just a winboard chess engine question...

[Cheszekiah]

Geschrieben von: / Posted by: Cheszekiah at 30 May 2000 02:15:05:
This may sound like a stupid question to most but I'll ask it anyway. I've played Crafty against other chess engines (Comet, Phalanx, AnMon, etc.) under the same time constrictions and it "seems" to me like whichever engine losses, learns to adapt and chooses a different path (same opening but different line somewhere in the middle game) from the previous game. If this is the case, then isn't it fair to say that these engines play better the more you play them? With that assumed, whenever I match any one of these engines against Chessmaster (slightly different time controls - say 1/1 for the winboard engines, and 1/3 or 1/5 with Chessmaster; I do this in hopes of improving the play of the engines by beating them repeatedly), I ALWAYS get the same result - a Chessmaster win or draw (CM has NEVER lost in the 60 or so matches that I've simulated; and yes, the engines do choose a different line somewhere in the middle of the game but they still end up losing, and often times, losing in embarrassing fashion - down by 3+ pawns is rather embarrassing to me). And when I do play the engines against CM with the same time controls, the win-loss ratio is usually 6:4 or 7:3 in favor of CM. So I wonder if these engines adapt and play a meaner game of chess with every loss?

[Pete Galati]

Geschrieben von: / Posted by: Pete Galati at 30 May 2000 03:50:16:
Als Antwort auf: / As an answer to: just a winboard chess engine question... geschrieben von: / posted by: Cheszekiah at 30 May 2000 02:15:05:

This may sound like a stupid question to most but I'll ask it anyway. I've played Crafty against other chess engines (Comet, Phalanx, AnMon, etc.) under the same time constrictions and it "seems" to me like whichever engine losses, learns to adapt and chooses a different path (same opening but different line somewhere in the middle game) from the previous game. If this is the case, then isn't it fair to say that these engines play better the more you play them? With that assumed, whenever I match any one of these engines against Chessmaster (slightly different time controls - say 1/1 for the winboard engines, and 1/3 or 1/5 with Chessmaster; I do this in hopes of improving the play of the engines by beating them repeatedly), I ALWAYS get the same result - a Chessmaster win or draw (CM has NEVER lost in the 60 or so matches that I've simulated; and yes, the engines do choose a different line somewhere in the middle of the game but they still end up losing, and often times, losing in embarrassing fashion - down by 3+ pawns is rather embarrassing to me). And when I do play the engines against CM with the same time controls, the win-loss ratio is usually 6:4 or 7:3 in favor of CM. So I wonder if these engines adapt and play a meaner game of chess with every loss?

Not a stupid question.
I think all 4 of the Winboard engines that you mention have the ability to learn. That's a very time consuming way to try to improve the strength of the programs by manually passing moves back & forth between Winboard & CM. You would maybe gain more by playing the Winboard engines againsts themselves in matches, and going and reading a book, or getting outside and getting some exesize.
Also, CM is known for hogging the shit out of your computer's resourses, leaving relatively little for the Winboard engine to run with. So it's not really a contest at all for CM, since it has the games pretty much fixed in it's favor unless you're running the programs from separate computers. Not to mention that CM is a very powerfull program (even though it doesn't cost very much)
Pete

[pete]

Geschrieben von: / Posted by: pete at 30 May 2000 08:15:08:
Als Antwort auf: / As an answer to: just a winboard chess engine question... geschrieben von: / posted by: Cheszekiah at 30 May 2000 02:15:05:

This may sound like a stupid question to most but I'll ask it anyway. I've played Crafty against other chess engines (Comet, Phalanx, AnMon, etc.) under the same time constrictions and it "seems" to me like whichever engine losses, learns to adapt and chooses a different path (same opening but different line somewhere in the middle game) from the previous game. If this is the case, then isn't it fair to say that these engines play better the more you play them? With that assumed, whenever I match any one of these engines against Chessmaster (slightly different time controls - say 1/1 for the winboard engines, and 1/3 or 1/5 with Chessmaster; I do this in hopes of improving the play of the engines by beating them repeatedly), I ALWAYS get the same result - a Chessmaster win or draw (CM has NEVER lost in the 60 or so matches that I've simulated; and yes, the engines do choose a different line somewhere in the middle of the game but they still end up losing, and often times, losing in embarrassing fashion - down by 3+ pawns is rather embarrassing to me). And when I do play the engines against CM with the same time controls, the win-loss ratio is usually 6:4 or 7:3 in favor of CM. So I wonder if these engines adapt and play a meaner game of chess with every loss?

