|
PureBytes Links
Trading Reference Links
|
----- Message d'origine -----
De : Bilo Selhi <citadel@xxxxxxxxxxxx>
À : <omega-list@xxxxxxxxxx>
Envoyé : vendredi 18 août 2000 23:39
Objet : Re: Re: Fuzzy Logic vs. System Dynamics
> i am attaching a zip file that contains a txt
> file with in depth views about nns, fls, etc
> and trading system design in general...
> in layman's terms ( sorry it's over 15k limit )
>
> bilo.
> ps. i checked out sirtrade and safir...
>
>
I have also checked your text ( all , yes, I can read very fast, even in english)
Overall interesting,but I disagree with this assumption:
<<HYBRIDS:
there are plenty of those that combine FL and NN in one package.
the key word is neurofuzzylogic. or NN, FL and GA and you get
a neurofuzzylogic optimizer... most of those tools are designed
to curvefit the crap out data by any means, be it tweaking the inputs
to get the desired outputs or tweaking the insides of the nn or the rules
or all of the above to get the desired outputs to create an illusion
of something that's working...
basically they are a fad designed for sale.
they are poor as modeling tools since they are not designed for that purpose.
the more non-linear the tool becomes the easier it gets to curver fit the data...>>
First , there are not so much tools available in the neurofuzzy logic (NFL) field.
Second, f it does not generalize, this means that the data set was not correct or the problem impossible to solve.
Unlike what you think, using NN in the background of FL greatly helps ( we have performed tests here on both when building the NFL engine).
The result is that plain FL is unable to properly learn anything that has no solution.
The NN alone can ( but will of course never generalize).
The NFL structure is in between but closer of the FL than the NN in terms of results.
It only provides an advantage over complex non linear problems, and keeps the barrier against overfitting that FL is.
Because of the NN structure that acts at every step of th FL process ( I cant say more) it's FL assisted by NN, not a FL engine and a NN added ( It's not a FL inputs coding to load a NN in the background).
If you are not yet convinced, try the following:
Build a training set with 4 indicators that do not work with NFL a la Safir-X ( this is of course possible).
Take the same file and train with a plain FL system: It will not work too.
Train it with a NN: They are chances that it will work.
In all of the 3 cases above anything will fail on unseen data.
Then draw then your own conclusions.
That's aside, the toolbox is maybe a good idea... for people who like research.
But is it the case of anyone here who wants to take advantage from this technology ?
Rgds
PO
|