neural networks" we should look for parameters to minimize prediction error and only their (NS) application for "tradons" allows profit optimization....
In short, let's go gather beads :)