Author: Ed Schröder
Date: 00:38:15 11/19/99
Go up one level in this thread
On November 18, 1999 at 18:35:02, Thorsten Czub wrote: >What i wanted to say is: > >We have programs like > >Mchess >Hiarcs >CSTal >Tiger. > >They all evaluate positions in a big range of scores. >Higher positional or speculative scores than other programs. > >But - >Most often when you see Mchess/Hiarcs/Cstal evaluate the positions >and also from iteration to iteration and often within an interation >you see the score jump and fail-high, fail-low, changes etc. > >Tiger is different. >It has the same high range to evaluate, but from first second of computation >to the play out the move the jumps and the score is not changing much. > >the score increases or decreases very close to the end-result. >it gets a fail-high or a fail-low or whatever window, and the next >big main-line is only 1/100 of a pawn better or weaker. >it all goes more smooth. more like a wave, not like a quantum leap. >not jumping but flowing like a river. > >anmon has this kind of evaluation too, but still different to tiger. > >IMO what happens when cstal score jumps or hiarcs score jumps or mchess >score jumps is that the evaluation get more precise, >but the search gets unprecise or say better: ineffective in these moments. > >Not with tiger. >Tigers score get more and more precise, but in the same moment the >search gets NOT corrupted by score-jumps. >It can't because it does not JUMP-. > >So the search is not inefficient although the score changes. > >This seems to be optimal IMO. >it leads to the fact that tiger goes deeper, and wins. >If another program gets deeper than tiger, it does not matter because >the tree of the other program is not accurate. the lines the other >program that comes deeper than tiger construct are not reality, but dreaming. > >so the strength of tiger comes IMO from a combination of >a) very efficient search >b) very clever handling of HOW to use the knowledge the way >that it differenciates different branches, but does not corrupt >the search process. A kind of clever selective mechanism. > >Therefore tiger can play positional AND can be outsearch >(e.g. Fritz6). > >I have played and seen tiger kill fritz6 in 40/120 games >on my machines 11.5-6.5 and i got 50% against fritz5.32 >(fritz5.32 searches deeper than fritz6). >Tiger is something special. >it will be a class better than the old programs. >you will see. >i am sure all other programs will have to copy christophes ideas >sooner or later, otherwise they will not be able to stop him. > >rebel can much profit from the joint-forces of christophe and ed. >tiger gets all the cleverness of old fox ed and his tiny tricks >that change the program massively, and ed gets many ideas from christophe >how to make search and comp-comp more efficient. > >What worries me the most in the moment is the opening book. >Jeroen has to get that human-chess opening-theory (game vs. nimzo, >game vs. diep, game vs. quest) has nothing to do in a tournament >opening book for comp-comp tournaments. > >these modern and new and long lines do STOP the engines from >showing their possibilities. > >I hope jeroen would make a new book with NO human-opening theory within. >or almost low percentage. >and look for forced but balanced positions, and not >kings-indian main-stream stuff and other openings like the shit big >theory line against nimzo. You are quite right on the book issue. Jeroen has a double task now as he has to maintain 2 books from now on, one for Rebel (just normal and sharp opening theory) and one for Tiger a modified version that is more solid. The Rebel book plays all human theory as sharp as possible including all kind of sacrifices. I often see Rebel go out of book with a -1.50 score having a good compensation of course. This is fine for playing humans but not (always) such a good idea playing against other computers. It's Jeroens intention to avoid these kind of lines in the Tiger book as we noticed very soon that Tiger does a lot better in comp-comp if more solid opening lines are played. In the final version of Rebel-Tiger you already will have a more solid Rebel book. However it will take Jeroen several years to finish this job. Ed
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.