RE: [amibroker] SPSO vs Trib vs CMAE, was: random optimization?

[Fred Tonetti <ftonetti@xxxxxxxxxxxxx>]

Ton,

 

Yes, I understand the situation.  But I’m not sure you understood my answer.

 

I pointed you at a particular section of the UKB part of which states …

Given that unlike Exhaustive Search an Intelligent Optimization methodology will by its nature not examine every possible combination of parameter values, it would be unlikely without some additional influence or direction that the Intelligent Optimization process would have picked for parameter values those that were not particularly sensitive.  This is because the processes of judging or calculating parameter sensitivity is typically performed after the optimization is finished which when using Exhaustive Search is all that is required as we then have a chance to view the results of all combinations.

In order to ensure that parameter sensitivity is taken into account when looking for parameter values with high fitness utilizing an Intelligent Optimizer, it is necessary to have a methodology for evaluating how sensitive parameter values are and to have that in turn impact the fitness calculation during the optimization process so that the process is led to a more robust set of parameter values.

Given that the Intelligent Optimization process as implemented in IO sends parameter values to AmiBroker for evaluation by its optimizer and retrieves results back from AmiBroker to determine how it should alter its search pattern in the next generation, this is more straight forward then it would first appear.  This is accomplished then by between one generation of regular optimization and the next looking at the results coming back from AmiBroker and for those points that are worth further examination performing some tests for Sensitivity that are not dissimilar to the methodologies used to generate the bar charts above.  The IO options and mechanics for this while not difficult for the user to employ are varied and fairly sophisticated and as such rather than discuss all of them here I would recommend for those who are interested that you read the sections on Sensitivity in the full documentation.

If you haven’t already I would suggest you read the doc the above UKB points at that’s contained in the zip in the files section ( link below ) and in particular pages 36 – 41 and the related directives in the couple of preceding pages.

 

http://f1.grp.yahoofs.com/v1/EL6NSYzFbi5lzNRJnIL0sUgNocERaBTYHOEzhIJrkt7ziqXnfXe7mn5ZagTedsTtnFLl6mOa1qJAkJlmGOwM7RmyEbCRqvc/IO%20v3.0.2.zip

 

---

From: amibroker@xxxxxxxxxxxxxxx [mailto:amibroker@xxxxxxxxxxxxxxx] On Behalf Of Ton Sieverding
Sent: Saturday, February 07, 2009 5:08 AM
To: amibroker@xxxxxxxxxxxxxxx
Subject: Re: [amibroker] SPSO vs Trib vs CMAE, was: random optimization?

 

Thanks for yr answer Fred. Our problem is that we are no longer sure that your 'you dont' is correct. An answer that we also got from Randy Haupt ( Practical Genetic Algo's ) as well as Pardo ( Eval&Opt of Trading ) and others. They all agree with what you are saying. And this also was our initial opinion. But we got some very nasty results. Please look again to the picture I already have sent to Steve ...

 

 

On the left the result of NOO with CMAE. Only after having done the OO we got the desired area in the middle ( Prairies in stead of peaks ). The problem with NOO for us is that you're missing robust area's. We hoped to get this by using CMAE. What we got was a poor result. That at least is our experience of using the NOO's included in AB as well as your's. So we decided to do it in the old fashioned way ( OO ) and use a special algorithm that is looking for area's in stead of points ... And that complete job of course takes at least 5 times the NOO time ... Let's hope Intel is coming soon with the next generation of MultiCore's ...

 

Regards, Ton.

 

 

 

 

----- Original Message -----

From: ftonetti@xxxxxxxxxxnet

To: amibroker@xxxxxxxxxps.com

Sent: Friday, February 06, 2009 9:57 PM

Subject: Re: [amibroker] SPSO vs Trib vs CMAE, was: random optimization?

 

I'd be the first to agree that there is a signficant difference between 5,000 minutes and 500,000 minutes.

 

As far as needing exhaustive search is concerned ... you don't ... what you need is a statisitically significant number of randomly generated points in the +/- % range for the value arrived at in optimization ... This will give you a very good feel for how robust the parameter values are ...

 

What it won't do without help is force optimimzation to find areas in the solution space like this.  That requires testing like what is described above for every "fairly good" solution so that in the next generation it uses the results "adjusted" by the sensitivity or robustness to decide where to look next.

