|
PureBytes Links
Trading Reference Links
|
> -----Message d'origine-----
> De : Bilo Selhi [mailto:citadel@xxxxxxxxxxxx]
> Envoyé : samedi 9 juin 2001 05:46
> À : David Rosenthal; Omega List
> Objet : Re: Fuzzy Logic
>
>
> the right approach would be:
> - to know the right variables to use, not letting the machine
> select those for you but knowing which ones to use from
> your knowledge domain ( market rule base )
This is not forbidden by the automated software approoach...
> - to know the right rule base to use for those variable to
> generate buy and sell signals, not letting the machine find
> the rule but to define those rules yourself...
This is what we did with fuzzy logic 5 years ago, and IMHO you are missing
some points:
If you define the fuzzy sets yourself, build the rule verify what rule has
to be fired or not, and so on, you are going to spend days if noy weeks to
set up a simple FL system.
Also, you must consider that the number of fuzzy rules is given by the
product of the number of fuzzy sets for each input.
If you use 4 inputs and 5 fuzzy sets for each, the theoretical number of
rules is:
4^5 = 1024 fuzzy rules...
Not to speak of the problem of drawing the fuzzy sets ( they do not have to
be equaly spaced).
Not to speak of the target pergformance summary that you want to maximize
or minimize), and that will depend on the above.
Using FL in such a manner is time costly, and very disappointing when you
see that something has to be modified after testing.
What will be modified by hand ? Fuzzy sets location ? which one ? All ?
number of fuzzy sets? for what input... and so on...Which rule has to be
modified amon 1024, and why...
This is a multidimentional problem that cannot be properly achieved by an
human being ( at least those who I know).
However, the software achieve this goal within a few minutes.
>
> - know the key variables*
With some experience, you will see that all the variables are key variables
if they are not redundant.
> - know the key rules
This is understandable when using crips logic. Not with FL, because the
fuzzy rules depend from all the variabes to be used
> the results should then match the specs...
>
> instead of
> - let the machine curve fit the variables
We do not do that...
> - let the machine curve fit the rules
> so that the results match specs... ( ultimate curvfit )
Believe me or not, this is the solution, and this is what you attempt to do
with a FL toolbox, without the benefits of automation.
>
> you are the knowledge domain engineer and YOU
> define the vars and the rules...
Yiu have to define the vars, that's right, but defining the rules will not
be a piece of cake (see above)
>
> trade size selection, risk, reward can be considered
> fuzzy vars also, so no big new revelation there...
>
Yes.
This will be added to the next Safir-X Pro version, as well as stop or limit
orders, exits and so on.
The problem will be even more complex to solve...
That aside, I agree with you with the benefit of using FL instead of crisp
logic, and only want to point out that the difficulties highlighted here are
for real, that we have had them and finally gave up with the half hand
coding solution, and are very happy with the automated search, even if we do
not know what and why the software does what it does.
Only the result is of importance.
Going to your solution is like willing to build a NN and determine the
weights by understanding them.
None is doing this as well as none is currently using an expert system like
those advocated in the 80's ( a FL rule base system is close to an expert
system).
Sincerely,
Pierre Orphelin
www.sirtrade.com
TradeStation Technologies representative in France
>
|