From what I have observed Booklearning works great against opponents of similar strength but makes not much sense against much better or much worse opponents .
I try to explain with two very simple book-learners :
1.) The avoid losses book-learner
Exclude lines that lost games out of your book . But what happens if you lose a lot because your opponent is better ? You end up playing 1.a4 in the end maybe
This can be helped by starting with a huge book and doing the learning starting from the last point availlable but the general effect should be the same .
2. ) The percentage book learner
Play moves that succeeded in the past . This has also drawbacks ; assume you play hundreds of games against lesser opponents you can win against with _any_ opening , then from time to time you play against a titled player and get a good spanking . This loss will have virtually no effect to your book .
It seems having a not to deep but relatively wide book like CM has ( which has no booklearning but still performs well against the best engines with their refined learners ) is not much worse .
I should add this is in no way an expert opinion but just a user' s idea like yours , I have the impression though your "training" of the engines with CM won't help them much .
pete

[pete]

Geschrieben von: / Posted by: pete at 30 May 2000 12:05:39:
Als Antwort auf: / As an answer to: Re: just a winboard chess engine question... geschrieben von: / posted by: pete at 30 May 2000 08:15:08:
I've played Crafty against other chess engines (Comet, Phalanx, AnMon, etc.) under the same time constrictions and it "seems" to me like whichever engine losses, learns to adapt and chooses a different path (same opening but different line somewhere in the middle game) from the previous game.
I didn't read carefully enough ; I think if the first change is somewhere in the later middlegame it probably has nothing to do with learning at all .
You play at rather fast time controls and operate manually ; therefor the engines will get slightly different time for each move when you play another game ; when adding the normal randomness of search I think the engines didn't learn but in fact just choose another move instead ; they might very well use the 1st one again in another game .

[Aaron]

Geschrieben von: / Posted by: Aaron at 30 May 2000 19:30:21:
Als Antwort auf: / As an answer to: Re: just a winboard chess engine question... geschrieben von: / posted by: pete at 30 May 2000 08:15:08:

This may sound like a stupid question to most but I'll ask it anyway. I've >>played Crafty against other chess engines (Comet, Phalanx, AnMon, etc.) under >>the same time constrictions and it "seems" to me like whichever engine >>losses, learns to adapt and chooses a different path (same opening but >>different line somewhere in the middle game) from the previous game. If this >>is the case, then isn't it fair to say that these engines play better the >>more you play them? With that assumed, whenever I match any one of these >>engines against Chessmaster (slightly different time controls - say 1/1 for >>the winboard engines, and 1/3 or 1/5 with Chessmaster; I do this in hopes of >>improving the play of the engines by beating them repeatedly), I ALWAYS get >>the same result - a Chessmaster win or draw

From what I have observed Booklearning works great against opponents of >similar strength but makes not much sense against much better or much worse >opponents .
I try to explain with two very simple book-learners :
1.) The avoid losses book-learner
Exclude lines that lost games out of your book . But what happens if you lose >a lot because your opponent is better ? You end up playing 1.a4 in the end >maybe
2. ) The percentage book learner
Play moves that succeeded in the past . This has also drawbacks ; assume you >play hundreds of games against lesser opponents you can win against with _any_ >opening , then from time to time you play against a titled player and get a >good spanking . This loss will have virtually no effect to your book .
I should add this is in no way an expert opinion but just a user' s idea like >yours ,
I have the impression though your "training" of the engines with CM won't help >them much .

