Hi,

I have question about the optimization with Zorro.

this is a first parameter:
...


after continue train - parameter 4


After testing p4 is a worse output like p1

I'm not understand, howto train Zorro parameters.
Each with Each
or One by One?

Thanks

Zorro does not select the best parameter. It selects the parameter value with the lowest gradient. So it can indeed happen that p4 causes a worse result than p1.

Hi,

i playing with the Workshop5_3.

After default train i have this:



but after the move Stop from the row 25 to row 18 ( now is a first optimizing parameter) i get the worse output.



I thinking, for the one from new version create matrix optimizing parameters and trying find the better output with mix the parameters.
[P1 - P2 - P3] ... first step
[P2 - P3 - P1] ... next step

.... try all variant.
-----------------------

Next test: change row 24:
var Threshold = optimize(1.1,0.5,1.5,0.1); // from 1 to 1.1
i get the better output:



Also idea for the new version. Set in percent how big can be change for the parameters between WFO cycle.


In this i see
between WFO 2-3 is is big change. If i set max change can be aprox. 10%?

Zorro is best for the develop AOS. Idea with WFO is excelent, but optimizing and testing with Zorro is need ( in me opinium) improve. For my is still like black box.
P.S: Sorry for my english. I'm learn only from manuals.

As I come from Mt4 I compare it with their method.

As far as I remember (that's why I don't have link) during the optimization (here training) in "Genetic-Mode" a table is created ordered acc. to the chosen value that has to be maximized.

At the beginning the parameter setup seems to be quite randomly but later most of variations from the top of that table are chosen. Once and a while a 'lucky punch' is tried from the middle of that table. In general even if the amount of possible setups is very, very high (~trillion) after 10'000 to 15'000 passes (seems to be the stop criterion) you get a quite useful result.

There is only one problem. The 'ini-part' can return a "INIT_PARAMETERS_INCORRECT" but this is taken by their Genetic algorithm as a valid pass (with result 0) in influences therefore the stop criterion which is of course influenced by the amount of passes done.
Without this little problem it seems to be a very simple and conducive solution.

This is quite helpful if you develop you system as you can start with a lot of parameters to optimize to find out which of them matters and which can be set as constants.

Calli