From: chcant on
Hi, I'm trying to implement a modified MESA identification through a
Genetic Algorithm.
In particoular, during the coefficients optimization I'm not able to
understand how I can solve the problem of the filter stability.
I mean, how can you find the constraint that you have to put on the
coefficients values during the GA in order to abtain an asymptotic stable
AR model?

I also tyied to optimize on the poles position in order to overcome this
kind of problem but with very poor results.
Do you think that a genetic algorithm couldn't be a good algorithm to
optimize over the poles domain?

Thanks in advance
Christian
From: Vladimir Vassilevsky on


chcant wrote:
> Hi, I'm trying to implement a modified MESA identification through a
> Genetic Algorithm.
> In particoular, during the coefficients optimization I'm not able to
> understand how I can solve the problem of the filter stability.
> I mean, how can you find the constraint that you have to put on the
> coefficients values during the GA in order to abtain an asymptotic stable
> AR model?

Don't know what is MESA identification, however the simplest way to
guarantee the stability of AR model is by implementing it in the form of
lattice structure. The reflection coefficients have to be in the range
(-1,1). You can project (+/- infinity) to (+/- 1) using a function like
arctg(x); this will also make AR model well behaved wrt parameters.

> I also tyied to optimize on the poles position in order to overcome this
> kind of problem but with very poor results.
> Do you think that a genetic algorithm couldn't be a good algorithm to
> optimize over the poles domain?

Genetic algorithm is bad regardless. If you have to use genetic
algorithm, that means you don't have any idea what to do.


Vladimir Vassilevsky
DSP and Mixed Signal Design Consultant
http://www.abvolt.com