Gear shifting for adaptive filters

Dear All,

Recently I had posted a thread where I asked about SNR penalty due to
linear interpolation. I managed to solve that problem but there are some
more doubts.

My receiver has adaptive doppler correction and adaptive DFE using LMS.
Now, the filter update constants that I had kept intially are as follows:

Training mode:
     mu1=0.01 -> for Feedforward filter
     mu2=0.003 -> for Feedback filter
     K=2e-5   -> to update interpolation factor
Decision Directed Mode:
     mu1=0.001 -> for Feedforward filter
     mu2=0.003 -> for Feedback filter
     K=2e-5   -> to update interpolation factor


And I was getting my BER plot 6 dB away then what I should get.

And then I tried:

(1) Use square pulse, instead of raised cosine.
(2) In decision directed mode, change mu1=0.002 and K=1e-5

Why using square pulse I get expected results and why not when I use
raised cosine?

But then what if my channel changes during DD mode, which it will change
for sure, will these constants be able to track the change?

Or do I have to keep the value of mu1 same throughout the processing?

I am asking about this because I was stuck at this problem for long time
and just by making the changes as mentioned above, I got expected results.

Your opinion matters a lot.

Thanks.

Chintan

On Mar 26, 6:59&#4294967295;am, "cpshah99" <cpsha...@rediffmail.com> wrote:
> Dear All,
>
> Recently I had posted a thread where I asked about SNR penalty due to
> linear interpolation. I managed to solve that problem but there are some
> more doubts.
>
> My receiver has adaptive doppler correction and adaptive DFE using LMS.
> Now, the filter update constants that I had kept intially are as follows:
>
> Training mode:
> &#4294967295; &#4294967295; &#4294967295;mu1=0.01 -> for Feedforward filter
> &#4294967295; &#4294967295; &#4294967295;mu2=0.003 -> for Feedback filter
> &#4294967295; &#4294967295; &#4294967295;K=2e-5 &#4294967295; -> to update interpolation factor
> Decision Directed Mode:
> &#4294967295; &#4294967295; &#4294967295;mu1=0.001 -> for Feedforward filter
> &#4294967295; &#4294967295; &#4294967295;mu2=0.003 -> for Feedback filter
> &#4294967295; &#4294967295; &#4294967295;K=2e-5 &#4294967295; -> to update interpolation factor
>
> And I was getting my BER plot 6 dB away then what I should get.
>
> And then I tried:
>
> (1) Use square pulse, instead of raised cosine.
> (2) In decision directed mode, change mu1=0.002 and K=1e-5
>
> Why using square pulse I get expected results and why not when I use
> raised cosine?
>
> But then what if my channel changes during DD mode, which it will change
> for sure, will these constants be able to track the change?
>
> Or do I have to keep the value of mu1 same throughout the processing?
>
> I am asking about this because I was stuck at this problem for long time
> and just by making the changes as mentioned above, I got expected results.
>
> Your opinion matters a lot.
>
> Thanks.
>
> Chintan

Have you tried normalised LMS?

>Have you tried normalised LMS?
>

%%%%

Yes. But more or less same result.

Also the interesting thing I found was:

We have recorded two sets of data for off line processing..

For set 1, Normalised LMS gives better result than LMS
For set 2, LMS gives better result than Normalised LMS.

And what I have seen that some real time modems use LMS and they work
fine.

So I dont know and I am more confused about this step size selection.

From my exp, I have seen that, in training mode I keep mu=0.01 and in
decision directed mode mu=0.002.

Have u come accross these kind of problems?

Any idea???

Thanks for ur reply.

Chintan