Author: Thorsten Czub
Date: 22:26:43 10/22/00
Go up one level in this thread
On October 22, 2000 at 22:06:34, Bruce Moreland wrote: >With Chris we were talking about positional evaluations that seem to have been >essentially limitless. yes - limitless. -infinite to + infinite IF this brings us towards the target: mate the king. why 1 pawn ? why one piece ? why one queen ? when this LIMIT stops the program from calculating plans in the tree that direct the opponent into a mate?? I do not understand what you always mean with "exact" and "right" and "accurate" when we all don't know what the exact score is ? if you value a human beeing, in numbers, how can it be exact ?? everywhere where you have fuzzy stuff, where you talk about qualities or subjectice things, you have inaccurate stuff. so in chess IMO. Mchess is a good example. Marty tried very early to make a program that understands about chess. AIchess and later Mchess was a revolution at its time. Same for SuperConstellation. but those ideas were later overtaken by programs like frans and conchess. the reason was: the hardware in those times was very limited. we could name other efforts, Levy's MarkV-VI and Nitsche/Henne MephistoIII, Chrilly Donninger once tried with Nimzo, before he made a bean-counter out of it. how much is an attack worth ? IMO - if it brings you towards the target, this is more worth than all material. if you limit the value, because you dont want to risk something, the program will play limited chess, and not TRY to win by mate. it will try to play exact. i remember when chess-tiger 11.2 had to play gandalf. chess-tiger played very succesful at this time at the paderborn tournament, and was shortly before winning the tournament. but gandalf, as a new-paradigm program, instead of looking for material wins, THE WHOLE TIME IN THE GAME made a very clever attack. it never made this attack OPEN, i mean, not gandalfs search SAW those attacking moves, but gandalfs EVALUATION brought those moves into consideration. there WAS nothing to SEE for the search, because the "attacking" moves not really lead to a material-win. only to something you cannot measure into pawns and pieces. because it cannot be measured. [Event "7th IPCCC"] [Site "Paderborn"] [Date "1998.02.15"] [Round "7"] [White "ChessTiger 11.2"] [Black "Gandalf3"] [Result "0-1"] 1. Nf3 d5 2. g3 e6 3. d4 Nd7 4. Bg2 f5 5. O-O Bd6 6. b3 Qe7 7. c4 c6 8. a4 a5 9. e3 Ngf6 10. c5 Bc7 11. Re1 b6 12. cxb6 Bxb6 13. Ba3 c5 14. Bf1 O-O 15. Nc3 Bb7 16. Rc1 Rac8 17. Bb5 e5 18. Re2 Qe6 19. Nxe5 Nxe5 20. dxe5 Qxe5 21. Bb2 Qe6 22. Rec2 Ng4 23. Qe2 Qh6 24. h4 Rcd8 25. Nd1 d4 26. exd4 cxd4 27. Qe7 Be4 28. Re2 Qf6 29. Bc4+ Kh8 30. Qxf6 Rxf6 31. Rd2 Ne5 32. Be2 Rff8 33. Kh2 Bf3 34. Bb5 Bd5 35. Be2 f4 36. gxf4 Rxf4 37. Kg3 Rdf8 38. Ba3 d3 39. Bh5 Rf3+ 40. Kh2 R8f5 41. Rc8+ Bg8 42. Be8 Bxf2 43. Bd6 R3f4 44. Rxd3 Rxh4+ 45. Kg2 Nxd3 46. Nxf2 Nxf2 47. Bd7 Ne4 48. Bxf5 Nxd6 49. Rxg8+ Kxg8 50. Be6+ 0-1 look into the game. ChessTIger was a strong program at this time, it managed to nearly win the tournament. but in the above game, you can see gandalf PLANNING for a mate. Tiger made arround 100.000 NPS and came very deep. at those days it had no real king-safety term. it was forced to get all its strength out of the search. a bean counter :-))) although even at those times playing very nice chess. gandalf was, at those times, a very slow program. but it tries to mate by bringing its pieces into position. while i operated the game i could really feel how gandalf increased with any move the attack-potential. it had this plan in the evaluation that increased with any move that lead it near to the target. tiger saw nothing. and when it saw the stuff with the search: it was too late. Christophe has IMO learned his lessons over the years. he has learned that you can have a very strong program, but when it is not doing anything, when it does not evaluate for the target, and instead for the BEST MOVE instead for the BEST GAME plan, you will lose the game. I think one main reason why tiger12 soandso, participating in paderborn-championship where shredder made number one, was not getting close to the title was not that tiger played weak, but that it was not focused on mating the king, and instead of seeing the deepest into search. this is not the target ! it can be a way, but the way does not bring us to the target. also it creates a style that is very boring. this is what chris calls world war 1 chess. when troups run arround without having a plan. tactical. but not strategical. IMO gambit-tiger, although still a fast program, has now a very clever search, no doubt that christophe still trusts in search, but he has IMO understood that the program needs to know about the target. and that you don't have to limit the score when it comes the point of the game. > But what I think was interesting about that program is >that it achieved positions where its positional terms could be used. right. old-tiger versions lost this ability. therefore christophe lost tournaments. christophe learned that he has to do something, and that the values are not limited. because when it comes to mate, you cannot give EXACT scores. >It is hard to create something that has a natural "pre-sacrifice" style. but the saccing style comes automatically when the program understood that not material-winning is the goal but mating. >bruce look - i always liked programs playing good chess. for me it was never important how the programs DID it. therefore i had absolutely no problem to work with christophe on tiger. in the old days tiger was a plein fast searcher. it has knowledge, but an unbelievable good playing-style. over the versions tiger lost the aggressiveness. and suddenly the sword got blunt. this was the time when tiger participated in tournaments but made much draws, lost won games etc., not because the search was suddenly weaker, but because the evaluation-function did not lead anymore to a game-plan. we tried to push christophe: make it more aggressive christophe, where is the old agressivess that was a part of your program, bring back this old style that was succesful, ... and one day i guess he was himself so frustrated about his strong but colorless way the program played or lost, that he decided to work on the problem. gambit-tiger was thought to be a "not so strong" but interesting program. and suddenly we all recognize that the en passant made interesting-program has a new quality. IMO this new quality that is difficult to measure is not coming out of new search ideas, that maximize material wins. but it comes out of something chris whittington earlier preached. i do not say christophe read chris' stuff and followed chris way by will. he maybe never read anything from chris or never considered about it. but chris on the other hand brought his program out of the market, after many years of work he tried it out. and it was not too bad. and this changed something. because CSTal has no limit in evaluating king-attacks. it manages to attack ! Up into the fog, where it likes fishing in the mud. thats all. there is no real tactical win. its only a bigger chance in this mud to mate the king ! so : drive direct into this fog. if this increases the mating chances, why not . i am looking much forward how the others will react on christophes approach. i hope that computerchess changes. and that we all learn out of this. i dont think the others will adapt easily. cause how to do want to adapt on situations like Rc6 ? normal search based programs cannot see Rc6. they are blind. they see rc6 is not winning anything, so they don't register the move is doing something to the position. it is as if you don't recognize that your wife is betraying you with another man, if this happens over years, and you don't register that something is wrong in your relationship, maybe one day the wife is gone without you and your search having recognized it.
This page took 0.01 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.