Hi Steve and all, A few more thoughts about my idea on how we might select erasure vectors most likely to lead to successful decodes. This is potentially important because the "right" erasure vector leads to a nearly instantaneous decode; nearly all of the processing time in sfrsd2 is now devoted to failed processing of randomly selected "wrong" erasure vectors. If we can tilt the balance toward testing likely-to-be-right erasure patterns sooner than unlikely ones, processing time will be reduced.
With this motivation I compiled histograms of the number of symbol values retained (i.e., not erased) as a function of where each symbol fell in the list ranked by p1, its estimated probability of being correct. Two cases are shown in a graph posted here: http://physics.princeton.edu/pulsar/K1JT/erasures.pdf The dotted curve shows the fraction of symbols retained at each rank position for all erasure patterns tried (about 2 million of them) in a run with 1000 simulated "transmissions" and ntrials=10000 per transmission. The solid curve is the corresponding quantity for the much smaller number of 809 erasure patterns that led to correct decodes. Several features are clearly evident in the histograms. You can see the effect of our present empirically-determined boundaries at rank indexes i = 32 and 38. Correct decodes generally occur with fewer retained symbols (more erasures) in the region i > 38 than for trials that did not decode. This seems to suggest increasing the probability erasure for large i. We should probably soften the present step-wise boundary at i=38, for which there is no physical motivation. We might also look at how these histograms may change when restricted to distinct ranges of p2/p1. Comments? -- Joe ------------------------------------------------------------------------------ _______________________________________________ wsjt-devel mailing list wsjt-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/wsjt-devel
