Discussion of article "Population optimization algorithms: Saplings Sowing and Growing up (SSG)" - page 10

New comment
 
Sort of cross validation selects the best needle or surface. And to get many needles, you can optimise across different pieces of history. The same ph-id will remain the same.
 
Andrey Dik #:

I can't be responsible for other articles on MQL, but there people did something and shared something, unlike you.
wikipedia is just a simple writing of people who are not responsible, and also politicised writing.
As for my articles, what exactly doesn't work? Stop blabbering, do something.
Yeah, I ruled one wikipedia article myself when I did a promotion 😀 You send people to wikipedia to say look, it says the same thing. It's really funny.
 
Andrei, are there many more algorithms left? Does it make sense to stop at SSG, or are there potentially stronger ones? )
 
Maxim Dmitrievsky #:
Sort of cross validation selects the best needle or surface. And to get many needles, you can optimise across different pieces of history. The ph will remain the same.

here, by the way, is a way to filter out parameters that are always poured (enter them into owls as erroneous and the tester will skip them). select the areas that are most often poured. and then use these poured areas as your imagination prompts you.

 
Andrey Dik #:

Here, by the way, is a way to filter out parameters that are always draining (enter them into owls as erroneous and the tester will skip them). select the areas that are most often poured. and then use these poured areas as your imagination prompts you.

It's not quite clear why the tester doesn't have such an intuitive way to run on several pieces of history and average. Maybe it is done through frames somehow.
 
Maxim Dmitrievsky #:
Andrei, are there many more algorithms left? Does it make sense to stop at SSG, or are there potentially stronger ones? )

There are a lot of algorithms, I don't know if there are more powerful algorithms.

The table is alive, I add algorithms to them as I learn them, i.e. I can't say - that one over there is the coolest, I only know the ones I described))))

In fact, it was already possible to take ant, bee and weed, they are very good. wooden of course now tears all, what will be the next leader - I do not know.

I'll get to the hybrid ones when I've gone through all the important known ones, hybrid ones are very promising.

For now I am considering population types, but there are other types, it will be interesting to study them too.

 
Maxim Dmitrievsky #:
I don't quite understand why there is no such intuitive way to run on several pieces of history and average them in the tester so far. Maybe it's done through frames somehow.
with tester/optimiser in general a lot of new ideas come, I don't know if it is a dedicated developer who does it, or if any of the team who is available does it....
 
Nikolai Semko #:
Taken out of context. Read on.
I was saying that the correctness of the choice of a point on the OP does not depend on a hill or a depression and not even on the velocity of the local vector of motion in time, but only on the sign of the acceleration vector (derivative of velocity), one part of which is in the future, which is not known.
So, what I'm saying is that by observing the time variation of the OD,
we can make a prediction of the next OD, which means we can get the sign of the acceleration vector and the point and the fit....
So it's the future that's unknown.


OP is an optimisation surface
 

It is a big mistake to get ideas from looking at three-dimensional pictures. It is like drawing conclusions about the three-dimensional case from two-dimensional pictures.

With two parameters, the number of saddles roughly corresponds to the number of maxima - between two maxima there is one saddle (with one parameter there are no saddles at all). As the number of parameters grows, the number of saddles becomes much larger than the number of extrema and they become more diverse. And the main task of maximisation becomes not to take a saddle as an extremum, which is quite possible because of the limited number of calculation points.

If there are discontinuities in the dependence of the target on the parameters, then there is complete darkness and it is simply impossible to imagine all multivariate variants.

 
Aleksey Nikolayev #:

It is a big mistake to get ideas from looking at three-dimensional pictures. It is like drawing conclusions about a three-dimensional case from two-dimensional pictures.

With two parameters, the number of saddles roughly corresponds to the number of maxima - between two maxima there is one saddle (with one parameter there are no saddles at all). As the number of parameters grows, the number of saddles becomes much larger than the number of extrema and they become more diverse. And the main task of maximisation becomes not to take a saddle as an extremum, which is quite possible due to the limited number of calculation points.

If there are discontinuities in the dependence of the target on the parameters, then there is complete darkness and it is simply impossible to imagine all multivariate variants.

Yes, quite right. Three-dimensional pictures are the maximum that we can see, more dimensions cannot be seen. But we need to have an idea of the surface for AO tests.

I use three-dimensional test functions (two parameters), even where there are 1000 parameters in the tests it is 500 test functions.

If the FF is "heterogeneous" in parameters, as it is in the case of the Expert Advisor, then it is impossible to imagine the gyre surface at all, but it is not more difficult than "homogeneous" test functions. All the algorithms in the articles are tested for "chitting", as for example, you could actually option two parameters and copy them to all other parameters, then the test multivariate functions would click on one and two times.

There is also a method on "parallel-perpendicular" (I don't know how it is called exactly) tendencies of algorithms, it is when an algorithm solves better optimisation problems where vertices and troughs are located vertically and horizontally to coordinate axes, such algorithms fail tests on functions with rotation (take any test function and rotate it by 5-10 degrees).

1...345678910111213
New comment