Using artificial intelligence at MTS - page 23

New comment
 
Rosh:
You need to write a script or set different arrow colours for different MagicNumbers in the code itself.

That's a great idea, thanks to the "code to set different arrow colours for different MagicNumbers"!

But about the script:
Do you mean changing the colour of the arrows, or will the script be able to remove the "extra" orders?
 
Aleksey24:

But about the script:
do you mean changing the colour of the arrows, or will the script be able to remove "extra" orders?

It is as your heart desires. :) An example of how to process the arrows is given here http://www.alpari-idc.ru/ru/articles_mql4/25.html
You will need to check the descriptor - the tester writes MagicNumber in it.
 
Rosh:
Aleksey24:

What about the script:
do you mean changing the colour of the arrows, or will the script be able to remove "extra" orders?

It is up to your heart's will. :) An example of how to process the arrows is given here http://www.alpari-idc.ru/ru/articles_mql4/25.html
You will need to check the descriptor - the tester writes MagicNumber in it.

OK.
And in general "spectral diversification" is not such a simple thing.
I took only 2 parameters (two indicators) at a minimum to start with.
With 2 good values of local extrema of each.
I use arrays instead of parameter values.
A total of 4 MagicNumber (4 times in loop start())
Lot size is divided by 4 - i.e. 4 Expert Advisors at a time (4 times as many trades).
I will share my results.
 
Reshetov:
Integer:

Equivalent to smoothing AC using a cypher filter with some characteristics. The smoothing coefficients are not balanced, which is equivalent to a brick on the buy button. The brick (+ e.g. stochastic) works very well by itself, if only one knows when to buy it and when to sell. Also, taking into account that the AC can go down 2 times in 21 bars and the presence of 4 optimisable parameters......))))

But for me it sheds light on how neural networks work and why they are not as efficient as one would like.

I used to have a hobby in the beginning of creative period - to write EAs for working on m1 based on last week results (7200 bars, as opposed to 66000) - as much as 300 percents per week were shown in the tester.....

I wonder how many harmonics the price needs to be decomposed into a Fourier series to get a grail after optimisation?

Neural networks are effective only when identifiable objects are linearly separable, i.e. when one class of objects can be separated from another in feature space by a plane described by a linear equation.

As for the AC oscillator, the Expert Advisor not only looks at its last value (decisions based on the last values are most often used in technical analysis), but it studies the history, i.e. what were 3 other values of the indicator in the past. He is interested in the oscillator behaviour for decision making. This very behaviour gets to the input of the neural network. And on the output we get Buy or Sell.

Another new feature is not the standard neural network training, but selection of weights on historical data using the genetic algorithm. I tried both variants. Genetics gives slightly worse result and slower in time. But there is no built-in neuronics algorithm and learning it in MT4. But there is an optimization based on genetics. And some researchers in this area realized that dynamic learning is not very adequate if the situation changes drastically. If bulls prevail in the market, the system will retrain to the bull trend and forget about the bearish trend. And vice versa. Samuel A. L. 1959, "Some studies in machine learning using the game of checkers" (IBM J. Research and Devepopmend 3: 210 - 229), first encountered and described this monstrosity. He observed that if his programme had a professional opponent, it gradually moved to a professional level game. But if the opponent was a beginner, then the program "forgot" the previous level and started to pass to primitive play. Therefore it probably does not make sense to dynamically teach the neuron on its own mistakes and losses. It is easier to pass it through history, in order to develop a trading strategy adequate to the market.

As for the grails, you do not have to be very smart. You just need to fulfill a number of conditions:

1. The system must open positions either without any stoplosses, or with stoplosses at a very large distance, so that the probability of their operation is close to 0.
2. A powerful filter based on several indicators with triggering conditions separated by a logical AND (&&). And to pull a lot of input parameters of these very indicators into the MTS external settings, so that only a few positions were opened during several years of historical data on tests.
3. To all this add capital and risk management with a raised fraction


I am not an expert on neural networks but as far as I remember, what was said about linear separability refers to the simplest of the first networks on perseptrons. It was proven that they do not possess this property, as for neural networks in principle they were created to solve such problems as non-linear separability. Correct me if I'm wrong, there's a lot I just can't remember.
 
Do you have any interesting links on the subject of NS application in trading. There is already a large library of NS theory.
I plan to use NS not for prediction, extrapolation or interpolation, but only for searching for patterns.
I am specifically interested in technology of training, for example.
For example, with a trainer - suppose we give sets of signals for each bar on history as a training sequence, and the expected TC output is -1/0/1 (sell/0/buy).
And what, for each bar we should manually preset signals? How to avoid?
And how to apply training without a teacher in this case? How to obtain the result of trades on the entire history and achieve maximal profit?
What methodology is used for this purpose?
 
Dali:
Does anyone have any interesting links on the subject of NS application in trading. There is already a big library of NS theory.
I plan to use NS not for prediction, extrapolation or interpolation, but only for searching for patterns.
I am specifically interested in technology of training, for example.
For example, with a trainer - suppose we give sets of signals for each bar on history as a training sequence, and the expected TC output is -1/0/1 (sell/0/buy).
And what, for each bar we should manually preset signals? How to avoid?
And how to apply training without a teacher in this case? How to obtain the trade result on the whole history and achieve the maximal profit?
What methods are used for this purpose?

For pattern recognition try to use the kohonen map. But first you should either normalize the data or just code it.
 
Dali:
...
For example, with a teacher - suppose we give sets of signals for each bar on the history as a training sequence, and the expected output of NS: -1/0/1 (sell/0/buy).
And what, for each bar, we should manually preset signals? How to avoid?
...
I've also thought about it for a long time - the easiest way for me was setting signals on next bar's close: higher - buy, lower - sell, not very good quality, but "locality" of lows and highs is determined only by the timeframe used. I used H4, the result is OK, though the EA does not work by ticks, but "with explicit control of the bar opening".
 

Yes Kohonen cards are good, the most important thing here, imho, is to code the patterns correctly before feeding them into the network inputs.

 
A question has arisen. Does anyone have a criterion for determining whether a cochonen net is trained or not?
 
Vinin:
A question has arisen. Doesn't anyone have a criterion - how to determine if the cohonen network is trained or not.

If for N iterations examples are no longer divided into classes and migration of patterns has stopped, then we can say that learning is complete.
1...16171819202122232425262728
New comment