LMS
in [DSP]
|
From: HardySpicer on 5 Jul 2010 23:19 Does anybody know for ordinary LMS what the official computation is ie number multiplies,adds etc for n weights. Or is there a big O(.) quoted somewhere a bit like FFTs. Hardy
From: Muzaffer Kal on 6 Jul 2010 00:18 On Mon, 5 Jul 2010 20:19:18 -0700 (PDT), HardySpicer <gyansorova(a)gmail.com> wrote: >Does anybody know for ordinary LMS what the official computation is ie >number multiplies,adds etc for n weights. Or is there a big O(.) >quoted somewhere a bit like FFTs. Ordinary LMS being y = Sx*h, err = d - y and h = h + u*err*x so one needs 3N multiplies and 2N adds per output sample. -- Muzaffer Kal DSPIA INC. ASIC/FPGA Design Services http://www.dspia.com
From: HardySpicer on 6 Jul 2010 04:49 On Jul 6, 4:18 pm, Muzaffer Kal <k...(a)dspia.com> wrote: > On Mon, 5 Jul 2010 20:19:18 -0700 (PDT), HardySpicer > > <gyansor...(a)gmail.com> wrote: > >Does anybody know for ordinary LMS what the official computation is ie > >number multiplies,adds etc for n weights. Or is there a big O(.) > >quoted somewhere a bit like FFTs. > > Ordinary LMS being y = Sx*h, err = d - y and h = h + u*err*x so one > needs 3N multiplies and 2N adds per output sample. > -- > Muzaffer Kal > > DSPIA INC. > ASIC/FPGA Design Services > > http://www.dspia.com I agree with the adds/subtracts. However err and u(mu) are scalar so require only 1 multiply, then they get multiplied by vector x for another N so we have (N+1) + N (from x*h) = 2N+1..? Hardy
From: Muzaffer Kal on 6 Jul 2010 12:36 On Tue, 6 Jul 2010 01:49:40 -0700 (PDT), HardySpicer <gyansorova(a)gmail.com> wrote: >On Jul 6, 4:18�pm, Muzaffer Kal <k...(a)dspia.com> wrote: >> On Mon, 5 Jul 2010 20:19:18 -0700 (PDT), HardySpicer >> >> <gyansor...(a)gmail.com> wrote: >> >Does anybody know for ordinary LMS what the official computation is ie >> >number multiplies,adds etc for n weights. Or is there a big O(.) >> >quoted somewhere a bit like FFTs. >> >> Ordinary LMS being y = Sx*h, err = d - y �and h = h + u*err*x so one >> needs 3N multiplies and 2N adds per output sample. >> -- >> Muzaffer Kal >> >> DSPIA INC. >> ASIC/FPGA Design Services >> >> http://www.dspia.com > >I agree with the adds/subtracts. However err and u(mu) are scalar so >require only 1 multiply, then they get multiplied by vector x for >another N so we have (N+1) + N (from x*h) = 2N+1..? Actually it's even better than that in practice. LMS is quite robust so one usually implements mu as a (programmable?) shift and error as a single power of two operation so the whole mu*err looks like a large barrel shifter. -- Muzaffer Kal DSPIA INC. ASIC/FPGA Design Services http://www.dspia.com
|
Pages: 1 Prev: SI System urged to adopt "hella" Next: SI System urged to adopt |