Hi Steve,
Agreed on all counts.
A surprise discovery: it seems that all of our recent tests have been
using the original code in .../trunk/demod64a.f90, with its exp(x)
symbol metrics. I tried switching back to your simple p1/psum, p2/psum
mnetrics computed in .../trunk/rsdtest/demod64b.f90, with much degraded
results. I did not try re-tuning things accordingly.
-- Joe
On 9/30/2015 7:23 AM, Steven Franke wrote:
> Thanks Joe. Interesting. It could be, I suppose, that with large ntrials the
> mr2syms are helping us find more decodes, but they are rejected by the soft
> distance check.
>
> In any case, I agree that we should just stick with the current erasures-only
> scheme. but let's keep sfrsd3 on the side and see what it does when we get to
> testing real signals.
> Steve k9an
>
>> On Sep 30, 2015, at 5:27 AM, Joe Taylor<j...@princeton.edu> wrote:
>>
>> Hi Steve
>>
>> I played with your mr2 insertion code a bit. I agree that the
>> second-best symbols don't seem to help us much.
>>
>> I posted two new plots of decodes vs. ntrials here:
>>
>> http://physics.princeton.edu/pulsar/K1JT/decodes_vs_ntrials.pdf
>> http://physics.princeton.edu/pulsar/K1JT/decodes_vs_ntrials3.pdf
>>
>> The first one is like before, but extended out to ntrials=1,000,000.
>> Interestingly, only 10 more good decodes (plus 5 bad ones) were found at
>> trial numbers above 156,000. However, 44 good decodes (and no bad ones_
>> were found between 10,000 and 100,000.
>>
>> The second plot compares results with erasures only (solid line) and
>> with erasures and mr2 substitutions (dotted line). As you had noticed,
>> with substitutions we get to a stated number of decodes in a smaller
>> number of trials, up to around ntrials=1000. Above ntrials=10,000 there
>> seems to be no gain. Moreover, the substitutions code is 1.6 times
>> slower.
>>
>> It seems we're better off using erasures only, with the soft information
>> used to compute a soft distance for each potential codeword.
>>
>> -- Joe
>>
>> ------------------------------------------------------------------------------
>> _______________________________________________
>> wsjt-devel mailing list
>> wsjt-devel@lists.sourceforge.net
>> https://lists.sourceforge.net/lists/listinfo/wsjt-devel
>
> ------------------------------------------------------------------------------
> _______________________________________________
> wsjt-devel mailing list
> wsjt-devel@lists.sourceforge.net
> https://lists.sourceforge.net/lists/listinfo/wsjt-devel
------------------------------------------------------------------------------
_______________________________________________
wsjt-devel mailing list
wsjt-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/wsjt-devel
