MT5 terminal updated today and the "Optimisation" window does not show up during the test - page 7

New comment
 
Sergey Chalyshev:

Your words scare me, (especially the highlighted one).

It turns out that the genetic algorithm doesn't work and you advise to use a ramdomizer.

The purpose of genetic algorithm is not to use randomness, but to find an effective solution in less time.

p.s. and also? the new build has flipped all the enum variables, not good, now all the sets need to be changed, it's a mess.

I advise you to read my words carefully once again and then reread the above mentioned articles.

How long can you turn off your brain and press Start without understanding mechanisms of how the genetic algorithm works? It works absolutely correctly.

The randomizer is an integral part of the gene mixing mechanism. In addition to a targeted movement through crossing towards the target function, a mandatory mutation through randomisation is required.

 
Renat Fatkhullin:

I advise you to read my words carefully once again and then reread the said articles.

How long can you turn off your brain and press the Start button without understanding the mechanisms of how the genetic algorithm works? It works absolutely correctly.

A randomizer is an integral part of the gene mixing mechanism. In addition to purposeful movement through inbreeding towards the target function, mutation through randomization is necessary.

People are afraid of not randomness in algorithm (randomness is the essence of heuristic algorithms, and hopefully everyone understands this, be it GA, annealing, Monte Carlo, etc.), but that you recommend systematic restarting of search. But why? - Because there are mutations in algorithm that provide random search space, why more "to make random", you might as well use pure random with sorting by FF without using operators like crossing, because the result is still random and do not hope for satisfactory search results, which is actually what Sergey said:

Sergey Chalyshev:

...

It turns out that genetic algorithm doesn't work and you advise to use a ramdomizer.

...
 
Andrey Dik:

What scares people is not the use of randomness in the algorithm (randomness is the essence of heuristic algorithms and hopefully everyone understands that, be it GA, Burnout, Montecarlo, etc.), but that you recommend doing a systematic restart of the search. But why? - Because there are mutations in algorithm that provide random search space, why more "to make random", you might as well use pure random with sorting by FF without using operators like crossing, because the result is random anyway and don't hope for satisfactory search results, which is actually what Sergey said:

Enough of these old songs. It has been discussed several times.

And I recommend absolutely reasonable. With a large search area there are always many unidentified extrema which can be found by repeated attempts. You know this very well.

And here you've decided to cast deliberate, "just as well as random". You're just trying to make a big deal out of it, just like you did the last time you discussed genetics.

Sergey said on the basis of misunderstanding the mechanism of genetics. You know everything, but have decided to deliberately throw wood on the fire of questions of unprofessional participants. Not only that, you pretend that mutation is not random. Random in genetics is mutation. I have clearly written about it,



So as not to frighten people, I specifically gave a list of easy to understand articles on genetics. Your article is in there. Where you yourself explicitly write about randomisation:

Protopopulation creation. Genes are randomly generated within a given range.

In the NaturalMutation operator the mutation is to generate a random gene in the interval [RangeMinimum,RangeMaximum].


Here is a list of articles once again:

 
Renat Fatkhullin:

Enough of the old songs. It's been discussed several times.

And I recommend it absolutely sensibly. With a large search area, there are always a lot of unrecognised extremes that can be found by repeated attempts. You know this very well.

And now you've decided to throw in a deliberate, "might as well be random". You're just trying to make a big deal out of it like you did the last time we discussed genetics.

Sergey said on the basis of not understanding the mechanism of genetics. You know everything, but decided to deliberately throw wood on the fire of questions of unprofessional participants. Not only that, you pretend that mutation is not random. Random in genetics is mutation. I have clearly written about it,

...
So as not to scare people, I specifically gave a list of easy to understand articles about genetics. There's your article in there too. Where you yourself explicitly write about randomization:

That's right, mutation is random, that's what I'm saying.

I'm not trying to put anything in my head but I'm trying to clarify it. Most people would never understand why you need to consciously manually restart, you could scratch their heads and they wouldn't understand...

So I say - why make people do manual fiddling (independent pressing of buttons, which also requires brains), when you can use algorithm for random probing in unfamiliar areas in addition to all other operators?

so let me explain my thought more briefly - a man can not always trust deterministic processes, and one can not trust random processes either! - If you build probing in random directions into the algorithm, you will never get questions like "Why should the optimizer be restarted?", "How often should the optimizer be restarted?", "What can make you feel an acute desire to restart the optimizer?" - will never occur to users.