Agreed. A lot of times, the opening moves themselves are not totally wrong, just a squeunce of moves just are all slightly wrong that lead to a bad position. It will take a long time for today's book learning capabilities to learn what is right..
Besides I remember reading a article about Crafty without book..Amazingly it takes 4-5 tries before crafty concludes that 1 e4 e5 2 Nf3 Nc6 3 Bb5 a6 4 0-0? (force this obviously bad moves) is wrong..And play 4 Ba4!
1)Resulted based learning..
When Crafty loses a game, it goes back to the last line or something to try something different..
Crafty avoids this problem a little by weighting the effects based on opponent's rating..If Crafty managed to lose to a weaker player Crafty would normally win, it's likely that the move is bad..If Crafty loses to a higher rated player, the move is penalised less..
Also Evualtion based learning..Where engine analyses out of book and if the evulation is negative beyond a boundary, the engine marks the move as bad... This has the advnatage of indentifying not only objectively bad moves but also positions not suitable for the engine..For example, while the Black side of the KID is perfectly fine for KID players, a engine that does not now or appreciate the strenght of f5 will naturally score the position bad for black..
So, KID is not suitable for it..
For crafty it uses the average of the next X( i think x=10) moves as the evulation score..This is believed to give more accurate results than using a single evulation step..


Some humans do this!


Same here..
Altough the idea of self-adjusted learning opening books are really interesting to me..

I've had a similar idea of training crafty by getting large Book.lrn files to export into my books..So far I allowed Crafty to learn some 2 MB worth of book.lrn (from crafty users in chess servers) for twhat's it worth..

[pete]

Geschrieben von: / Posted by: pete at 30 May 2000 20:18:27:
Als Antwort auf: / As an answer to: Re: just a winboard chess engine question... geschrieben von: / posted by: Aaron at 30 May 2000 19:30:21:

Agreed. A lot of times, the opening moves themselves are not totally wrong, just a squeunce of moves just are all slightly wrong that lead to a bad position. It will take a long time for today's book learning capabilities to learn what is right.
Besides I remember reading a article about Crafty without book..Amazingly it takes 4-5 tries before crafty concludes that 1 e4 e5 2 Nf3 Nc6 3 Bb5 a6 4 0-0? (force this obviously bad moves) is wrong..And play 4 Ba4!

Exclude lines that lost games out of your book . But what happens if you lose >a lot because your opponent is better ? You end up playing 1.a4 in the end >maybe
2. ) The percentage book learner
Play moves that succeeded in the past . This has also drawbacks ; assume you >play hundreds of games against lesser opponents you can win against with _any_ >opening , then from time to time you play against a titled player and get a >good spanking . This loss will have virtually no effect to your book .
I should add this is in no way an expert opinion but just a user' s idea like >yours ,

1)Resulted based learning..
When Crafty loses a game, it goes back to the last line or something to try something different..
Crafty avoids this problem a little by weighting the effects based on opponent's rating..If Crafty managed to lose to a weaker player Crafty would normally win, it's likely that the move is bad..If Crafty loses to a higher rated player, the move is penalised less..
Also Evualtion based learning..Where engine analyses out of book and if the evulation is negative beyond a boundary, the engine marks the move as bad... This has the advnatage of indentifying not only objectively bad moves but also positions not suitable for the engine..For example, while the Black side of the KID is perfectly fine for KID players, a engine that does not now or appreciate the strenght of f5 will naturally score the position bad for black..
So, KID is not suitable for it..
For crafty it uses the average of the next X( i think x=10) moves as the evulation score..This is believed to give more accurate results than using a single evulation step..


Some humans do this!

Same here..
Altough the idea of self-adjusted learning opening books are really interesting to me..

I've had a similar idea of training crafty by getting large Book.lrn files to export into my books..So far I allowed Crafty to learn some 2 MB worth of book.lrn (from crafty users in chess servers) for twhat's it worth..