----- Original Message -----
From: Ton Sieverding
Date: Friday, February 6, 2009 7:46 am
Subject: Re: [amibroker] SPSO vs Trib vs CMAE, was: random optimization?
To: amibroker@xxxxxxxxxps.com

> Hi Fred. Please read Steve's text. He is talking about
> 'trillions' of pairs. I did not say that. I agree with you that
> 'most' IO's have a time in the range of several minutes. Your
> average of 5 minutes is fine for me. As long as I am using
> single combinations and doing the job with IO, I have no
> problem. My problem is when I want to optimize many AFLs with
> many Symbols. I gave 100/10 just as an example. Take that for 10
> users with different OF's. Don't forget that IO does not give me
> the complete landscape. Perhaps I should have given the
> environment we are really using : 100 AFLs with 1000 ETFs. The
> only way to handle this in an acceptable time is splitting the
> task over 6 PC's. And even that takes several days and nights ...
>
> Yes that's what we are doing. Trying to get an algorithm that
> optimizes on areas in stead of points. Searching for prairies in
> stead of peaks. See the picture I sent to Steve. But for our
> optimization job we need the complete landscape. So forget IO.
> It only gives me a small part of the total picture. We need
> Exhaustive Search. And thanks for the URL. That was the first
> thing I did. Make a study of Fred Tonetti's IO system ...
> Perhaps I should do that again -)
>
> Kind regards, Ton.
>
>
> ----- Original Message -----
> From: Fred T onetti
> To: amibroker@xxxxxxxxxps.com
> Sent: Thursday, February 05, 2009 11:19 PM
> Subject: RE: [amibroker] SPSO vs Trib vs CMAE, was: random
> optimization?
>
>
> See imbedded
>
>
>
> -----------------------------------------------------------------
> -------------
>
> From: amibroker@xxxxxxxxxps.com
> [mailto:amibroker@yahoogroups.com] On Behalf Of Ton Sieverding
> Sent: Thursday, February 05, 2009 11:59 AM
> To: amibroker@xxxxxxxxxps.com
> Subject: Re: [amibroker] SPSO vs Trib vs CMAE, was: random
> optimization?
>
> Sure, Steve. That's what I am doing also. Please check my
> underneath email and
>
> you will get the following sentence : "Let's say I want to
> optimize 100 different systems on 10 different Symbols or 1.000
> combinations times your trillions.
>
>
> It's a life time. I just do not have a solution for this. Do
> you ?" It's precisely this problem that kills me. I have no
> solution for that ... Especially not if you want to optimize say
> every quarter ...
>
>
> Most intelligent optimization runs take a few minutes for a
> single system on a single tradable. The number of combinations
> of optimizable variable values is typically irrelevant to run
> times for an intelligent optimization engine. As such
> optimizing 100 different systems on 10 different symbols at 5
> minutes each would be 5000 minutes or less than 100 hours . How
> do you arrive at "life time" run times ?
>
>
> And even that does not solve my 'area in stead of point'
> optimization problem. First I do not want to check the landscape
> visually and second that's only possible with 3D pictures. Above
> all, CMAE etc. does not give me the complete landscape. So even
> if I would like to check it, it's impossible. And the fact that
> I should check what CMAE is giving me shows the underneath
> picture. It's not true that CMAE is looking for area's. Perhaps
> it tries to do that but for me not really in a successful way.
> Therefore I really have a hard time with optimization ... In
> theory it works fine. Until you enter the 'Real World' ...
>
>
> The area you refer to can be checked mathematically by
> randomly selecting data points for each optimizable variable in
> some user selected +/- % range of the selected value. For a
> thorough discussion of this see
> http://www.amibroker.org/userkb/2007/08/13/4-io-robustness-a-
> sensitive-subject/ and the related full documentation.
>
>
> Regards, Ton.
>
>
> ----- Original Message -----
>
> From: Steve Dugas
>
> To: amibroker@xxxxxxxxxps.com
>
> Sent: Thursday, February 05, 2009 5:24 PM
>
> Subject: Re: [amibroker] SPSO vs Trib vs CMAE, was: random
> optimization?
>
> Hi Ton,
>
>
> I am talking about the ones built-in to AB - SPSO, TRIB abd
