Hi Steve,
Thanks for pushing ahead with the new soft-RS decoder. Very interesting
results, to be sure! I'm delighted to see that a composite matrix for
erasure probabilities works so well.
I've made good progress here, too. I have fully confirmed your results
using rsdtest, and for completeness I've extended the test runs to
include ntrials = 10^5 and 10^6. With s3_1000.bin as input data and
acceptance criterion nsoft<35, these two long runs produce 918 and 952
good decodes, respectively, and no bad decodes. Excellent!
For comparison, recall that WSJT10 (with BM+kvasd) yielded 809 decodes.
I think these results are truly state-of-the-art quality.
We'll probably find that acceptance criteria need to be adjusted
(tightened) somewhat when exposing the decoder to various types of real
data, to keep the rate of false decodes acceptably low. But such
adjustment was necessary with kvasd, also: it's one of the reasons that
the WSJT-X sensitivity for nearly ideal, threshold-level JT65A signals
is not quite as good as that in WSJT.
I still need to trace (and correct?) the reason that the identical
sfrsd2 gives somewhat fewer decodes when called from within WSJT-X than
when called from WSJT. Evidently some step in the path to evaluating
array s3() is now sub-optimal in WSJT-X -- at least for idealized,
threshold-SNR data.
One more idea: I think we should try a scheme similar to what we've done
in the WSPR decoder. After a successful decode of whatever signals are
found in the analyzed passband, subtract those signals from the raw data
and then make a second decoding pass. On a crowded HF sub-band, and
also in an EME pileup situation, I suspect that additional decodes will
be possible.
-- Joe, K1JT
------------------------------------------------------------------------------
_______________________________________________
wsjt-devel mailing list
wsjt-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/wsjt-devel