I will even say more, as there are no continuous processes in the market (quotes are discrete), so there is no question about the smoothness of the studied surface of the TC results, which means that even if we use pure random and sorting, we will get optimization results not much worse than GA (not critically worse), which means that by flipping a coin the probability does not change from previous results, which means that there is no point in manual restarts of the optimizer, results will still be the same as if we had not stopped optimizing

 
Andrey Dik:

So that's what I'm saying - why make people do handwashing (pushing buttons by themselves, which also requires a handwashing mind), when you can put in the algorithm besides all the other operators just random probing in unfamiliar areas?

Stop talking nonsense and suggesting that the company kills itself against the wall.

You may as well do random, you may as well scare people, and now you say "why bother with handicrafts". Again, you are clearly aware that this is exactly how genetics works. And just as clearly you know that you can't reliably genetically probe the entire space in any one or 2,3,4, passages. There is always a chance of a miss. So I'm absolutely right with the advice to check with repeated passes if you want more certainty.

In comment #22 I easily showed how genetics found better results on the fourth pass:



The strategy tester is an analytical tool. And the genetic algorithm is a means of drastically reducing the amount of computation. This tool has conditions of applicability and methodology of use.

If a trader has reached the point of using robots, he or she will have to learn both the subject area and the tools in any case.


Save my time, please. You are going to teach me with such primitive statements about a subject in which I have invested so many years.
 
Renat Fatkhullin:

Stop talking nonsense and suggesting that the company should kill itself against the wall.

Then it's "just as well random", then "you scare people", and now "why bother doing handicrafts". Again, this is a blatant "put probing into the algorithm", when you clearly know that this is exactly how genetics works. And just as clearly you know that you can't reliably genetically probe the entire space in any one or 2,3,4, passages. There is always a chance of a miss.


A strategy tester is an analytical tool. And genetic algorithm is a tool to drastically reduce the amount of calculations. This tool has the conditions of applicability and the methodology of use.

If a trader reaches the use of robots, he or she will have to learn both the subject area and the tools in any case.

I'm sorry, but there is some misunderstanding between us.

We do two simple experiments:

1. Run the optimizer 10 times for 100 passes.

2. run the optimizer for 1000 passes once.

It is clear that such experiments will not work (or will they work?), but the message is clear.

Unfortunately, I am once again convinced that it is dangerous to bring experience to people - they will stonewall them (why? - just in case)

 
Andrey Dik:

I'm sorry, but there is some misunderstanding between us.

We do two simple experiments:

1. run the optimizer 10 times for 100 passes.

2. We run the optimizer for 1000 passes once.

It is clear that such experiments will not work (or will they work?), but the message is clear.

Once again we make a test with deliberately incorrect conditions.

Trying to go into the realm of completely random starting protopopulations (100 passes is cheeky, 1,000 is cheeky) to completely disable genetics. Genetics will only give more or less reliable results for 25-30 generations, and that is from 10 000 passages.

The message is clearly understood. Cut the nonsense.

 
Renat Fatkhullin:

In comment #22 I easily showed how on the fourth pass the genetics found better results:

the ugly truth of life is that the best results could appear on the last run, or even on the first run (i.e. all subsequent runs can always produce worse results, even if all possible passes are made). there is no pattern to this. so it is equivalent to simply making all passes through history at once (within GA - optimization of course, rather than billions of passes).

i.e. we never know for sure if these are the best results, maybe the next runs will be even better. so why break your head with such questions, to which there is no and can be no answer? - give all the passes at once and that's it (10,000 GA passes).

that's it. I'm gone, sorry.

Those who have tried to mine coins using optimization algorithms clearly understand what I'm talking about, and you clearly understand it even if you haven't tried mining. every optimization task has a certain search complexity ceiling, once reached, the results will not be any different from a random search. In mining, this ceiling is very low (I managed to find up to 7 zeros in hash, then the ceiling), in trader tasks, this ceiling is much higher, can reach 90 or more percent of total number of passes (tasks are millions times easier than mining), but the essence of this does not change.
 
Renat Fatkhullin:


We'll fix the drawing of the optimization graph on Monday.

Regarding the optimization graph, take another look at the line chart. The information is not fully reflected.

Old build 1755

1

Build 1810

2

Test conditions as in post #34.

p.s. The rest of the indicators are back to normal.

 
Sergey Chalyshev:

Directly, all enum upside down, if it was 1,2,3 - now 3,2,1.

Better don't update if possible, wait for stable version to be released.

Yes, I'll clean by hand every time, not enough refresh button as automatic refreshing gets in the way

1234567891011121314...25
New comment