lichess.org
Donate

Forced checkmate sequence can increase in length as analysis goes deeper

If only we did not only get scores at root or first PV leftmost node as feedback. What if we could see the whole PV used for that score.

I am also puzzled by notions of forced anything. And why should a forced sequence necessarily be the shortest. But here it is the forced notion itself. And I am putting a draft post in my musing.. instead of brainstorming here..

I think, if this is reproduced, as last 2 posts seem to confirm (I should do it myself, but seen a bunch of deep searches before and would rather rely on others curiosity by now), that we might need better understanding of what and how exhaustive search engines go from a legal tree search, with an a-priori legal mate, to ensuring the mate is actually the best alternating min-max sequence when the whole searched tree is considered.. I would like to remind all of us, that engines work with partial trees. not complete trees.

Rambling elsewhere. got tired (typing and thinking endurance). The above is summary, post removing. might try to explain while asking question later. I might also be not understanding what engines are doing there well. But who can?
@schlawg said in #18:
> If someone will provide it, I can look into it.

I agree with u, I don't think it does. But it MIGHT be possible. I'll try to provide u with a FEN for the position
#31 confused me a lot. I understood almost nothing from it. It’s out-of-context, maybe?
I think we can agree that this number SHOULD NOT be revised upward in the same search at greater depth, otherwise what justification can exist for reporting it at all?

I have not seen it increase when using a fresh single pv search where hashfull values stay under 1000. By fresh, I mean directly after launching the executable. I do observe it with multipv > 1, at hashfull 1000, or on subsequent searches without clearing hash.

Maybe the stockfish devs can explain or isolate this further.
@schlawg said in #34:
> I have not seen it increase when using a fresh single pv search where hashfull values stay under 1000.

It doesn't seem to have anything to do with Multi PV or old data in the hash table. This is a newly started Stockfish 16 with no Multi PV. It happend in my first try. At depth 46 the distance to checkmate increased. The hash table was not more than 30% full at this time.

38/42 00:02 -35,89 bxa4
39/50 00:02 -96,05 bxa4
40/57 00:03 -96,05 bxa4
41/48 00:05 -M40 bxa4
42/44 00:05 -M40 bxa4
43/45 00:05 -M40 bxa4
44/46 00:05 -M40 bxa4
45/47 00:05 -M40 bxa4
46/49 00:05 -M41 bxa4
47/48 00:05 -M40 bxa4
48/55 00:06 -M39 bxa4
49/58 00:06 -M32 bxa4
Mind providing everything you sent after starting stockfish?
My guess is that Stockfish is not guaranteed to perform better at higher search depth. I can undertand that.

I don't believe that an announced forced mate can be revoked becuase it's incorrect. That would be a bug.
@schlawg said in #36:
> Mind providing everything you sent after starting stockfish?
Hm, yes, but it would be a lot. Should I post it in a private message?
I mean everything you typed, not everything stockfish generated.
<Comment deleted by user>

This topic has been archived and can no longer be replied to.