Using artificial intelligence at MTS - page 3

New comment
 
Itso:
Really, Reshetov is good, it seems that so far no one has made neural networks in MQL4, they just talk and brag about it.


Caesar posted a neural network in CodeBase - The Self-Learning Expert
 
Sorry - overslept - my apologies.
But still, I don't deny my reasoning.
And yet Rosh - how did it all end?
 
Not at all. You still need to grow to the necessity of using neural network (this conclusion I made for myself). That is, first you need problem definition (formalization of strategy, etc.), and only then mastering a tool (using neural network).

One more thing. There is a strong suspicion that my neural network (my head) at the moment is cooler than any artificial one, which I can create at the moment. Decided to return to this topic later, when the need arises. So far, the need has not been felt.
 

Dear Reshetov, I did not mean to belittle your achievement. Rather, I wanted to understand the essence of your Expert Advisor. I am firmly convinced about optimization: it should be conducted inside the Expert Advisor to make it a full-fledged neural network. My question about perceptrona was misspelled. I will try to rephrase it: why do you use linear combination of AC (plane) instead of conditions if(a1<x1 && a2>x2 && a3<x3 && a4<x4) that describe a polyhedron. Let me try to draw an analogy. Suppose I put a hat on when it's below zero outside. I want to predict if I will put a hat on tomorrow or not. That is, I should predict the temperature tomorrow: if it is below zero I will wear it. With your system, I take today's temperature and multiply it by 35. I add to the result the temperature a week ago, multiplied by 27. Then I take the temperature a fortnight ago multiplied by 84 and I take the temperature three weeks ago multiplied by 7. While the temperature makes perfect sense (as does the AC), the result of its linear combination described above loses its meaning. Of course I can adjust the coefficients of the model so it will predict temperature for tomorrow with some probability. But I think it is better to use conditions that make some kind of physical sense. For example, if today's temperature is below zero, and yesterday's was above zero, then we have a trend of temperature decrease, and it is possible that tomorrow's temperature will also be below zero. You can also add other factors (indicators) which influence tomorrow's temperature. For example, if today's temperature is below zero and it is cloudless, tomorrow's temperature will probably also be below zero. If we abandon this analogy and proceed to forex, why not select several different indicators that measure price movement and impose to them conditions such as if(IND1>x1 && IND2>x2 ...). The vast majority of Expert Advisors are built in this way. But there are very few Expert Advisors that are able to self-learn (adapt), i.e. to optimize x1, x2 ... in real life.

By the way, I also have a little experience of creating an Expert Advisor on a neural network. It was built by the Nearest Neighbor method. It was a lot of calculations but was of little use. I eventually abandoned it.

 
Yep, there they are, the true motives of gpwr's original bewilderment as to the choice of indicator supplied to the perceptron input. I quote my hunch:

Mathemat wrote (a):
Yuri, come on, why get so emotional. The question was almost certainly about the hidden meaning of this filtering, not the interpretation of the result... Roughly speaking: why this particular filtering? As applied to the trading system, it may not be the most adequate question, but you can try to reason with your choice - why AC and not some IACD...


Gpwr, you have already got the answer a page earlier: do not look for a hidden meaning in neural networks; look for rationality in the number of optimisable parameters and, of course, in the final results of the system and its more or less acceptable statistical validity. After all, many trend-tracking systems are based on muwings. Man likes to see the world around him as smooth rather than fractal, as a smooth world seems more predictable to him.

2 Rosh: I like your idea of three parallel neural networks, each trained on a different part of the graph, but I, also frustrated by the results of training my primitive NS, would prefer GA over NS. McCormick seems to think GAs are more promising than NSs in his encyclopedia of trading systems...

And in general, a normal system that pretends to work on any part of the chart must be adaptive to account for its own disasters. Roughly speaking, the perceptron weights in the branch author's EA should somehow adapt to the state of the market.
 
I will digress from the topic a little bit, although it is also quite in line. While driving from work today, it occurred to my dumb brain that we should all reconsider our attitude to indicators. It is in the perspective of their use in neural networks, but not only in them. In short, I want to start from the idea that an indicator is not a small tool for screen decoration, but a tool that helps to trade. In my opinion, the best way to help this process is to estimate the probability of price movement up or down by a certain amount of points without any wisdom. Therefore, let us consider that the indicator is a number (or rather the function of price series) that changes from +1 to -1. The sign of this number shows the supposed direction of price movement - '+' up, '-' down, while the module - the probability of reaching a significant amount of points in this direction, for example 30 (it will be better to make it an obligatory indicator parameter). I.e. all indicators have a unified and uniform interface. What they have inside them is entirely up to the authors' conscience. I came up with it especially for the purpose of connecting indicators to neural networks. In this case they are extremely easy to connect. But I think the idea has its own value. I won't have to deal with a new indicator written using this standard, its curve is immediately understandable. Otherwise, it may often happen that you see an indicator on the Internet. There is no description of it. And even if there is a source code, it is difficult to say what the author had in mind and what to do with it... Alas, such popular things as various moobs and Bollinger lose their right to exist with such an approach. But no one promised it would be easy... The advantages of such a standard, it seems to me, many times outweigh its disadvantages.
 
eugenk1 писал (а):

I came up with this precisely for the purpose of connecting indicators to neural networks.

Do you people know about the Nostradamus system? Or should I dig out the link from the archives again? For 100 indicators, prices from different timeframes, all in neural network....
 
gpwr писал (а):

Except that there are very few experts who are capable of self-learning (adapting), i.e. optimising x1, x2 ... in real life.


Who's not too lazy;-) Optimise the optimisation period in my Expert Advisor from the Championship. My Expert Advisor itself occasionally shows profit on M15, H1. I have not got time to experiment with it yet.
 
Integer wrote:
Who's not lazy;-) Optimise the optimisation period in my Expert Advisor from the Championship. My Expert Advisor itself occasionally shows profit on M15, H1. I have not yet got time to experiment with it.

If it is not a secret - how much time passed between the official announcement of the Championship and the end of registration?
12345678910...28
New comment