@Toscani said in #5:
> Thanks...
> I guess the best fen's for your site would come from a quiescence search.
> www.chessprogramming.org/Quiescence_Search
@SayanManu When you say evaluation on the site, could you specify the command line syntax used, so we could make sure that it is either a searched score or a static leaf evaluation of the FEN? ooops i saw #2 by Toscani.. same curiosity.
I took the habit of using score for at current position FEN that is being analysis by SF, say as on lichess, the output is a whole sub-tree searched numerical "measure" resulting from optimizing (min-max coarsely) all the tips of the branch of that partial tree.
So the comment by
@Toscani is based on understanding or assumption, that you are using the word evaluation, for the static evaluations. There is a version of command line syntax for SF that promise output being such a static evaluation of the input FEN. It use to give a very reduced view of the full matrix of output that the static evaluation used internal in root searches.
And example of some past version (unknown) of SF classical eval along with some version of NNue eval and their combination can be found on a site writting in javascript, not html, (so hard to link to the matrix directly you have to go there and find it).
That would be the evaluation function in the parlance I would choose to respect in constrast with a score given an input FEN which would result from many variations having been explorer to variable deep depths, result that come from the very last positions of those variation, during the single root (input position) search.
The post i quoted, is about such searches extending from input depth parameters, as only the leaves that satisfy the quiescence search criteria get the chance to get that static direct explicit function of the input FEN. the FEN at the end of extension search.
The FEN where i understand there are no more possibly capture move.. (am i wrong). so that the static evaluation does not happen in the middle of a bar fight. or the legal possiblity of it. As humans we do not choose to grab any material that can be grabbed. we look at all the interactions on the board.. we can go for a while before consuming threats. all along the engine would not try the static eval there, during a root position search (i use that phrase to distinguish from play which is also about many positinos along a game)..
I think the idea quote test above. is that fishtest may have been always optimizing parameterrs over only such quiescence abiding positions. other position would not be subject to selective ELO pressure in the fishtest devlopment pools many games. as never affecting the score of those games The score which among which the engine find its best candidate "move" given the sub tree searched, from input root.
lots of repeat maybe above.. but i think if your site were to explain how you find the numerical values, we would be able to understand the meaning of "evaluation" of a position in the title. or i missed a post above confirming it is the comman line static evalaution.
it might be interesting in that case, to compare with the web guide.. Stockfish evaluatino guide or something (in dev forums, for ages, the static evaluation was often referred to as the evaluation function, but in lichess and chess community I have noticed that there is a fog zone between score and evaluation. and no awareness of the explanation above trying to explain the difference. as we shoud be curious about that... instead of dismissing the amplitude of the score because we lost our material count rule of thumb with it. Now that SF16 is coming, that maximum score amplitude evololution, will not be visible anymore.
So the -10, and 10 i think i saw, might indicate i just babble in vain, w.r.t. to op.. sorry then.
With regard to the command line EVAL syntax, it is not clear that it is the one used within the single root subtree searched (i.e.. normal use of SF as oracle, e.g. in lichess).
one way to probe possible discrepancies would be to use the web guide live code.. SF classical is supposed to have been frozon for a while, and only NNue undergoing evoluatino in development across versions.
that leaves the question of which NNue is there. ok .. and also are we getting the parameters (i call them docking) of all the places in the code before leave eval gets propagated back in the optmization based on min-max recursively backward in the explored tree (whether the tree is forgotten to make space or not).
so we should trust that the EVAL command, is not further modified when receiving a main PV score compared to the possibly forgottent leaf eval value (possibly the value is not forgotten but the position that it came from, yes. why lichess does not give us the tail of PV. (pros and cons to that, it makes it look like there is no problem, unless one is in the knowing circles, not a conspiration, inertia, and lack of curiosity from us. credulity in face of high engine ELO.
I suggest such tests be shared if anyone has the energy to test the web guide.. you have to enter each FEN manually.
Also i wonder where the EVAL output could have the labels of the table explained. any links?