Author: Dann Corbit
Date: 14:29:36 10/06/01
Go up one level in this thread
On October 06, 2001 at 16:36:47, Bruce Moreland wrote: >On October 06, 2001 at 01:54:57, Slater Wold wrote: > >>Well Bruce, with all due respect, you're kinda funny on these things. >> >>You seem to have a "threshold" for a winning move. Which I guess makes some >>sense, but it is very arguable. >> >>If Program-A looks at any given position and says, "I am winning by 6+ pawns, >>and here is my mainline" then it's winning. However if Program-B looks at the >>same exact position and says, "I am winning by 3+ pawns, and here is my >>mainline" it would seem by your standard, it's not winning. What if the >>mainlines are the same? What if Program-A rewards certain things differently >>than Program-B? >> >>It's kind of like you disagreeing that DJ7 solved Nolot #3. It did. It >>followed Baudot's mainline to a *tee*. And you still don't think it's solved >>because DJ7 says it's only winning by "x" amount. Which I guess is not above >>your threshold. It thinks the move is best, it has the mainline just as >>described by Baudot, how is that not solving the problem? >> >>I agree, programs should strive to find the *best* possible move. But a winning >>move is a winning move, right? >> >>It's sort of like drag racing; it doesn't matter if you win by an inch, or by a >>mile. A win is a win. > >It is easy to make a position where one side will score very high, but the game >is drawn. > >In the root position of WAC 100, white has connected passed pawns in an ending, >and this is going to score big. > >There are plenty of positions where connected passed pawns won't win, and yet >the score is still big. > >A big score isn't enough, sometimes, especially in tricky cases like this where >white has a high score from the root. I can show you positions from real games >where one program said +7 or so and drew. The particular case I'm thinking, one >side had a protected passed pawn on the seventh, *and* an extra bishop, against >a pawn, and couldn't win. So this is KBPP (both pawns healthy and neither was a >rook-pawn) vs KP, drawn via 50-move rule. > >If white has no plan in WAC 100 and just shuffles pieces for 50 moves, it will >show +3 or whatever for the entire time. It is hard to say that a program has >found a win, when you see that happen in a 12-ply search or whatever. > >+7 is more likely to be a win. I didn't follow the main line out, but stuff was >advancing and things were being captured. I think it's clear that both Be3 and >b6+ win, but it's not true that a program sees a win if it plays Be3 in >particular, since the move isn't committal. b6+ is an indication that a program >sees something, but Kb3 indicates nothing, by itself. The score seen by the program is also important. This is especially true at short time controls. For instance, a program might "choose the right move" but show a score of -3.0 pawns. Later, it changes its mind to the *wrong* move. After a very good long think of an hour or so, it finally sees the right move again, this time for the right reason, and the score reflects this wisdom. Sometimes, especially with a fairly new test set, a solution is found that is *better* than the given solution. There are test positions from books where the answer is absurdly wrong (e.g. no matter what you do, you will be checkmated if the opponent plays properly. Or a mate in 4 or 5 but the book's move wins a pawn and might eventually win the game but it is not at all certain.) >In the case of Nolot 3, Junior had +0.7 and the right line. That indicates a >preference for the Nxg5 line. The opinion about the Bxg5 line is probably not >that much different. > >Fine, Junior preferred the key move, but it didn't feel that it could force a >position that it believed to be significantly better. > >I'd give it credit for finding the key move, but it's not like it came back with >a score of +5. > >Mostly, we are trying to compare search functions with Nolot, not evaluation >functions. There's a difference between "finds the line" and "solves the >problem". > >Some of the Nolot position have been "finds the line" by various programs for >years. > >The first time I published a result for a Nolot position, which was in 1994, I >indicated that one of the "solutions" my program found was not convincing, for >this reason. I showed that the key could be found, but not necessarily that it >was understood. > >I acknowledge that in Nolot 3, the key can be found. If the score does not reflect the value of the move, then the program did not understand it. Imagine a ten year old (non-prodigy) playing a super-GM. In a thorny situation, he happens to make the ideal move. Now, his plan was short-sighted and not with the correct goal in mind, though the first few moves follow the same sequence. In the end, a GM will see the true value of the move but an inexperienced or less capable person won't. In the same way, chess programs can "solve" chess positions without having a clue why it is such a great move. The computer's assessment is the ce. If the ce is negative, it thinks it is losing. If the ce is small, it sees a drawish board. If the ce is huge, it sees a winning situation. I tend to agree that unless a program has a commensurate ce, it has found the right move for the wrong reasons.
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.