Author: martin fierz
Date: 06:18:04 01/08/04
Go up one level in this thread
On January 08, 2004 at 08:36:49, Ed Trice wrote: >On January 08, 2004 at 04:27:01, martin fierz wrote: > >> >>i don't know whether that is correct. IIRC your experiments were done with >>wyllie (a weak program) and with an old version of kingsrow (which was >>absolutely no good at winning these kind of positions - i know this because i >>discussed this with the author of kingsrow, ed gilbert - of course, you couldn't >>know this!). >Ed Gilbert felt his program could win, and it did not. >Roberto felt his program could win, and it did not. > >This was an important result! For the first time, it was discovered that a >winning position in the game of checkers was difficult enough to be drawn in >practice. in fact, you demonstrated this for the first time. schaeffer already predicted this long before: it is already suggested in "one jump ahead" that this can happen. only somebody (=you & gil!) first had to compute perfect play databases of course, to really demonstrate it. >>after some discussion with ed gilbert, his program is now much better at >>converting wins. > >Yes, AFTER this was demonstrated to him. How was I supposed to play the program >version that he improved AFTER this experiment while we were conducting the >experiment? of course you couldn't :-) he always knew that this might happen. what he didn't realize at the time was that you can do much better than he did with his old kingsrow version. what made him improve his winning endgame heuristics were a number of games i sent him between cake and kingsrow, starting in tough winning positions (not that he couldn't win against you - he sort of believed that that was natural). cake won every single one on the strong side, and defended several on the weak side. this result made him think :-) of course, kingsrow was not defending perfectly like WCC could; so this is no proof at all that WLD-db programs can win complex winning positions. but it is certainly a demonstration that you can have better heuristics for such endgames than what kingsrow had at the time. >Well the entire solution to this 253 ply win has been published now. I dont >understand how you can diminish a result shown against two grandmaster strength >checker programs. Yes, Wyllie is the weakest of the lot, but it still would >easily exceed the minimum requirements for grandmaster status. this is not about diminishing your result. you said that "programs cannot win against the ppl database". you demonstrated this against two strong programs, that is quite true. but this still leaves the interesting question open whether it is possible to use a WLD database + good heuristics to win even the complex endings. i believe that very much can be done in this respect. however, i also believe that my heuristics, although better than those of kingsrow, are not good enough to win certain positions - we should try one day! finally, you know just as well as i do that this "winning a winning endgame" has nothing to do with the grandmaster status of a program or not - today, any program with an 8pc database and a reasonable evaluation function is grandmaster level, with a basic alpha-beta search and no search tricks at all. checkers is too easy :-) what i want to say is that you can have ZERO heuristics for winning such endgames and still be grandmaster level easily. this is all about whether you ever thought up some intelligent heuristics to win such endgames or not. >The longest win for 3x3 = 167 plies > >bk: 25 >bc: 5,9 >wk: 18 >wc: 17,30 > >black to move and win. I think it just plays into a 3x2 in a few plies, but it >is one of the longest 3x2 wins. You and I already played a 3x2 test in 2001, and >we learned that these are able to be handled by contemporary programs. i will try this one. we only learned that ONE position was handled properly - not more. it is quite possible that cake botches this one :-) cheers martin
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.