Author: blass uri
Date: 17:33:35 10/04/99
Go up one level in this thread
On October 04, 1999 at 18:46:11, Robert Hyatt wrote: >On October 04, 1999 at 14:17:06, blass uri wrote: > >>On October 04, 1999 at 11:52:43, Robert Hyatt wrote: >> >>>On October 04, 1999 at 10:30:40, blass uri wrote: >>> >>>>On October 04, 1999 at 09:41:30, Robert Hyatt wrote: >>>> >>>>>On October 04, 1999 at 04:26:17, blass uri wrote: >>>>> >>>>>>On October 03, 1999 at 23:44:31, Robert Hyatt wrote: >>>>>> >>>>>>>On October 03, 1999 at 23:17:29, Ricardo Gibert wrote: >>>>>>> >>>>>>>>[snip] >>>>>>>>> >>>>>>>>>my webster's defines 'sacrifice' as 'voluntarily giving up something of >>>>>>>>>value'. I have a hard time saying 'I will sacrifice a ten-dollar bill if >>>>>>>>>you will give me a 20 dollar bill in return...' >>>>>>>>> >>>>>>>>>:) >>>>>>>> >>>>>>>>Ok, you got me. I neglected to explicitly state I was refering to the _chess_ >>>>>>>>version of the term. >>>>>>>> >>>>>>> >>>>>>> >>>>>>>then here is a 3-move sequence. Sacrifice or combination? >>>>>>> >>>>>>>RxB, NxR, RxN. >>>>>>> >>>>>>>RxB obviously dumps a rook for a knight. or if you look to the end of the >>>>>>>combination it wins two pieces for a rook which is a significant advantage. >>>>>>> >>>>>>>Sacrifice or combination? >>>>>>> >>>>>>>How is that different from QxP+, RxQ, RxR#?? >>>>>>> >>>>>>>Dumping a queen for a pawn? Or winning the king? >>>>>>> >>>>>>> >>>>>>>>> >>>>>>>>>But I don't object to the term being used.. I just think that for a computer, >>>>>>>>>the concept 'sacrifice' is wrong. It is just a perfectly computable >>>>>>>>>combinational tree search... >>>>>>>> >>>>>>>>You can give up a bishop to obtain a draw by perpetual check and because you >>>>>>>>never get the material back, it is a called a sacrifice. I know it seems trivial >>>>>>>>and is not what people generally have in mind when they use the term >>>>>>>>"sacrifice", but I do believe it's use in such cases is fairly universal. >>>>>>> >>>>>>> >>>>>>>in the case of a computer, it isn't 'sacrificing'. It _sees_ that it can >>>>>>>draw or that it can win. IE it isn't giving up _anything_. A human might >>>>>>>toss a bishop 'thinking' (but not sure) than he can force a perpetual. But >>>>>>>a computer either 'proves' that it can force it, or it won't ever go for the >>>>>>>move in the first place. >>>>>> >>>>>>Not truth. >>>>>> >>>>>>Some programs use also selective search. >>>>>>I believe that Fritz evaluates positions based on some average between >>>>>>The evaluation based on selective search and the evaluation based on brute force >>>>>>search. >>>>>> >>>>>>If the selective search show perpetual check and the brute force does not see it >>>>>>then Fritz (in a bad position) might 'think' that he have chances to do a >>>>>>perpetual check without proving it and play for it. >>>>>> >>>>> >>>>> >>>>>However, that is a _bug_ and not a _sacrifice_ because the program searched and >>>>>found the perpetual. Even though it was wrong. But the _search_ said draw, and >>>>>the tree it searched 'proved' to the program that it was a draw. Unfortunately, >>>>>if this is the way Fritz searches (I don't believe it does this personally, >>>>>because it would be so horribly inefficient to do both kinds of search, that >>>>>Fritz would not be nearly as tactically strong as it is today) then the sac is >>>>>the result of a bug, not because of a computer 'speculating'... >>>> >>>>I know that Fritz is speculating and it is not a bug. >>> >>>Sorry, but I don't believe that. It either searches and 'sees' something >>>or it searches and 'doesn't see' something. I know of no algorithm that can >>>just 'guess' at a result, and fold this into the alpha/beta search along with >>>a normal deep null-move search, and then somehow combine those two different >>>results. >> >>The fact that you do not know does not prove that it does not exists. >> >>I also do not know if it exists and only guess because I had no explanation to >>some strange behaviour of the evaluation function that I saw(not often). >> >>It is possible that this