[Scilab-users] Problems with Arx-algorithm

[A Khorshidi]

Hi.

>From the simulation point of view, the resulting model may not be
satisfactory due to two main reason; 1: inappropriate model structure,
2:Nosiy data.

If you intend to use only the ARMAX model, then you should change the orders
of the polynomials so that a good estimated model can be obtained. 
So, you've also change the orders four times. Why not try more other orders? 

As far as I know, the fitness value between 85-90% is acceptable. So, as a
side note, the normalized root mean square error can be used to compare
estimated data (ysim) with measured data (y): 
fitness_percentage = 100*(1-(norm(y-ysim)/norm(y-mean(y)))) 

However, if the fitness value was not admissible, the measured data may be
contaminated by noise. Then, you should filler the measurements before
starting the parameter estimation. Use the frequency response of an
estimated model to find the frequency range which needs to be filtered.  


HTH
Mehran
_




--
View this message in context: http://mailinglists.scilab.org/Problems-with-Arx-algorithm-tp4030499p4030512.html
Sent from the Scilab users - Mailing Lists Archives mailing list archive at Nabble.com.