Or the opening itself was perfect but you were for example a little crafty on a P133 playing some 4-processor monster-crafty to stay completely in the one-program world although this also probably will be true for games against _very_ strong humans .
.
I think this is a misinformation ; the move 0-0 is just too unattractive for a chessprogram :-O ; it is true though that crafty was very fond of 4. Bc4 ; I remember some calling this the Crafty-Ruy


First I think it is enjoyable to think about book learners as you if user or if programmer have near perfect information , I mean it is quite easy to develop own ideas and follow others more knowledgeable ones when talking about it .
The crafty book-learner is pretty advanced I think and I tried to use the two simple book-learners to explain what I still think is a general problem though probably not to the one who asked the initial question .
The "first simple" book-learner for quite some time seemed to be just the one Chessbase programs used ; it was funny , I tried this with Junior5 on a slow P133 I think ; lost with 1. d4 => never 1.d4 again etc . ; and after a few games the book was completely screwed so it doesn't seem to be completely unusual .
The idea with the ratings is nice but it really is a thing only working on an ICS .
Also the idea with the time control or the one with learning bad moves even if you won the game eventually is really fine .
But perfection seems to be difficult in this area too . The problem with this approach is the dealing with draws I think .
In Crafty narrow book is used when playing computer ; in fact extremely annoying narrow book I think
But this is also an open field for "clever" abusers .
Against all crafties using default book since 17.0 you can get a draw as black every time when you are marked as "computer" by simply playing
1. e4 e5 2. Nf3 Nc6 3. Bb5 a6 4. Ba4 Nf6 5. 0-0 Be7 6. Re1 b5 7. Bb3 d6 8. c3 0-0 9. h3 Re8 10. d4 Bb7 11. Ng5 Rf8 12. Nf3 Re8 13. Ng5 Rf8 14. Nf3 Re8 1/2
The narrow book doesn't give crafty a chance to escape ( reported bug )


In fact I think this is a book learner many use . For example Shredder3 is a major example I recently found out ; after winning some 50 games against the Sicilian with 2.c3 it will never ( or at least not very soon ) give up on this variation again ; this is not what you will usually like as a user and also seems sub-optimal . Shredder4 seems to be able to also recognize refutations now .

same here ..
I think the Crafty book learner is _very_ good . For example it sits at ICC all day and plays whatever prog with cleverly tuned book might show up , then plays all the clever GM and IM players seeking holes . I think it is not very likely that there is a better crafty book than the one Dr Hyatt frequently uploads to his ftp-Site .

[Aaron]

Geschrieben von: / Posted by: Aaron at 31 May 2000 14:59:03:
Als Antwort auf: / As an answer to: Booklearning revisited geschrieben von: / posted by: pete at 30 May 2000 20:18:27:

Or the opening itself was perfect but you were for example a little crafty on >a P133 playing some 4-processor monster-crafty to stay completely in the one->program world although this also probably will be true for games against >_very_ strong humans .

Besides I remember reading a article about Crafty without book..Amazingly it >>takes 4-5 tries before crafty concludes that 1 e4 e5 2 Nf3 Nc6 3 Bb5 a6 4 0->>0? (force this obviously bad moves) is wrong..And play 4 Ba4!

2. ) The percentage book learner
Play moves that succeeded in the past . This has also drawbacks ; assume you >play hundreds of games against lesser opponents you can win against with _any_ >opening , then from time to time you play against a titled player and get a >good spanking . This loss will have virtually no effect to your book .

Some humans do this!
I've had a similar idea of training crafty by getting large Book.lrn files to >>export into my books..So far I allowed Crafty to learn some 2 MB worth of >>book.lrn (from crafty users in chess servers) for twhat's it worth..

