Hi Joe, > JT65-HF 2329 24 BM + kvasd > WSJT-X r5912 2249 0 BM + kvasd > WSJT-X r5955 2114 0 BM + sfrsd > WSJT-X r5955 1816 0 BM only
For what it’s worth - on a batch of my HF files, r5922, using either kvasd or sfrsd and with the ncount threshold set to 0, produces significantly more decodes than our current sfrsd-based 5955. It’s something like 736 total decodes from r5922 vs 686 from r5955. Even after I attempted to tune 5955 I managed to move it up to only about 700 decodes, still significantly below what r5922 produced. And r5922 uses your very simple formula for setting the erasure probabilities. With the ncount threshold removed, r5922 does suffer from the problem of very high spurious symbol-vector flux at the decoder input and it often bogs down - but now I’m thinking that perhaps that flux of mostly spurious vectors is accompanied by a number of good ones that we may now be losing after the sync-related fix that you implemented to lower the flux. Another possibility is that r5955 is using tighter constraints on nhard, nsoft, etc. at the decoder output. In any case, I suspect that r5922 with sfrsd (with ncount threshold removed) might compare more favorably with your second line, above. Steve k9an ------------------------------------------------------------------------------ _______________________________________________ wsjt-devel mailing list wsjt-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/wsjt-devel
