Maybe that contempt setting needs to be AI auto adjusted depending on the material advantage, so that it avoids the draws when the material end games show it is unfavorable to exchange because the end game table says with material similar to what is there will cause a draw or unfavorable win ratios. An engine should not have to wait until the end game to reference the end game tables to make a call on the piece to move. It should start using the odds of winning, if the game ends with that material.
It should not be contempt to use a prefixed number and not look at the end game table for similarities to the present position. Sometimes pieces are locked down in mobility and cannot directly affect the mobility of the other pieces, so the end game tables could be used beyond 7 pieces. The tables should be used as a guide to reach a proven end game. The end game tables are in a way solid positions that should not be ignored until they reach that level of pieces. The end game tables should be able to influence what to play before they reach that 7 men.
https://www.chessprogramming.org/Contempt_Factor#
It looks as though you as going deep. I am slower than you. (also a bit dispersed a the moment, as often).. but i will keep my goals minimal.
let me know if you find out that the deep runaways search of the begginning, the ones you put in post have to be revised.
otherwise. my priority is to find back jomega's example with verbose SF in details (pruning error, caught in time by killer move list), and look at your position example.. see if there could be same mechanisms at play, without the killer move list catching the error in time.
killer move list is dependent on the current position under exploration in the subtree expansion to still have candidate moves in it. it is fleeting.
Hypothesis under my current consideration.
> There could be many ways the first example repair could not have happened.
@Toscani please contact me with links to your posts you would ask need for my point of view. I am aware of the web guide site, but I have given up about figuring out a critical view of what it outputs. that is also a long story. but some of the chess theory feature present are worth noticing.. just not their weight with respect to material counting components.
keeping that long story about parameter optiinzation of the past for later. may need iterations. Since i am also curious about chess theory as it is accepted in SF, I would like to keep talking about the human understanding of them, it might give me some clues about the interaction. but my non-engine recent interest was focussing on more curated sets of thems such as the puzzle system here on lichess with restricted tactical scope where consensus about themes meaning something is automatized and roughly not disputed (but possibly for semantics preferences, all of which can be adjusted later i think). For the other positional themes or chess theory complement. my current convinced objective (hard to achieve, for me) is getting through PCC table, which is way smaller than the web guide set of features that are uncurated (for interactions,intersections or double couting etc...). PCC main effort according to me is to curate a sufficiently explaining for target chess player learning level, so that it could be use as a set of analytical component for evaluation of positions beyond material counting signals.
Lots of creative guesses about how engines work in this thread. I'm sure this already got mentioned but I'll recommend reading https://www.chessprogramming.org/Main_Page to replace some of those guesses with accurate information.
And just responding to that last guess, no, contempt has nothing to do with speed, it is to avoid the case where the engine is behind against a weaker player(a human for instance) and would normally play for a draw, since a draw is 0 and the current score is worse than that. But with contempt it will treat draws as being worse than 0, and will avoid them unless the position gets even worse than the contempt value. Contempt is not useful for a human chess analytical tool, it is used to win games against a flawed opponent.
@Toscani said in #42:
> It should not be contempt to use a prefixed number and
You have totally misunderstood contempt. AS @ProgrammerAngrim told is a small trick to play for win when it safe to assume opponent is lot weaker.
Without that grandmaster can accidentally get a draw agains top engine. lets that top variation gives -1 centipawn and there exist forces draw. in such cases all engines take the draw UNLESS there is contempt set which means it will a move that can create winning chances at cost slightly worse position.
Before one can propose sensical improvements to current search procedures developed for about 70 years is a must to study the current state of art. No matter how laborious it is. otherwise just use the darn things as they very complex to understand
all heurestics have been introduced in scienfistic papers. Stockfish does have list on search features listed. programmin wiki and wikipedia both have links to original papers and sometimes to some further ramifications of them. Like singular extension as implemented on -89 is not used by anyone as is but it is/was used by SF and there is plenty of articles about it
https://www.chessprogramming.org/Singular_Extensions
as there is no all combinen article about SF it is not really needed. they are doing alpha-beta(or negamax) search of the tree withlisted heurestics. then if you are interested on given heurestic you will find links to being properly formulated easy enoug. Almost of them are result on open science work - not all. and if you like more holistic description I am not aware any modern presentation. https://www.amazon.com/Computer-Gamesmanship-David-N-Levy/dp/4871878058 only one know of but that if from mid eighties so it covers all truly essential parts but misses quite bit late developments.
Dave levy was the IM that made bet thta computer cannot win so-so timeframe and finally lost to Deep Thought
This topic has been archived and can no longer be replied to.