I think this is a misinformation ;
the move 0-0 is just too unattractive for a >chessprogram :-O ; it is true >though that crafty was very fond of 4. Bc4 ;
First I think it is enjoyable to think about book learners as you if user or >if programmer have near perfect information , I mean it is quite easy to >develop own ideas and follow others more knowledgeable ones when talking about >it .
The crafty book-learner is pretty advanced I think and I tried to use the two >simple book-learners to explain what I still think is a general problem though >probably not to the one who asked the initial question .
The "first simple" book-learner for quite some time seemed to be just the one >Chessbase programs used ; it was funny , I tried this with Junior5 on a slow >P133 I think ; lost with 1. d4 => never 1.d4 again etc . ; and after a few >games the book was completely screwed so it doesn't seem to be completely >unusual .
The idea with the ratings is nice but it really is a thing only working on an >ICS .
Also the idea with the time control or the one with learning bad moves even if >you won the game eventually is really fine .
But perfection seems to be difficult in this area too .
The problem with this approach is the dealing with draws I think .
In Crafty narrow book is used when playing computer ; in fact extremely >annoying narrow book I think
But this is also an open field for "clever" abusers .
Against all crafties using default book since 17.0 you can get a draw as black >every time when you are marked as "computer" by simply playing
1. e4 e5 2. Nf3 Nc6 3. Bb5 a6 4. Ba4 Nf6 5. 0-0 Be7 6. Re1 b5 7. Bb3 d6 8. c3 >0-0 9. h3 Re8 10. d4 Bb7 11. Ng5 Rf8 12. Nf3 Re8 13. Ng5 Rf8 14. Nf3 Re8 1/2
The narrow book doesn't give crafty a chance to escape ( reported bug )
In fact I think this is a book learner many use .
For example Shredder3 is a major example I recently found out ; after winning >some 50 games against the >Sicilian with 2.c3 it will never ( or at least not >very soon ) give up on >this variation again
this is not what you will usually like as a user and also seems sub-optimal .
Shredder4 seems to be able to also recognize refutations now .
I think the Crafty book learner is _very_ good .
For example it sits at ICC all day and plays whatever prog with cleverly tuned >book might show up , then plays all the clever GM and IM players seeking >holes . I think it is not very likely that there is a better crafty book than >the one Dr Hyatt frequently uploads to his ftp-Site .

Yes , see comment on ratings..

Perhaps.. I doubt it was misinformation since it was a article by the creator of Crafty himself on a March issue of the ICCA journal i think..Of course , i could have misintrepreted it..It was a good article, one of the few that i understand (fully) or so i though..
True. Crafty would never play that on its own..But I believe Dr Hyatt inputted that loser move into the "small" book to see and demostarte how book learning worked..
Since crafty only starts to think when out of book after 4 0-0 , it doesn't know whether the poor evaultion(result based learning is another matter) that results is due to 4 0-0 or someelse down the line without a few tries..
Perhaps you should test this out..

The above sentence doesn't really have sense to me due to some typos . But I suspect that statement is not very complimentary to me somewhat..
Anyway, none of the above ideas are mine..They are actually implemented in Crafty..

Okay. I wasn't trying to correct you or something.I'm well aware your explaination is sufficent .When I first started out, I was also impressed by claims that chess programs now could "learn". Like the orginal poster I thought that this would make chess programs a lot stronger.
Now, I know better, the so called learning is still slow compared to human learning..

Just that I always wanted to chat about book learners and being a relative newbie here and in CC board , my posts are usually ignored espically when i start a thread..So I'm just taking advanatge of this chance..

Kinda dumb idea espically i think..Sorta of like a beginning player who shifts from opening to opening when the loss is not due to the openings..
Crafty which "understands" that it is expected to lose anyway against higher rated guys, is less dismayed by such losses and and is at least better than such novice players ..



Guess where, Crafty operated by Dr Hyatt hangs out at most of the time?
I believe in sole autogames you can also signal to crafty whether it's playing a computer plus it has a internal list of certain names like ferret, that warns it that it's opponent is a good one..

Yes so fine that it's in Crafty now! I'm not sure if other commericals use this as well, but it's really a brillant idea IMHO by DR Hyatt..

Perfection? Why are we talking about it?
Well learning even by Humans is trial and error. Even humans sometimes learn the wrong things..For example a beginner might stick with the stonewalk attack and win initally, but this stuns your future chess progress..


