change a position's mate score from 20 to 29 #60
Conversation
|
Just realized that for this particular 7men position it's enough to download a single wdl and dtz file. So you can keep this pr open for a bit. Will try to repeat the analysis later and report back. Afk now. |
|
I'm also running a local check on this, but without TB. Will let it run till tonight. |
|
So with the additional I assume the low |
|
I had Followed by ~20h of no update. |
|
Is |
|
no, once a mate is found, search doesn't explore deeper than that (since there is no deeper line that can improve the score). It is called mate distance pruning: https://github.com/official-stockfish/Stockfish/blob/070e564c389eb2c263f3982060ab5899b67d0a62/src/search.cpp#L580-L589 I guess in exceptional cases, it could be beneficial to search deeper to find a shorter mate (e.g. populate hash with a new mating line, that helps find the shorter one), but that's not the rule as far as I know. |
|
Right, thanks. I was thinking about something like that, but had noticed in my original output the last few lines: (note the occasional |
|
That occasional higher seldepth means that in that iteration that mate 29 had not yet been found by search, before, in that iteration search had explored a PV with greater depth. |
|
Matefish's ProofNumberSearch finds a mate line pretty quickly but can't prove it with 1 GB Hash (the current maximum) only. Here is the line: I have not verified the PV for correctness ... |
With proving it, do you mean to prove that there cannot be a shorter one? The PV you show is for mate in 29, confirming our earlier findings. |
|
Ich habe übrigens auch mal eine Analyse mit der aktuellen Entwicklungsversion vom Huntsman gestartet. Hier beobachte ich immer wieder den gleichen Ablauf. Eigentlich findet die Suche ein Matt, verliert dieses aber gleich wieder. Hier mal ein Ausschnitt:
Eigentlich sollte dies nicht passieren ... irgendwas passt da nicht. (Um nicht gleich von einem Bug zu sprechen.) |
Nein. Es geht eher darum, dass noch nicht alle Möglichkeiten vollständig geprüft wurden und sicher ist, dass Schwarz kein Schlupfloch mehr hat. |
|
Ok. For the mate in 29 I would take the reported mate scores from SF found by myself and vondele as "proof" then. |
|
Even deeper analysis: |
This PR proposes to raise the
bmvalue for the position8/8/3p4/N2p4/1k6/p2K4/8/1B6 w - -from 20 to 29, see also #52. The new entry inmatetrack.epdreadsOver the last couple of days I have also done a local analysis, using the command
printf "uci\nucinewgame\nposition fen 8/8/3p4/N2p4/1k6/p2K4/8/1B6 w - -\nsetoption name threads value 2\nsetoption name hash value 6000\nsetoption name SyzygyPath value /disk2/syzygy/\ngo infinite\nucinewgame\nquit" | ./stockfish >& log.txt &. The selected output of which is as below:Output
Things to note: this is a 7men position, analysed with 6men syzygy TBs. The score
mate 29and the associated PV is stable fromdepth 79todepth 119. At the end of the analysis the hash is only 81.5% full.All of this together seems to suggest that the original position is unlikely to be a mate in 20. Of course, once 7men DTM tablebases become readily available (again), we will get a final confirmation. For now I think it is best to go with the best mate score (and its PV) that we have.