HNY All! I’m writing to report the results of some tests on r6330, which includes Joe's new scheme for assigning a quality metric to the codewords that are found by the soft-decision Franke-Taylor decoder.
First, results for my batch of 333 hf files: r6041: 4181 (ntrials=1000, robust sync) r6330: 4283 (naggressive=0, ntrials=1000, robust sync) I also compared results from r6330 to our earlier benchmark results for simulated data in AWGN. The attached figure shows the difference between r6315 and r6330 on simulated data, using naggressive=10 and T=10^5. It is clear that the changes had the most significant effect for the lowest SNRs, where the new metric does a much better job of separating good from bad codewords, allowing us to run with the acceptance criterion opened up.
fig_wer3.pdf
Description: Adobe PDF document
Interestingly, there is a slight improvement over the previous version at high SNRs as well. In the second figure, the latest results from r6330 are shown as a dashed line with filled triangle markers.
fig_wer4.pdf
Description: Adobe PDF document
All-in-all - across-the-board improvements are evident and, based on my test results, r6330 has the best jt65a decoder performance of any version that we have produced to date. Congratulations Joe! Steve k9an
------------------------------------------------------------------------------
_______________________________________________ wsjt-devel mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/wsjt-devel