strange behaviour is a bug > >I don't "know" that fritz doesn't do this either. That is why I clearly >wrote "I don't believe...." which is quite different from saying "It >absolutely does not..." > >>> >>>IE CSTal doesn't 'speculate' in that form... it just has large positional >>>scores it tosses into the mix when it sees certain things going on on the >>>board, such as the king too exposed or whatever. And deep/fast searchers >>>generally are able to spot the fatal flaw in such speculation and pounce on >>>it with both feet. I have _never_ seen Fritz behave in this manner because if >>>it did, it would get crushed by programs that didn't behave like that... >> >>It is possible that usually the selective search does not lead to mate so the >>number of the selective search does not have big influence on the evaluation >>function. >> >>If you use 0.08*selective search score+0.92*brute force search score >>then you will see problems only when the brute force search score leads to mate. >> > > >But who does _both_ and merges the scores together? That is the part that makes >no sense from a tree-searching point of view. Because the two search spaces >overlap a _lot_ and it is a lot of wasted effort... > > > >>You also can use a different formula that is not linear. >>> >>> >>> >>> >>>> >>>>In a case the selective search show draw by perpetual check and the brute force >>>>search does not see it the evaluation is probably going not to be 0.00 but >>>>something between 0.00 and the evaluation of the brute force search. >>>> >>> >>> >>>Again, I don't believe that fritz is doing _two_ searches, one selective and >>>one non-selective. It might be adding some selectiveness on to the end of the >>>normal search, as that has been done as far back as the original greenblatt >>>program... However, Thorsten has reported seeing lots of 0.00 scores when they >>>are simply wrong. I have played fritz on the servers and had the opponent say >>>"I am seeing a draw" while Crafty was seeing +3.00, and in many cases, the 0.00 >>>was wrong... >>> >>> >>> >>> >>>>I do not remember cases of speculating perpetual but I remember cases of >>>>speculating when it saw a win for itself in some selective lines and decided >>>>to do a sacrifice(sometimes it may be right sacrifice and it also may be >>>>a wrong sacrifice). >>>> >>>>I guess that it does an everage between selective search and brute force >>>>because I saw some evaluations that I can explain only by this theory. >>>> >>>>I remember a case when the evaluation started to go down slowly from a big >>>>advantage for white 7-8 pawns towards no advantage and >>>>The sequence of evaluations was arithmetic sequence. >>> >>> >>>that happens. It simply means that the evaluation is grossly faulty, or that >>>the search is faulty... we all have that problem from time to time... I have >>>lost +5 games on ICC and won -5 games, against computer opponents.. >> >>I remember an evaluation that cannot be explained by the position >> >>It started from +8 or +5 (I am not sure about the exact number and got down by >>0.31 every iteration(again I am not sure about the exact number) >> >>evaluation like +3 pawns could not be explained by a logical evaluation >>because if you see that you win the queen it should be at least +8 pawns and if >>you do not see it because of null move problems the evaluation should be close >>to 0. >> >>Uri > > >Not necessarily... you can hold off such losses at times by giving up a bit of >positional compensation.. a sort of 'positional horizon effect'. But each >iteration takes you a ply deeper and you have to give up more to hold the >loss beyond the horizon... and down, down, down goes the eval... The relevant position is: 7k/4K2p/7P/3p4/8/4Q3/1q6/8 w - - 0 1 The first evaluation above 0 of Fritz5 is +5.16 pawns for white I do not believe that it can be explained by positional compensation. If it is because of a bug then Fritz3 and Fritz4 have the same bug(I do not know if Fritz5.32 shows similiar behaviour) Uri
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.