I'm sure this has been brought up somewhere, but has there been any serious work done to try to refine engine personalities to the point where they can indicate who would win in a match of say, Capablanca vs. Fischer, Fischer vs. Kasparov, etc? I actually think this has probably not been possible until the last couple years, because it has only been in this time that computers have definitely become better than the best humans of all time. Before that, it was obviously impossible for a computer to predict all the moves of a stronger grandmaster.
My idea is to add a lot of well thought out parameters to a program ( max search depth, playing style, etc, along with perhaps even parameters of statistical distributions to quantify randomness in play), and then use optimization techniques (e.g. simulated annealing, evolutionary algorithms) to define the parameters to maximize the match rate with moves the players played in their games. If the match rate could somehow get high enough (>90%), it seems this could work.
The inspiration for this question are the Chessmaster personalities - but when one pits these against each other they don't really produce intuitive or consistent results it seems.
Using engine personalities to determine the best of all time
Moderators: hgm, Rebel, chrisw
-
- Posts: 10282
- Joined: Thu Mar 09, 2006 12:37 am
- Location: Tel-Aviv Israel
Re: Using engine personalities to determine the best of all
I doubt if it is possible to get match of >90% because I think that the algorithm of humans is simply too random for it.Peteshnick wrote:I'm sure this has been brought up somewhere, but has there been any serious work done to try to refine engine personalities to the point where they can indicate who would win in a match of say, Capablanca vs. Fischer, Fischer vs. Kasparov, etc? I actually think this has probably not been possible until the last couple years, because it has only been in this time that computers have definitely become better than the best humans of all time. Before that, it was obviously impossible for a computer to predict all the moves of a stronger grandmaster.
My idea is to add a lot of well thought out parameters to a program ( max search depth, playing style, etc, along with perhaps even parameters of statistical distributions to quantify randomness in play), and then use optimization techniques (e.g. simulated annealing, evolutionary algorithms) to define the parameters to maximize the match rate with moves the players played in their games. If the match rate could somehow get high enough (>90%), it seems this could work.
The inspiration for this question are the Chessmaster personalities - but when one pits these against each other they don't really produce intuitive or consistent results it seems.
You cannot predict even 10% of the moves of a random player for obvious reasons.
Top players are more deterministic than the random player but I do not think that their algorithm about choosing a move is deterministic enough to allow you to predict their moves with 90% accuracy even in the hypotethic case that you know their algorithm.
Uri
Re: Using engine personalities to determine the best of all
Thanks for the reply.
Yeah, there would be a lot of noise I suppose. I guess one accomplishment would be to create a personality that you could show is statistically better than all other personalities at predicting a particular player's moves. I'm not sure that's even been done yet. Again, I'm not sure what went into the Chessmaster personalities.
Yeah, there would be a lot of noise I suppose. I guess one accomplishment would be to create a personality that you could show is statistically better than all other personalities at predicting a particular player's moves. I'm not sure that's even been done yet. Again, I'm not sure what went into the Chessmaster personalities.
-
- Posts: 1357
- Joined: Wed Mar 08, 2006 10:15 pm
- Location: San Francisco, California
Re: Using engine personalities to determine the best of all
Most of the settings are not considered to be specific to that GM, but are rather relative to the settings for other GMs. Alekhine being heavier on the attack than most, Capablanca being more of a defender, etc.Peteshnick wrote:Again, I'm not sure what went into the Chessmaster personalities.
The goal is to make interesting play that is "in the style" of that GM, but not to really attempt to guess at what the GM would play in that move. One example is the Fine personality, which has a "Contempt for Draw" setting of -1, which means that he'll accept a draw even if he feels like he is ahead by a full pawn (+/- a few other minor factors). This is bordering on ludicrous, although Fine was known to accept draws in positions in which he clearly had the advantage.
A large portion of the "style of play" of the Chessmaster GM personalities comes from their opening book, which consists entirely of moves actually made by the GM (and nothing else). So, if you begin a game with an opening that the GM played a lot of published games in, you are very likely to come out of the book with something that is easier to believe is the GM playing.
jm
[/list]