Hi Joe, I did some tests on WSJT-X r5954 this evening. I added both sets (JT and SF) of erasure probabilities and ran some different combinations. I didn’t change any settings other than to comment or un-comment the necessary lines in demod64a and in sfrsd2.
1. matched: jt symbol metrics and jt erasure probs: 607 2. mismatch: sf symbol metrics and jt erasure probs: 523 3. matched: sf symbol metrics and sf erasure probs: 594 Overall, numbers are still less than WSJT, but at least now the numbers are the right way around. And all results are much better than the best result (448 I think) that I got when I tried it yesterday. So whatever knob you turned, just turn it a bit more… If there are differences in the way that the s3 array is treated in WSJT and WSJT-X, such as differences in flattening algorithm and maybe noise level estimation, then I suppose that it’s possible that a re-tuning of the erasure probs might improve things a bit. Eventually, it might be useful to add back in the code that accumulates statistics (in the 8x8 matrix form) and also enable sfrsd2 to read in the accumulated statistics without manual intervention, i.e. push-buttonize the self-tuning capability of the system. This would make it easy to tune up the algorithm for HF-type files. I quickly tried your spurious-vector fix on one of my 20m batches and it seems to be very effective, reducing the number of spurious vectors by a large factor. There are still some files that produce 8 or more false decodes and only 2 decodes, but overall it is much better. Last thing - where do I go to turn off the blank lines for the failed decodes? I’d like to disable those and then run my HF files to compare with previous results. Steve k9an ------------------------------------------------------------------------------ _______________________________________________ wsjt-devel mailing list wsjt-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/wsjt-devel
