Discussion of article "Population optimization algorithms: Saplings Sowing and Growing up (SSG)" - page 2
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
ZY. An interesting question for all those who are interested in this topic: what is the difference between local extrema and global extrema (not taking into account their difference in FF values)?
Nothing.
A few needle antennas.
I don't really understand what you want to do so I don't guarantee the quality of the advice...
that's what you need:
that's what it takes:
Not really. Suppose a GA took 100 steps on a function like the one in the picture. Of those, 90 of them will end up near the global. That's the cluster of close ones that is worth taking.
If we are dealing with a hedgehog, we will get a lot of mini-clusters around some points. Those points are what we need. The GA could refine the coordinates of the clusters through the narrow space around them.
Roughly speaking, we need to classify the GA results into clusters and then finish each cluster with narrow optimisation. We will get a set of input parameters "interesting" for TC.
Not really. Suppose the GA took 100 steps on such a function as in the picture. Of those, 90 of them will end up close to the global. This is the cluster of close ones that is worth taking.
If we are dealing with a hedgehog, we will get a lot of mini-clusters around some points. Those points are what we need. The GA could refine the coordinates of the clusters through the narrow space around them.
Roughly speaking, we need to classify the GA results into clusters and then finish each cluster with narrow optimisation. We will get a set of input parameters "interesting" for TC.
right?
is that it?
Yes. I think that if after each optimisation you cut out a chunk of the space (like 80% of the input, what's around) of the found global, that's how everything is found.
Figure 5: Forest test function.
An excellent visualisation of what can be seen during a full TC enumeration. Of course, 3D is the two input parameters here. But the slides/spikes are clearly visible. For TCs, spikes are generally evil. Hilltops, on the other hand, are the most interesting.
Regarding spikes being evil. For TC, they are randomness - a tight fit (regardless of the optimisation criterion).
Yes. I think that if after each optimisation you cut out a chunk of the space (like 80% of the input, what's around) of the found global, that's how everything is found.
Such a chunk is characterised by a given interval for each input parameter. So if you have the data of the cut region, you can very easily (even in the standard Tester) perform optimisation without this piece of space.
But I completely lack the competence how to define the area around the found global maximum in the GA results.
A simple example. We ran an optimisation on some TC. It has finished by outputting sets of inputs. We need to find the most noticeable (the number of points divided by the minimum radius of the sphere they fit into) cluster of multidimensional points among these sets.
What is needed is a mode that finds all hills and gives these ranges for all parameters that can be worked with further.
Subsequent optimisations can only be done within the ranges of such robustness hills.
it's the hilltops that are the most interesting.