Mike --
On 9/29/2015 5:29 PM, Michael Black wrote:
> So the failures are a separate section?
No. Have you looked at what the decoder is doing?
If 25 or fewer of 63 received symbols are in error, the deterministic
Berlekamp_Massey (BM) algorithm is guaranteed to succeed. With more
than 25 errors BM may still succeed if we know (or can guess) which
symbol values are likely to be incorrect.
We use a stochastic approach, making many "partly random" guesses about
which symbols (typically anywhere from 37 to 51 of them) might be in
error. These selected symbols are said to be "erased", and the BM
algorithm assigns their values a "don't care" status.
The "erasure vector" (list of symbols to be erased) is only partly
random because we can rank the list of 63 symbols by their relative
reliability. Clearly it makes sense to assign low-reliability symbols a
higher probability of erasure. We also assign higher erasure
probability to cases where the second-best symbol value was "almost as
good" as the first-best value.
When the BM algorithm succeeds in returning a potentially valid
codeword, we measure its "soft distance" from the 63 received symbol
values. If the estimated reliability of a erroneous symbol was high,
the incremental soft distance contribution is large, and vice-versa.
The inner program loop keeps trying different erasure vectors, seeking
the minimum soft distance. When a minimum is found that is low enough,
a decode is declared. If no small-enough minimum is found after
"ntrials" attempts, a decoding failure is declared.
> Question would be how often in real life do failures occur?
For all modes and decoders in WSJT and WSJT-X, upstream parts of the
program make basic tests of signal quality, symbol reliability, etc.,
that will minimize the number of decoding attempts that are almost sure
to fail.
> And would you make the # of trials dependent on # of cores?
No. But some user control over the attempted "depth" of decoding may be
provided.
-- Joe, K1JT
------------------------------------------------------------------------------
_______________________________________________
wsjt-devel mailing list
wsjt-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/wsjt-devel
