ForumSections

Discussion of article "Population optimization algorithms: Saplings Sowing and Growing up (SSG)" - page 4

New comment
Andrey Dik #:

Forest is the FF. i.e., it is some requirements that were imposed on the system to be optimised. if the requirements to the system are changed, the FF will change, but the system has not changed, right?
It's as if the user tried to change the requirements in various ways and still got Forest. integral FFs look exactly like Forest, such as balance, for example.
It is necessary to try not to use integral FFs, if possible, and if it is not possible, to make NADstroika over FFs, i.e. to apply FFs to FFs in order to avoid sharp peaks.
well, let's take an example. we have taken a balance FF. it looks (presumably) like Forest. you can fish in the murky water of the results of optimisation by balance for particles of sense that lie somewhere near, and you can go the other way, you can superstructure the balance FF so that the surface is no longer acutely finite, and all the necessary parameters lie near and at the same time on the topmost gentle hill!
In short, we can say that if the FF is acutely finite, then either this is really all that can be squeezed out of the problem, or the researcher has made a mistake.

That's a whole other topic: when and how it makes sense to consider hills/spikes. To get to it, one must first learn how to find them on arbitrary functions.

fxsaber #:

This is a completely different topic: when and how it makes sense to consider hills/spikes. To get to it, we must first learn how to find them on arbitrary functions.


OK. Then we can try to reason like this......
though, what if it turns out that the "plateau" (maximum concentration of close points) is much lower than the global extremum? where is the level of acceptability of the plateau height?
I am trying to push to the fact that what is to be obtained can be described in some way, i.e. in iteg the problem leads to the search of the global extremum, where all the nearest acceptable solutions are concentrated).
fxsaber optimisation criterion).
I was interesting the same thing, here is my question to smart people here
Is this what you want to achieve?

Andrey Dik #:

okay. then we can try to reason along the following lines.....
although, what if it turns out that the "plateau" (maximum concentration of close points) is much lower than the global extremum? where is the level of acceptability of the plateau height?
I am trying to push to the fact that what is to be obtained can be described in some way, i.e. in iteg the problem leads to the search of the global extremum, where all the nearest acceptable solutions are concentrated).

Yes, it all boils down to finding the global. The last steps of a GA, like, have to be around that global. So they will basically be the area to throw out on the next GA runs.

That is, we always search only for the global using any of the proposed algorithms. Then we simply discard the area where the global has fallen and repeat.

mytarmailS #:
I was wondering the same thing, so here's my question for the smart people.
h ttps://stats.st ackexchange.com/questions/566930/optimization-taking-into-account-the-shape-of-the-optimization-surface
Is this what you want to achieve?

Yes, that. And even more than that.

In this picture, we will find both red and green regions if we act according to the principle of throwing out all regions of previously found maxima.

For the case of the picture we need five optimisations: four will find red and one will find green.

Then we pass the five points through the analogue of TesterDashboard and immediately see who is worth what.


ZЫ It is strange that "smart" people with suggestions of smoothing do not understand the nature of surface object formation.

fxsaber #:

Yes, that. And more.

In this picture we will find both red and green areas if we act according to the principle of throwing out all areas of previously found maxima.

For the case of the picture, we need five optimisations: four will find red and one will find green.

Then we pass the five points through the analogue of TesterDashboard and immediately see who is worth what.

Yeah,
So you want to find a few peaks and test them all, right?

I thought you wanted to find a smooth peak, ignoring the sharp peaks.


In the first case, you just need to run the OA a few times with a small number of iterations.


In the second case, it's the "optimising a noisy function" section.

A specialised domain with specialised AOs.

From the general ones, you'll do AO:
Simulated Burnout, Bayesian Optimisation.


AO - optimisation algorithm
fxsaber #:

Yes, that. And more.

In this picture we will find both red and green areas if we act according to the principle of throwing out all areas of previously found maxima.

For the case of the picture, we need five optimisations: four will find red and one will find green.

Then we pass the five points through the analogue of TesterDashboard and immediately see who is worth what.


Here! I am just trying to convey that it is possible to avoid five optimisations and get into the green region at once, within one optimisation. for this you need to introduce the main FF over the minor FF. the main FF should describe the green region and is the global maximum.
So, the FF on the picture is not what we need, we need a FF where we need to look for the global maximum.
It is difficult to explain it on fingers. but you can, in the article))))
mytarmailS #:
So you want to find a few peaks and test them all, right?

Yes. Checking more than 20 peaks, as a rule, does not make sense. If there is something robust, it should be found among these 20.

Of course, you can theoretically imagine a hedgehog with a shaved hill. Then there will be no hill among a hundred optimisations based on the emission principle. But this situation is far from practice.

As a rule, we form such a hedgehog ourselves so that it does not have a bald spot.

Andrey Dik #:

Here! I'm trying to tell you that you can avoid five optimisations at once.
He doesn't need the green area.
fxsaber #:

Yes. Checking more than 20 peaks, as a rule, makes no sense. If there is something robust, it should be found among these 20.

Of course, you can theoretically imagine a hedgehog with a shaved hill. Then there will be no hill among a hundred optimisations based on the emission principle. But this situation is far from practice.

As a rule, we ourselves form such a hedgehog to have a bald spot.

Then it's quite simple, as I wrote above.
You need 20 peaks, just run AO 20 times.
1234567891011...13
New comment