> CMAE. But, did you see Mike's reply about running them on all
> stocks? After seeing that, it hit me that I usually run my
> opt's on 1 ticker at a time in order to see detailed results for
> that ticker, Mike pointed out that it can take days or weeks if
> you run it on a large number of tickers. I don't know how many
> tickers you run your opts on but anyway I am sorry if my answer
> was misleading...
>
>
> Steve
>
> ----- Original Message -----
>
> From: Ton Sieverding
>
> To: amibroker@xxxxxxxxxps.com
>
> Sent: Thursday, February 05, 2009 6:01 AM
>
> Subject: Re: [amibroker] SP SO vs Trib vs CMAE, was: random
> optimization?
>
> Thanks for yr answer Steve. Your are talking about "IO
> Engines". Can you tell me which engines you are using ? AB's,
> Fred's IO or ...
>
>
> Regards, Ton.
>
>
> ----- Original Message -----
>
> From: Steve Dugas
>
> To: amibroker@xxxxxxxxxps.com
>
> Sent: Wednesday, February 04, 2009 7:31 PM
>
> Subject: Re: [amibroker] SPSO vs Trib vs CMAE, was:
> random optimization?
>
>
> Hi Ton - It is the norm for me, when I start an
> optimization, for AB to tell me that it will take several months
> or more to run, I think I remember a few where it said 50 or 100
> *years*, something like that. But with the IO engines you can
> go ahead and run it anyway, for me they always finish probably
> within an hour, sometimes much quicker. The times can vary a
> bit, I think maybe it depends on where in the opt space they
> start, what paths they take from there and what it leads them
> to... Sometimes they will finish and report results in 5 or 10
> minutes, other times can be an hour or maybe a little more, most
> will be somewhere in the middle. At least that is how it always
> works for me. I definitely agree with you, I am not looking for
> peaks like the one you posted either but sometimes there are
> smaller and more profitable plateaus that are tradable for
> months and I like to at least find them and know they are there...
>
>
> Steve
>
> ----- Original Message -----
>
> From: Ton Sieverding
>
> To: amibroker@xxxxxxxxxps.com
>
> Sent: Wednesday, February 04, 2009 11:27 AM
>
> Subject: Re: [amibroker] SPSO vs Trib vs CMAE, was:
> random optimization?
>
>
> Your are talking about ".... up into the trillions and
> more.". How are you handling the time problem ? These should be
> optimizations of several months ... Even with CMAE etc. it will
> be still a question of weeks ( about 1/4 of the time ). Let's
> say I want to optimize 100 different systems on 10 different
> Symbols or 1.000 combinations times your trillions. It's a life
> time. I just do not have a solution for this. Do you ?
>
>
> And again optimization on points is not what I would
> like to do. Because of the underneath picture ... I would like
> to optimize on areas in stead of points. I hoped to get this
> with CMAE. The result was negative ...
>
>
> Regards, Ton.
>
>
> ----- Original Message -----
>
> From: Steve Dugas
>
> To: amibroker@xxxxxxxxxps.com
>
> Sent: Wednesday, February 04, 2009 12:21 AM
>
> Subject: Re: [amibroker] SPSO vs Trib vs CMAE, was:
> random optimization?
>
>
> Hi Ton - The 2 MA crossover system was just a simple
> example for illustration purposes. In real life I would do an
> exhaustive opt on that one since it would only have maybe 100 x
> 100 = 10,000 combinations, and perhaps the small optimization
> space is why CMAE was able to find the peak. The systems I test
> with the IO engines generally have at least millions of possible
> combos and some up into the trillions and more. FWIW, I have
> done lots of these tests and I will have to stand by my earlier
> remarks because that is my honest experience, but perhaps others
> may see different results...
>
>
> Steve
>
> ----- Original Message -----
>
> From: Ton Sieverding
>
> To: amibroker@xxxxxxxxxps.com
>
> Sent: Tuesday, February 03, 2009 3:40 PM
>
> Subject: Re: [amibroker] SPSO vs Trib vs CMAE,
& gt; was: random optimization?
>
>
> Interesting ...
>
>
> ..... results, then ran lots of IO tests and
> compared them to the exhaustive
> results to see what the IO's found and also what
> they missed. You could say
> that CMAE seems to take the "safe" approach, IMHO
