Adaptive digital filters - page 3

 
Mathematician, sometimes I don't read what you write in Russian as well (I don't draw the same conclusions as you expect :-)). But here in English. Although I understand what it says here that it is used by the military to escort moving objects. I can't read it verbatim (I learnt English in my childhood :-() And redrawing and FFT are different things, I can use FFT and nothing will be redrawn. I'll go to Coshi now and read it. And military knows how to track the target :-). Only jamming and decoys save from guaranteed hit, and not always :-) (There has been a battle of wits and algorithms going on for some time now, the pilot in the aircraft needs only to push the button and ..... It seems that everything else is algorithms).
 
Prival:
Also redrawing and FFT are different things, you can use FFT and nothing will be redrawn. For Cauchy I will go now, to read.

Oh, how interesting. And I thought that all pseudo-machines based on the principle of "transformation - frequency filtering - reverse transformation" are necessarily redrawn...

And don't waste any time on Cauchy, its p.d.f. - a/(b^2 + (x-m)^2). With normalisation by one, of course. But the integral of p.d.f. multiplied by the variable x is already divergent (this is m.o.).

 
Mathemat:
Prival:
Yes also redrawing and FFT are different things, you can use FFT and nothing will be redrawn. For Cauchy, I'll go now, to read.

Oh, how interesting. And I thought that all pseudo-domains based on the principle of "conversion - frequency filtering - reverse conversion" are necessarily redrawn...


Prival is right, it all depends on which filter structure/scheme to use. If really interesting, for example, like this:


There is no redrawing. I used to have a lot of fun with these filters myself.

 
Something about the JMA, like the best, adaptive, etc., hit me. (all eaten up, how). And we have a good job :-). And the left-handed like Russia is no more, but I do not believe it.
I look, look at him - some strange formulas, and the avatar is not something like :-) I like it better :-).
(Compare http://www.jurikres.com/catalog/ms_ama.htm#top). Our aeroplane is better :-).

That`s why I suggest to try to make a better indicator, more adaptive. Maybe something good will come out.

The idea is the following.
1. We take this indicator as a base ('Kaufman optimized AMA: Perry Kaufman AMA optimized'), many people have already worked on it. The theory of this indicator is described in the file (file attached). We take one part of this indicator (idea). Calculation of the ER efficiency ratio (varies from 0 to 1). It will determine the averaging (sampling) period from 2 to N (N is set as an input parameter in the algorithm). The rest is a bit trickier.
2 We don't use EMA (exponential moving average) but a polynomial. The maximum power of the polynomial n (also set as an external parameter). In principle, we can stop and vary n and run in the tester, I think we can already get good results. But IHMO the flea is not yet fully trained, so let's move on.
3. If it's adaptive, then let it be adaptive to the fullest extent. In addition, the next one - the degree of polynomial is also calculated (chosen the best one by some criterion). Since we have no a priori information on noise. I suggest using the criterion - the coefficient of determination. The logic of selecting the optimal polynomial according to this criterion is described in the file (see pp. 12, 13 and 14). There is even a program written in MathCade, how to do it.

If anyone is interested, I am ready to program and recheck point 3 in MathCade. I will also help you to create such indicator in MQL due to my modest capabilities.
Files:
 
And also don't forget that the inductor simply has to be non-linear. Anyway, that's what Djuric's team decided in order to satisfy all four requirements of an ideal adaptive filter. And there's also something of information theory involved... grasn, do you have any ideas about a non-linear filter scheme?
 
to Prival, Mathemat
<br / translate="no">The idea is the following.
1. As the basis we take this indicator ('Kaufman optimized AMA: Perry Kaufman AMA optimized'), many people have already worked on it. The theory of this indicator is described in the file (file attached). We take one part of this indicator (idea). Calculation of the ER efficiency ratio (varies from 0 to 1). It will determine the averaging (sampling) period from 2 to N (N is set as an input parameter in the algorithm). The rest is a bit trickier.
2. we do not use EMA (exponential moving average) but a polynomial. the maximum degree of the polynomial is n (also set as an external parameter). we can stop and vary n and run it in the tester, I think we can already get good results. But IHMO the flea is not yet fully trained, so let's move on.
3. If it's adaptive, then let it be adaptive to the fullest extent. In addition, the next one - the degree of polynomial is also calculated (chosen the best one by some criterion). Since we have no a priori information on noise. I suggest using the criterion - the coefficient of determination. The logic of selecting the optimal polynomial according to this criterion is described in the file (see pp. 12, 13 and 14). There is even a program written in MathCade, how to do it.


My humble self-taught opinion is this: the proposed "adaptive filter" model will not work, I will not waste my time on it. This is anything but adaptive filtering. There is a coherent, coherent, proven theory of adaptive filtering. And if you want to make exactly adaptive filter - you'd better use exactly this theory.

If you don't have time to understand this theory and design AF, then take MathLab and build required filter (if not an expert in adaptive filtering, MathLab will do it much better). Further you have two ways: either generate dll or use m-files to convert them to MQL, thanks God they are open.

 
grasn:

My humble self-taught opinion is this: the proposed "adaptive filter" model will not work, I will not waste my time on it. This is anything but adaptive filtering. There is a coherent, coherent, proven theory of adaptive filtering. And if you want to make exactly adaptive filter - you'd better use exactly this theory.

If you don't have time to understand this theory and design AF, then take MathLab and build necessary filter (if not an expert in adaptive filtering, then MathLab will do it much better). Then there are two ways: either generate dll or dig in m-files, transferring them to MQL, thank God they are open.

It's hard not to agree with this opinion.
 
I suggest that instead of fooling around, hoping for the best of luck, you should try the clones of Djuric's inductor, posted here, and check whether they are that good, firstly using a simple pseudo-moving crossing system as an example.
 
NorthernWind:
grasn:

My humble self-taught opinion is this: the proposed 'adaptive filter' model will not work, I will not waste my time on it. It is anything but adaptive filtering. There is a coherent, coherent, proven theory of adaptive filtering. And if you want to make exactly adaptive filter - you'd better use exactly this theory.

If you don't have time to understand this theory and design AF, then take MathLab and build required filter (if not an expert in adaptive filtering, MathLab will do it much better). Then you have two ways : either generate dll or go through m-files and put them into MQL, thank God they are open.

It's hard to disagree with that opinion.


I wonder what's the opinion. That I don't know DSP and in particular one of the topics I used to read lectures on (adaptive digital filters). Or that it is better to do it in Matlabe? I think the author is wrong there and there. I have a "little" knowledge in this field, and there is a better programming language than MathLaba. I don't need any dll to send calculation results to MT4 terminal (I just need komposter).

It seems to me that writing about my suggestion and saying that there is no adaptive filtering there, grasn is wrong. And will not be able to answer where, when and for what reason, say, it is necessary to apply Hemming's window, and when its application only harms. What is the difference between Wiener adaptive filter and Widrow-Hopf filter when analyzing their FFC or Butterworth filter and Chebyshev filter, when it is necessary and possible to apply the first filter, and when the second one.

I'm sorry if I came across harshly, but you can't just dismiss ideas as passé. It takes me 1-2 hours at most to program all I have written in MathCade and I don't need anyone's help for that. Wanted to help others to show the direction to dig, if they want to get an adaptive filter, and ready to help them in this case. Adaptive filters are a sea and a small trolley of them.

That you aren't so angry let me give you as MathLaba lover a book about DSP, there is 989 pages about this thing DSP, a lot of examples in this programming language, but in my humble opinion MathCad is better :-)

Files:
read_me.zip  9488 kb