Joe - I like those results! But your S/N axis is shifted by 1 dB relative to what I would expect. As I recall, we were getting something like 837 decodes on the earlier set of -24dB files. In this latest plot, it looks like sfrsd is producing almost 1000 decodes at -24 dB. Steve k9an > On Nov 19, 2015, at 11:03 AM, Joe Taylor <[email protected]> wrote: > > Hi Steve and all, > > I've attached a plot comparing the sensitivities for decoding JT65A in WSJT > 10.0 (r6088), WSJT-X v1.6-rc1, and WSJT-X v1.7 (r6131). The simulated > channel is AWGN; the demodulators all use noncoherent measurement of symbol > amplitudes. Correlation decoding and message averaging were not used. > > Test measurements were made using 1000 simulated JT65A transmissions at each > of the SNRs -20, -21, -22, ... -29 dB. I ran each program on the full set of > files and counted the decodes for each generated SNR. > > WSJT 10 uses the KV decoder and was tuned (a long time ago) for best > performance with EME signals. WSJT-X v1.6 also uses KV, but in this case we > tuned for decoding speed, good performance, and few false decodes on crowded > HF bands. WSJT-X v1.7 uses the sfrsd decoder. For these tests I set "Random > erasure patterns" to 8, which means ntrials=10^4. > > As you can see in the plot, in these tests WSJT-X v1.7 is around 0.7 dB > better than WSJT10 and 1.3 dB better than WSJT-X v1.6. I think this is truly > superb performance. > > For EME applications we'll want to implement message averaging and > correlation decoding. This should buy us another 4 dB, more or less. These > decoding approaches are already present in JT4 mode, and they work well. > > I also need to make comparable tests and measurements for a simulated > Rayleigh-fading channel, and for submodes JT65B and JT65C. I will turn some > attention to these tasks, as time permits. > > -- 73, Joe, K1JT > <JT65A_decoding.pdf>------------------------------------------------------------------------------ > _______________________________________________ > wsjt-devel mailing list > [email protected] > https://lists.sourceforge.net/lists/listinfo/wsjt-devel
------------------------------------------------------------------------------ _______________________________________________ wsjt-devel mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/wsjt-devel
