# Compare LMS NLMS

Started by April 22, 2005
```Hi

I have some Problems with the NLMS. I have written a MATLAB skript to
simulate the LMS and NLMS now I will compare these two algorithms.
Especially the convergence speed. Here I will compare the learning
curves.
Inputsignal: White noise (Matlab randn())

My major Problem is the choose of the different Parameters like step
size and so on.

The coefficient-update: w(n+1)=w(n)+mu(n)*e(n)*x(n)

mu(n)=alpha/(L*P(n))

P(n)=(1-beta)*P(n-1)+beta*x(n)*x(n)

alpha= normalization between 0 and 2
beta = windowsize for Powerestimation

So has anybody experience with the NLMS und the choose of the
Parameters?

Thanks Markus
```
```P&#2013265925; 22 Apr 2005 00:47:20 -0700, skrev Gie78 <Markus.Schweikhardt@web.de>:

> Hi
>
> I have some Problems with the NLMS. I have written a MATLAB skript to
> simulate the LMS and NLMS now I will compare these two algorithms.
> Especially the convergence speed. Here I will compare the learning
> curves.
> Inputsignal: White noise (Matlab randn())
>
> My major Problem is the choose of the different Parameters like step
> size and so on.
>
> The coefficient-update: w(n+1)=w(n)+mu(n)*e(n)*x(n)
>
> mu(n)=alpha/(L*P(n))
>
> P(n)=(1-beta)*P(n-1)+beta*x(n)*x(n)
>
> alpha= normalization between 0 and 2
> beta = windowsize for Powerestimation
>
> So has anybody experience with the NLMS und the choose of the
> Parameters?

My best advice is that you experiment with those parameters. For random
signals, mu(n) generally should be small. There is also another method
that I used:
P(n) = 1/L*x(n:-1:n-L+1)*x(n:-1:n-L+1)

Sverre
```