Author: Dann Corbit
Date: 10:26:47 02/09/99
Go up one level in this thread
On February 09, 1999 at 13:00:43, Bruce Moreland wrote: [snip] >If there is more than one best move that leads to a desired game-theoretical >result, there is nothing inherent in chess that would suggest a choice between >them. > >The goal of chess is to maximize the game-theoretical outcome. If you can mate >you should mate, otherwise you should draw. It doesn't matter how long it takes >you, a mate in 45 is as good as a mate in 1. > >Most people would think that mating in a shorter time is important, but this is >a layer of veneer that humans have added to the game. There is elegance in a >short win, and there is practicality in it, since you want to score the point >without offending your opponent, boring yourself, or risking losing it due to >some external factor. I agree that a mate is a mate is a mate. If your program knows for sure it can win with a given sequence it is as good as any other -- even if it takes one hundred ply. >In a chess program, returning "MATE - ply" as a score will guarantee the >shortest mate the program can see, and this is an excellent heuristic if you are >winning the game, but it is less good if you are losing, as anyone who has had a >program throw all of its pieces away against you, before you even realized you >had a win, understands. > >I think it makes sense to score some moves as better than others that lead to >the same game-theoretical result, because you can talk about practical chances >versus specific categories of opponent. The choice of move is necessarily >subjective, but everyone can agree that a particular move would be most >effective against the known classes of opponent. Is there an effective way to measure this? >For instance, assuming no counterplay by your opponent, we can agree that it is >almost always a good idea to win a queen and mop up quickly, and we'd score such >a move as the best move in a test suite. But if you can also choose to win a >pawn and go into a technically difficult but provably won ending, there is no >difference, in a sense. > >We are really testing these programs against hypothetical opponents who have >approximately known characteristics. These are all very excellent points. My real concern is with some EPD problem that suggests something like bm Bxh3 A program eats that row and spits out: <goo>; bm Bxh3; ce -1291 pv Bxh3 Nd4... Now, the program chose the right move, but it really getting its butt kicked, according to its own opinion. I have had a lot of emails for the rockpile where I get "solutions" just like the above. If the move is supposed to be such a dandy, why are we down a queen and 3 pawns? If we really are that far in the hole it is time to resign. So there is no best move. Or if there is one, we *really* have no idea why it is good. In other words, I think a best move should: 1. Result in a draw or checkmate for the one making the move. 2. Result in a large positive change in ce for the one making the move. 3. If, even after the large positive change, you are still down 3 pawns or more, then there *is no best move* for that position. A lost position has no best move. If somehow, you are not really down 3, it still does not count unless your program can see it. In other words, if you accidentally fall on top of a solution, it should not count. In all cases, the program claiming to have found the best move should be able to justify the choice.
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.