Discussion of article "Practical Use of Kohonen Neural Networks in Algorithmic Trading. Part II. Optimizing and forecasting" - page 2

 
Maxim Dmitrievsky:

GA will never converge to a global optimum, it's an evolutionary method aimed at diversity. It does not have to converge to anything at all.

https://www.monographies.ru/en/book/view?id=707

It's not that it converges differently. It's that it's limited to 10K runs. It takes me less than 2 minutes to do an Optimisation. Why you can't automatically do a hundred such optimisations and show the total result is a mystery.

[Deleted]  
fxsaber:

It's not that it converges differently. It's that it's limited to 10K passes. It takes me less than 2 minutes to do an Optimisation. Why you can't automatically do a hundred such optimisations and show the total result is a mystery.

then you can put bruteforce and do a random search of parameters inside the bot, and save it in a table.

 
Maxim Dmitrievsky:

then you can put bruteforce and do a random search of parameters inside the bot, and save it to a table.

Randomisation and optimisation are different things.

[Deleted]  
fxsaber:

Randomisation and optimisation are different things.

naturally ) any selection criteria are added and that's it, distributions are narrowed gradually

 
Maxim Dmitrievsky:

naturally ) any selection criteria are added and that's it, the distributions are narrowed gradually

Let's not reduce to the obvious - implement any optimisation algorithm yourself.

[Deleted]  
fxsaber:

Let's not reduce it to the obvious - implement any optimisation algorithm yourself.

The question is not that, but whether naked genetics is needed at all. But if you come up with something interesting, good.

 
Maxim Dmitrievsky:

That's not the question, it's whether you need naked genetics at all. But if you come up with something interesting, good.

Well, I can't go through 30 million passes like that with full brute force. And to get a good result in 70 seconds is probably good. But there are no methods of improvement, except manual start.


While looking at variants on the topic of optimisation(this one, for example), I came to the conclusion that it probably won't be difficult to create auto-optimisation on real ticks. It became interesting, whether the results of TC improve?

 
fxsaber:

it is not difficult to create auto-optimisation on real ticks.

I encountered the problem of repeatability: it does not exist from pass to pass. What are some ways to get the same autopotimisation passes?

[Deleted]  
fxsaber:

Faced with the problem of repeatability: it does not exist from pass to pass. What are some ways to get the same autopotimisation passes?

MathSrand() if there is an HSC there.

 
Maxim Dmitrievsky:

MathSrand() if there is a DST.

Thanks, it helped.