Author: Robert Hyatt
Date: 13:31:42 08/21/98
Go up one level in this thread
On August 21, 1998 at 10:54:04, blass uri wrote: > >On August 21, 1998 at 10:33:44, Robert Hyatt wrote: > >>On August 21, 1998 at 09:56:10, blass uri wrote: >> >>> >>>On August 21, 1998 at 08:59:25, Robert Hyatt wrote: >>> >>>>On August 21, 1998 at 07:58:21, Robert Henry Durrett wrote: >>>> >>>>>There have been quite a few threads this month dealing with test positions which >>>>>test verious aspects of the software. >>>>> >>>>>But there is one question about the software which nobody seems to want to >>>>>mention: Why do the chess engines have trouble with the initial position where >>>>>no pieces have been moved yet? [Position at the start of all chess games] >>>>> >>>>>If the programs can be "tweaked" so that they play good openings, then all of >>>>>the opening books will become unnecessary! >>>>> >>>>>Why not work on the "initial position" and get the programs to find the best >>>>>move for that position? >>>> >>>>The primary reason this fails is you will get killed doing it. As a test, >>>>try Crafty with either book=off, or book random 0, which means it will play >>>>the exact same opening moves if you don't vary the time control. If you lose, >>>>you can pick *any* move of yours you made, and replay the game to that point >>>>and vary, knowing the program will walk down the same line again, since they >>>>are "deterministic". You will eventually find a move that wins, and then it >>>>is all over when word gets out, because everyone will play that move. >>>You can use some learning and this is impossible to win by the same game. >>> >>>for example after a game you lost you can have a bad opinion against the moves >>>you played(a panelty of 0.1 pawn against these moves is enough to change the >>>first move and you cannot repeat the same game). >>>another idea is to use singular extensions to the moves in the games that you >>>lost. >>>It is interesting to check what is the value of the opening book in ELO in the >>>ICC if you use learning so you do not lose the same game twice. >>> >> >>against *what* move? If a program doesn't have a book, and plays totally >>on its own, as suggested above, then *what* move do you assume is bad >>when your opponent plays the same first move again? Do you assume your >>first move was the loser? The second? the 50th? > >I do not know so I decide to have an opinion against all the moves I played >and I substract 0.1 pawn from the evaluation of the moves I played in the game I >lost >It can change the first move from 1.e4 to 1.d4 but not to 1.a4 and after some >games I lose with white I play 1.e4 again because even with the -0.1 pawn >it is better than alternatives without -0.1 pawns >If the opponent play the same then I change the first move it is reasonable to >change. > >Uri And you'd make the same mistake I made in the first verion of Crafty's learning, too. :) IE it learned that *no* book moves were playable, because it only learned bad things, and there is *no* move that someone can't beat you with... So it gets *very* difficult to pull this off. That's why the current version also learns "positive" things... Without the book, your idea would also cause problems, because the move at move 20 might have been brilliant, while the move at move 30 was an outright lemon. You penalize the brilliant move as well as the lemon, and play *worse* moves everywhere in the game given another chance... it would seem to me that this would fail even worse than not learning, because this is "random" learning with no idea of what was good or bad..
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.