Author: Roger D Davis
Date: 18:29:25 08/11/99
Go up one level in this thread
I mentioned doing something like this a while back, and actually mentioned using genetic algorithms to breed styles that are closer to the human player of interest. I'm sure this same idea has been entertained many times by many people. On August 11, 1999 at 15:10:04, Christopher R. Dorr wrote: >This is an idea that has interested me for a while, and I was interested in what >others thought of it, and of how it might be implemented. > >CM6K (among others) has various 'personalities' that are supposed to play like >various players. Generally, they don't seem to. The changes in playing >parameters are often too coarse, and often seem to be ideas that someone had >about what a player values, rather than being based upon something objective. > >What I'd like to do would be to take the entire corpus of a players games (most >great GM's didn't play more than a few thousand recorded games at most), and >analyze them with a very strong program. Have the program determine what it >would play in every given position that the player had to face, and score the >correlation. A score of 1.0 would indicate that the computer chose every move >the player did, a score of .5 would mean that it chose 1/2 the moves the player >did, and 0 would mean that it chose none of the moves that the player did. > I don't think you need to compute a correlation. Correlations usually deal with continuous data anyway. You wouldn't want to deal with all the moves in all the games a particular player played, either, since some of those moves will be forced, and others will be part of opening variations popular at the time the player lived. You just need a way of selecting moves that reveal individual differences between playing styles, that is, a way of detecting nonforced moves outside the opening, and then turning these into the equivalent of a test suite, and then maximizing the score on this suite. That's where the genetic algorithms come in. If you wanted to get really sophisticated, you could select a pool of players you wanted to emulate, and then have master's level players rate each pair of players in the pool on a scale from 1 to 10 in terms of similarity. You could then perform a multidimensional scaling procedure to extract the underlying dimensions of "playing style", whatever those are. Your genetic algorithms would then need to be constructed not only to make sure that Tal is as much like Tal as possible, but also to make sure that Tal is as different from Morphy as the judges believe. Your methodology thus has both a convergent and discriminant aspect, so that each player is "fixed" both relative to himself and others in the multidimensional space which represent playing styles. >Then, we modify a or some variables that the computer uses to evaluate the >position, i.e. center control bonuses, open file bonuses, pawn structure >penalties, piece/pawn values, depths and extensions, and do it again, possibly >using a genetic algorythm approach. Eventually, we will have better and worse >approximations of the player's style. While we may never have an exact >electronic Petrosian, the set of variables that scores a .90 correlation is a >better representative of his style than one that scored .75 (over the body of >his professional games). This would of course take many iterations, and a ton of >computing time to evaluate 3000 games at a reasonable time control, but I feel >that it would generate some useful information. > >Once we have some good 'proxies' of great players, we can evaluate some things >about playing strengh. One of my favorites was the old "Morphy was a 'better' >player than Tal" debate, that could never be answered because of a lack of a >common frame of reference. > >If my "95% Morphy" plays my "95% Tal", then we can get some objective judgement >about the validity of their respective approaches and philosophies of chess. > >At some point, this could have some instructive value as well. When I was >actively teaching chess, I often had an intuitional feel about where a student >needed to improve, but sometimes, I couldn't get a handle on their issues. If I >could 'pollute' my model to play 90% like my 1600 student, I could determine >such things as whether it was consistency, an undervaluation of positional >considerations, or an overvaluation of the attack relative to material, and >hence give some objective targets to attack. > >I'd be interested to hear whether anyone else feels that this is feasible, or >just a bad idea, or actually has some promise. > >Thanks, > >Chris Roger
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.