Alayan wrote: ↑Sun Jan 05, 2020 9:42 am
lucasart wrote: ↑Sun Jan 05, 2020 7:02 am
SF uses skip depth scheme, so the reported depth is misleading and massively understated.
This is wrong. Depth skipping has been removed from Stockfish over 9 months ago, because it scales awfully at longer TCs. See
https://github.com/official-stockfish/S ... /pull/1972
Current Stockfish doesn't do it.
Indeed:
Code: Select all
commit 66818f2e85732644708e23b3f2c2e544abfbc3b0 (HEAD)
Author: CoffeeOne <peter.zsifkovits@gmx.at>
Date: Wed Mar 20 14:50:41 2019 +0100
Skip skipping thread scheme (#1972)
Several simplification tests (all with the bounds [-3,1]) were run:
5+0.05 8 threads, failed very quickly:
http://tests.stockfishchess.org/tests/view/5c439a020ebc5902bb5d3970
20+0.2 8 threads, also failed, but needed a lot more games:
http://tests.stockfishchess.org/tests/view/5c44b1b70ebc5902bb5d4e34
60+0.6 8 threads passed:
http://tests.stockfishchess.org/tests/view/5c48bfe40ebc5902bca15325
60+0.6 4 threads passed:
http://tests.stockfishchess.org/tests/view/5c4b71a00ebc593af5d49904
No functional change.
Saying it scales awfully is a bit exagerated. All we can say is that:
* it clearly helps at short tc
* as tc increases, it helps less and less
* at very long tc, it becomes practically useless (un-mesurable, hard to distinguish signal from noise these sprt(-3,1) results are difficult to interpret, especially in the presence of p-hacking).
But yes, it seems to be asymptotically useless. In fact, all evidence points to this simple conclusion:
pure SHT is asymptotically optimal. Yes, this is a very bold claim, a flamewar bait, which I am sure will not fail
In fact, I wouldn't be surprised if all this (ugly) thread voting logic in SF is also asymptotically useless.
I've tried to remove the d/d+1 scheme in demolito, and it failed hard:
http://chess.grantnet.us/viewTest/4170/
So, probably a clear regression, although the test bounds were not decided (in advance of the test) to prove as much.
I'll re-spin it by doubling the tc from 4+0.04 to 8+0.08. Of course, the draw rate will also increase, so I can't just compare elo in 4" to elo in 8", but perpahs comparing bayes elo, or wilo, or Michel's normalized elo metric...
Theory and practice sometimes clash. And when that happens, theory loses. Every single time.