> it finds the broad
> plateaus pretty well but as you might guess they
> are usually far from the
> most profitable. In my experience, the other two
> IO engines will generally
> find those too but they also find a lot of the
> smaller and more profitable
> ones, which you can then run a mini exhaustive opt
> on to get a more complete
> picture ....
>
>
> Is that true ? Does CMAE really take the 'safe'
> approach ? Look to following
>
> picture and see what CMAE gave me as an optimum ...
>
>
>
>
>
> I got the left peak and hoped to get the plateau
> in the middle ...
>
>
> Regards, Ton.
>
>
>
>
> ----- Original Message -----
>
> From: Steve Dugas
>
> To: amibroker@xxxxxxxxxps.com
>
> Sent: Tuesday, February 03, 2009 7:54 PM
>
> Subject: Re: [amibroker] SPSO vs Trib vs CMAE,
> was: random optimization?
>
>
> Hi Steve - Once you have done an IO and found
> some results that look
> promising, then you can run a mini exhaustive
> opt if you want. For a simple
> example, you run an IO on a MA crossover system,
> testing both MA's with
> periods from 1 to 100. You won't see all
> possible combos reported but maybe
> the results show that MA1=10 and MA2=20 might be
> good. So to see all the
> other values in that neighborhood you could then
> run a little exhaustive
> opt, say MA1 = 5 thru 15 and MA2 = 15 thru 2 5,
> something like that, which
> will run in a reasonable time.
>
> To test the built-in IO engines, I ran a few
> exhausive opts and saved the
> results, then ran lots of IO tests and compared
> them to the exhaustive
> results to see what the IO's found and also what
> they missed. You could say
> that CMAE seems to take the "safe" approach,
> IMHO it finds the broad
> plateaus pretty well but as you might guess they
> are usually far from the
> most profitable. In my experience, the other two
> IO engines will generally
> find those too but they also find a lot of the
> smaller and more profitable
> ones, which you can then run a mini exhaustive
> opt on to get a more complete
> picture.
>
> Regarding the trade-off you mentioned, I would
> think it is a matter of
> personal taste. How greedy are you? 8 - ) How
> risk-averse? I am
> inclined to try the smaller and higher plateaus
> first, as long as they have
> a little play on each side and are doing well
> right now, and knowing that
> they will fail eventually and I need to keep a
> close eye on them... Good
> luck!
>
> Steve
>
> ----- Original Message -----
> From: "Steve Davis" <_sdavis@xxxxxxcom>
> To:
> Sent: Monday, February 02, 2009 5:01 PM
> Subject: [amibroker] SPSO vs Trib vs CMAE, was:
> random optimization?
>
> > Steve,
> >
> > I would like to hear more about your system
> optimization process. How
> > were you able to determine the size of the
> plateaus discovered by the
> > built-in optimizers, and how did you decide
> which solutions had the
> > best trade-off between plateau size and
> profitability? >
> > Thanks,
> > another Steve
> >
> > --- In amibroker@xxxxxxxxxps.com, "Steve
> Dugas" wrote:
> >>
> >> Hi - I have spent lots of time playing with
> the built-in intelligent
> >> optimizers, in my experience SPSO will return
> the same results every
> > time if
> >> the settings are the same. Trib and CMAE will
> probably return different
> >> results each time. FWIW, I find CMAE to be
> the worst of the three and I
> >> don't use it anymore, it will find plateaus
> but nearly always misses
> > the
> >> much more profitable but smaller plateaus.
> Using a quad-core I can
> > run 4
> >> simultaneous instances and I find that by
> running 1 SPSO and 3
> > Trib's and
> >> then comparing the 4 results together, it
> will generally point me to
> > some
> >> pretty good param values. Good luck!
> >>
> >> Steve
> >>
> >> ----- Original Message -----
> >> From: "gabriel_id@..."
> >> To:
> >> Sent: Monday, February 02, 2009 7:25 AM
> >> Subject: [amibroker] Re: random optimization?
> >>
> >>
> >> > OK..
> >> >
> >> > Can u give me what type of engine and with
> what kind of settings will
> >> > get the same results when i optimize this lines:
> >> >
> >> > N = Optimize("N-minutes", 33, 1, 60, 1);
> >> > TimeFrameSet( N * in1Minute );
> >> > MA1 = MA( Close, 10);
> >> > MA2 = MA( Close, 20);
