Author: Bob Durrett
Date: 08:15:34 11/20/03
Go up one level in this thread
One could consider starting without an opening book and then let the chess-playing program develop it's own opening book based on what it learned in it's past and ongoing games. [Essentially, the program would develop it's own opening repertoire.] This could be done with or without human assistance. Also, actual learning might occur during play of the games or occur later in special post-mortem learning sessions. Human involvement in the process would be relatively easy if the "post-mortem learning" option were chosen. How does a human improve his/her opening repertoire? Essentially, there are two ways: (1) by individual creative innovation and (2) by relying on outside sources &/or assistance. The same could be done by a chess-playing program. The following assumes that the "post-mortem learning" option is chosen and that one wishes to have the chess-playing program rely solely on it's own individual creative innovation for finding new opening moves. After a game, a wise human does a post-mortem analysis and tries to figure out where he/she "went wrong." In our present-day "dumb" [: )] engines this might have to be a matter of identifying the moves which caused the position evaluation scores to go bad. [Hopefully, future engines will be able to recognize and process long-term considerations as well.] Suppose, as an example, that the program playing white tentatively "decided" that it's first weak move was its fifth move. In the post-mortem learning session it would then do a "deep think" first to verify it's tentative conclusion and then to find the very best alternative fifth move. It would then store the original white fifth move(s) with performance history for the move(s), and also store the new alternative fifth move in it's own "private" opening book and play the new move on it's next opportunity. During the next post-mortem learning session, the chess-playing program would compare how well it did after the initial and new fifth moves. It would then do more "thinking" to decide whether or not (a) A new fifth move was needed, (b) the problem was before the fifth move, (c) improvements were needed in moves following the fifth move. The interested programmer must now be thinking about the underlying logic and candidate algorithms needed to implement this idea. As the book grew, stored history information and it's processing might become somewhat complex. Admittedly, the programming task would be challenging but there seems no fundamental impediment to implementing this idea into software. It would be most interesting to see the openings the software came up with. Naturally, learning would be best/faster if the opponents played very strong chess, at least in the opening phase. ["Garbage in, garbage out"] A top GM might be better able to "teach" the silicon "student," although there is no reason why the "teacher" could not be "silicon" too. [Maybe the human programmers could be replaced with "silicon" programmers too. : ) ] Bob D. _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ On November 20, 2003 at 09:54:12, Bob Durrett wrote: >On November 20, 2003 at 09:27:12, Mathieu Pagé wrote: > >>On November 20, 2003 at 03:34:54, swaminathan natarajan wrote: >> >>>Hi >>> >>>the main problem is that engine-engine matches without opening book plays same >>>moves for infinite games when given same time control >>> >>>i have kept 6 games series >>>all the games played shows the same result >>>match ends in a 3-3 tie........same moves till endings and insuffiecient >>>material in the position >>> >>>i hope if the book is selected for the engine,the games will not be the same and >>>the result may change? >> >>Yes, you are right, except if both engines take the same opening line again and >>again. >> >>Some sort of opening learning will totaly avoid this behavior since the engines >>will try to find a line that will give them something better than a draw. >> >>Mathieu > >The idea of using "some sort of opening learning" to cause chess-playing >programs to explore new openings in engine-engine matches "without an opening >book" is intriguing because it would be very interesting to see what the >"engine" might come up with. One reservation, however, is that the choices may >not be truly without human influence. It would depend on how the idea is >implemented in software. If human programmers were, in effect, making the >decisions then the findings would be tainted by the scourge of "the human >tendency to err." > >Perhaps it would be enlightening to review how opening learning is accomplished >[implemented in software] in current-day chess programs. My impression was that >opening learning is accomplished by adjusting the probability settings in an >opening book. It doesn't have to be that way. A programmer could find another >way. > >Hopefully the "real" chess-playing program programmers will comment on how they >might implement Mathieu's idea. > >Bob D.
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.