Drawes. Result based learning does not affect this..Sometimes Crafty wins but still marks the chosen move as "bad", when the evluation based learning kicks in.Basically crafty "knew" that it was lucky to win because it's evaulation showed that it was basically a piece down on average for the few moves out of book.
Both have their drawbacks , but together they cover quite a few bases..

Indeed..

Well that was good enough for Karpov in one game I believe..(He was Black)

No I mean, some chessplayers actually anlayse whether certain openings have better %wins and just play that..
Yes Fritz6 does something like that but is less extreme..

Okay..50 games against Shredder3...Wow!

Actually it is what i might do..I wouldn't play the sicilian after getting thrashed for some 50 or so games..But course who i lost to would also be a consideration..

What do you mean? You mean you actually have a refutation for the siclian with 2 c3! ???


Agreed..But Human learning is still faster..
Not really..The more games crafty plays the faster it learns..Of course, some learning will be bad, but statisically speaking a large sample is always better..
That's why Dr Hyatt implemented the feature of allowing people to import and sent book.lrn files..He wanted to allow crafty to learn faster if everyone helped shared what all the crafties learns..But unfortunately few people i know actually email their book.lrn files to Dr Hyatt for him to add to his book.lrn files..
Ultimately, the idea was for a crafty server, that allowed all versions of crafty to automatically upload and download new learning..So once one crafty learns something, all other connected ones learn instantly or at least very quickly..
Let me repeat none of the above is based on my ideas..I'm basically just repeating what I read..I have some ideas of my own, but i suspect that if you consider the ideas (which i believe are actually implemented in crafty) a little hard to achieve, my own will be laughed off..

[pete]

Geschrieben von: / Posted by: pete at 31 May 2000 16:29:34:
Als Antwort auf: / As an answer to: Re: Booklearning revisited geschrieben von: / posted by: Aaron at 31 May 2000 14:59:03:

Or the opening itself was perfect but you were for example a little crafty on >a P133 playing some 4-processor monster-crafty to stay completely in the one->program world although this also probably will be true for games against >_very_ strong humans .

Besides I remember reading a article about Crafty without book..Amazingly it >>takes 4-5 tries before crafty concludes that 1 e4 e5 2 Nf3 Nc6 3 Bb5 a6 4 0->>0? (force this obviously bad moves) is wrong..And play 4 Ba4!

2. ) The percentage book learner
Play moves that succeeded in the past . This has also drawbacks ; assume you >play hundreds of games against lesser opponents you can win against with _any_ >opening , then from time to time you play against a titled player and get a >good spanking . This loss will have virtually no effect to your book .

Some humans do this!
I've had a similar idea of training crafty by getting large Book.lrn files to >>export into my books..So far I allowed Crafty to learn some 2 MB worth of >>book.lrn (from crafty users in chess servers) for twhat's it worth..

I think this is a misinformation ;
the move 0-0 is just too unattractive for a >chessprogram :-O ; it is true >though that crafty was very fond of 4. Bc4 ;
First I think it is enjoyable to think about book learners as you if user or >if programmer have near perfect information , I mean it is quite easy to >develop own ideas and follow others more knowledgeable ones when talking about >it .
The idea with the ratings is nice but it really is a thing only working on an >ICS .
Also the idea with the time control or the one with learning bad moves even if >you won the game eventually is really fine .
But perfection seems to be difficult in this area too .
The problem with this approach is the dealing with draws I think .
In Crafty narrow book is used when playing computer ; in fact extremely >annoying narrow book I think
But this is also an open field for "clever" abusers .
Against all crafties using default book since 17.0 you can get a draw as black >every time when you are marked as "computer" by simply playing
1. e4 e5 2. Nf3 Nc6 3. Bb5 a6 4. Ba4 Nf6 5. 0-0 Be7 6. Re1 b5 7. Bb3 d6 8. c3 >0-0 9. h3 Re8 10. d4 Bb7 11. Ng5 Rf8 12. Nf3 Re8 13. Ng5 Rf8 14. Nf3 Re8 1/2
The narrow book doesn't give crafty a chance to escape ( reported bug )
In fact I think this is a book learner many use .
For example Shredder3 is a major example I recently found out ; after winning >some 50 games against the >Sicilian with 2.c3 it will never ( or at least not >very soon ) give up on >this variation again
this is not what you will usually like as a user and also seems sub-optimal .
Shredder4 seems to be able to also recognize refutations now .
I think the Crafty book learner is _very_ good .
For example it sits at ICC all day and plays whatever prog with cleverly tuned >book might show up , then plays all the clever GM and IM players seeking >holes . I think it is not very likely that there is a better crafty book than >the one Dr Hyatt frequently uploads to his ftp-Site .

