Algorithm Optimisation Championship. - page 9

New comment
 
Реter Konow:

1. So - the FF value area is not just a range with two boundaries, between which there is only emptiness and lonely peaks of peaks. It's a full surface with relief that needs to be probed all the way through?

2. Does the FF transfer the "surface relief curves" to the algorithm?

3. So, the algorithm has to access the FF a huge number of times to get a minimal "idea" of the "surface" topography.

4. Until now, I imagined it in a two-dimensional array space, which simply stores some values to be found in a limited number of tries, but judging from the pictures, the search space is actually three-dimensional...

In other words, the number of values to be searched is several orders of magnitude higher. So, the more times you access the FF (look at the surface) to make a "relief map", the more accurately the vertices of the surface will be found. But the number of references should be reduced by competition rules... Something I understand... :)

5. So, if we address the surface a maximum number of times, we may create a perfect copy of relief.

6. But then, the fewer times, the worse will be the result?

1. Inside the FF can be anything, even a translated War and Peace, or a human genetic sequence. Anything.

2. If you mean the FF formula, no. Only the result. Parameters - > Result.

3. In the examples above, the functions are of the form F(x1, x2). That is three dimensional search space - 2 parameters. But earlier I said target 100-500 parameters, which means the search space will have much more dimension than 3.

4. The search space is multidimensional, much more than 3. Parameter variants are innumerable (limited only by properties of double value). We need to apply search strategies so that we don't have to do a complete search, that's the point.

5. It is not necessary at all. Of course, if we mean not a single blind poke.

 
Andrey Dik:

1. Anything can be inside the FF, even a translated War and Peace number, or a human genetic sequence. Anything.

2. If you mean the formula of the FF, no. Only the result. Parameters - > Result.

3. In the examples above, the functions are of the form F(x1, x2). That is three dimensional search space - 2 parameters. But earlier I said target 100-500 parameters, which means the search space will have much more dimension than 3.

4. The search space is multidimensional, much more than 3. Parameter variants are innumerable (limited only by the constraints of double value). You need to apply search strategies so that you don't have to do a complete search, that's the point.

5. It is not necessary at all. Of course, if we mean not a single blind search.

Since according to championship conditions we are searching for maximal values, their analogy to vertices seems better to me. Besides, the picture depicts it that way. If there are maximal values of FF, there are also minimal ones. top and bottom. space. physics still knows 4 dimensions. The others are only in theories. In space, there is emptiness and matter.

If you say that in our case, the search space can be completely filled with some disordered content, inside which you have to find points with maximal values, the representation of surface and relief disappears, and a representation of chaos of points with different values appears.

Applying a search strategy under such conditions doesn't seem possible to me. Another thing is if it is a familiar three-dimensional surface with relief and vertices on it...

I understand that the search strategy is the key to effective results, but in my opinion, for the algorithm to work, you need the surface, not the chaos of point values (like DNA or digital War and Peace) .....

 
Life tasks do not owe anyone anything and therefore can be anything, including not having to be bound to the physical world.
We define by numerical abstractions. And even if the search space is filled with random noise, such a space also has a global maximum. The degree of accuracy in determining such a maximum in a limited number of trials is the most important indicator of the quality of an algorithm in terms of search capabilities.
 
Andrey Dik:
Life's tasks don't owe anyone anything and therefore can be whatever they want, including not having to be tied to the physical world.
We define by numerical abstractions. And even if the search space is filled with random noise, such a space also has a global maximum. The degree of accuracy of determining such a maximum in a limited number of trials is the most important indicator of the quality of the algorithm in terms of search capabilities.

I agree with you, life sets us tasks without regard to our capabilities.

Do you think it is possible to apply a search strategy in chaos (random noise)?

I'll give it a try. :)

 
Реter Konow:

Do you think you can apply a search strategy in chaos (random noise)?

You can, who forbids it? ))

However, if we do not know beforehand what the search space is, it is better to use some search strategy than to search at random. The number of solved problems will be greater and the solution will be of higher quality, regardless of the nature of the problems.

 
Andrey Dik:
You may, who forbids it? ))

It's not going to be very easy.

So, there is some numerical chaos in the limited FF array space, which, when accessed, needs to be read.

Using a preconceived search strategy, it's necessary to reduce the number of calls to the realm of values, but still calculate the closest values to the maxima hidden in the depths of the array...

There's no ordering in FF's array...

However, it is not easy. :)

 
Реter Konow:

It's not going to be easy.

So, there is some numerical chaos in the limited FF array space, which, when accessed, needs to be read.

Using a preconceived search strategy, it's necessary to reduce the number of calls to the realm of values, but still calculate the closest values to the maxima hidden in the depths of the array...

There's no ordering in the FF array...

However, it is not easy. :)

It's not there, it could be there. Anything can be there in FF. Think of poor old Schrodinger's Cat. You can't tell in advance what's in the box unless you try to find out.
 
Andrey Dik:
It's not there, it could be there. Anything can be there in FF. Think of poor old Schrodinger's Cat. You can't tell in advance what's in the box unless you try to find out.
Let's try... )))
 
Andrey Dik:
Do you want to participate?
I was just browsing and came across 2 Igor Volodin, but then I saw that he was paying attention himself. That's why I deleted my empty post. And in terms of participation I don't possess that level in programming. Sorry for the inconvenience! Good luck to everyone in this interesting competition!
 
Реter Konow:

It's not going to be easy.

So, there is some numerical chaos in the limited FF array space, which, when accessed, needs to be read.

Using a preconceived search strategy, it's necessary to reduce the number of calls to the realm of values, but still calculate the closest values to the maxima hidden in the depths of the array...

There's no ordering in FF's array...

However, it is not easy. :)

Not all functions have noise. But some do, so the gradient descent method fails.

It's not for nothing that the name "genetic" appeared, analogies from nature work well: crossbreeding, mutation.

At first, I also wanted to at least partially use the gradient descent method, but I gave it up completely.

12345678910111213141516...133
New comment