> >> > BuySignal = Cross( MA1, MA2);
> >> > sellSignal = Cross( MA2, MA1);
> >> > TimeFrameRestore();
> >> >
> >> > Buy = TimeFrameExpand(BuySignal , N*in1Minute);
> >> > Sell = TimeFrameExpand(sellSignal , N*in1Minute);
> >> >
> >> > I tried cmae, 5 , 1000, have variable
> results.. on walkforward
> >> > i tried spso, 5, 1000, same variables results..
> >> > and also trib, 5, 1000..
> >> >
> >> >
> >> > --- In amibroker@xxxxxxxxxps.com, "Mike"
> wrote:
> >> >>
> >> >> Tribes is a non exhaustive optimizer,
> meaning that it does not
> >> >> evaluate every possible combination.
> >> >>
> >> >> As such, it is possible that it will find
> different "optimal"
> >> >> solutions every time, depending on the
> nature of the surface being
> >> >> optimized. For example; If the surface has
> many similar peaks, it may
> >> >> land on a different one each time (local
> optima) instead of the one
> >> >> true optimal solution (global optima).
> >> >>
> >> >> Try increasing the number of Runs and/or
> MaxEval. If you have more
> >> >> than 2 or 3 optimization variables, 1000
> MaxEval is not enough.
> >> >>
> >> >> http://amibroker.com/guide/h_optimization.html
> >> >>
> >> >> Mike
> >> >>
> >> >> --- In amibroker@xxxxxxxxxps.com,
> "gabriel_id@" wrote:
> >> >> >
> >> >> > hi there,
> >> >> >
> >> >> > i am a bit confused, i run the same
> optimization process.. on same
> >> >> > data range.. and i got different results
> each time :)
> >> >> >
&g t; >> >> > and the engine was trib, 5, 1000...
> >> >> >
> >> >> > thx,
> >> >> > GV
> >> >> >
> >> >>
> >> >
> >> >
> >> >
> >> > ------------------------------------
> >> >
> >> > **** IMPORTANT ****
> >> > This group is for the discussion between
> users only.
> >> > This is *NOT* technical support channel.
> >> >
> >> > *********************
> >> > TO GET TECHNICAL SUPPORT from AmiBroker
> please send an e-mail
> > directly to
> >> > SUPPORT {at} amibroker.com
> >> > *********************
> >> >
> >> > For NEW RELEASE ANNOUNCEMENTS and other
> news always check DEVLOG:
> >> > [BLINK: http://www.amibroker.com/devlog/]
>
> >> >
> >> > For other support material please check also:
> >> > http://www.amibroker.com/support.html
> >> >
> >> > *********************************
> >> > Yahoo! Groups Links
> >> >
> >> >
> >> >
> >> >
> >>
> >
> >
> >
> > ------------------------------------
> >
> > **** IMPORTANT ****
> > This group is for the discussion between users only.
> > This is *NOT* technical support channel.
> >
> > *********************
> > TO GET TECHNICAL SUPPORT from AmiBroker please
> send an e-mail directly to
> > SUPPORT {at} amibroker.com
> > *********************
> >
> > For NEW RELEASE ANNOUNCEMENTS and other news
> always check DEVLOG:
> > http://www.amibroker.com/devlog/
> >
> > For other support material please check also:
> > http://www.amibroker.com/support.html
> >
> > *********************************
> > Yahoo! Groups Links
> >
> >
> >
> >
>
>
>
> -----------------------------------------------------------------
> -------------
> I am using the Free version of SPAMfighter
> We are a community of 5.9 million users fighting spam.
> SPAMfighter has removed 642 of my spam emails to date.
> The Professional version does not have this message
>
>
>
>

---

I am using the Free version of SPAMfighter
We are a community of 5.9 million users fighting spam.
SPAMfighter has removed 642 of my spam emails to date.
The Professional version does not have this message

__._,_.___

**** IMPORTANT ****
This group is for the discussion between users only.
This is *NOT* technical support channel.

*********************
TO GET TECHNICAL SUPPORT from AmiBroker please send an e-mail directly to
SUPPORT {at} amibroker.com
*********************

For NEW RELEASE ANNOUNCEMENTS and other news always check DEVLOG:
http://www.amibroker.com/devlog/

For other support material please check also:
http://www.amibroker.com/support.html

*********************************

Your email settings: Individual Email|Traditional
Change settings via the Web (Yahoo! ID required)
Change settings via email: Switch delivery to Daily Digest | Switch to Fully Featured
Visit Your Group | Yahoo! Groups Terms of Use | Unsubscribe

__,_._,___