Yes , see comment on ratings..
Perhaps.. I doubt it was misinformation since it was a article by the creator of Crafty himself on a March issue of the ICCA journal i think..Of course , i could have misintrepreted it..It was a good article, one of the few that i understand (fully) or so i though..
True. Crafty would never play that on its own..But I believe Dr Hyatt inputted that loser move into the "small" book to see and demostarte how book learning worked..
Since crafty only starts to think when out of book after 4 0-0 , it doesn't know whether the poor evaultion(result based learning is another matter) that results is due to 4 0-0 or someelse down the line without a few tries..
Perhaps you should test this out..

The above sentence doesn't really have sense to me due to some typos . But I suspect that statement is not very complimentary to me somewhat..
Anyway, none of the above ideas are mine..They are actually implemented in Crafty..


Just that I always wanted to chat about book learners and being a relative newbie here and in CC board , my posts are usually ignored espically when i start a thread..So I'm just taking advanatge of this chance..


Guess where, Crafty operated by Dr Hyatt hangs out at most of the time?
I believe in sole autogames you can also signal to crafty whether it's playing a computer plus it has a internal list of certain names like ferret, that warns it that it's opponent is a good one..
Yes so fine that it's in Crafty now! I'm not sure if other commericals use this as well, but it's really a brillant idea IMHO by DR Hyatt..

Perfection? Why are we talking about it?
Well learning even by Humans is trial and error. Even humans sometimes learn the wrong things..For example a beginner might stick with the stonewalk attack and win initally, but this stuns your future chess progress..


Drawes. Result based learning does not affect this..Sometimes Crafty wins but still marks the chosen move as "bad", when the evluation based learning kicks in.Basically crafty "knew" that it was lucky to win because it's evaulation showed that it was basically a piece down on average for the few moves out of book.
Both have their drawbacks , but together they cover quite a few bases..

Indeed..

Well that was good enough for Karpov in one game I believe..(He was Black)

No I mean, some chessplayers actually anlayse whether certain openings have better %wins and just play that..
Yes Fritz6 does something like that but is less extreme..

Okay..50 games against Shredder3...Wow!
Actually it is what i might do..I wouldn't play the sicilian after getting thrashed for some 50 or so games..But course who i lost to would also be a consideration..

What do you mean? You mean you actually have a refutation for the siclian with 2 c3! ???
Agreed..But Human learning is still faster..
Not really..The more games crafty plays the faster it learns..Of course, some learning will be bad, but statisically speaking a large sample is always better..
That's why Dr Hyatt implemented the feature of allowing people to import and sent book.lrn files..He wanted to allow crafty to learn faster if everyone helped shared what all the crafties learns..But unfortunately few people i know actually email their book.lrn files to Dr Hyatt for him to add to his book.lrn files..
Ultimately, the idea was for a crafty server, that allowed all versions of crafty to automatically upload and download new learning..So once one crafty learns something, all other connected ones learn instantly or at least very quickly..
Let me repeat none of the above is based on my ideas..I'm basically just repeating what I read..I have some ideas of my own, but i suspect that if you consider the ideas (which i believe are actually implemented in crafty) a little hard to achieve, my own will be laughed off..

I think though this will _still_ spoil your book and probably it would be better to turn learning off if your opponent is something like 500 points stronger than you .
Interesting , so the misunderstanding was obviously on my side . Is this article already also printed on the net somewhere ?

Nope . They are not intended to say anything unfriendly about you or anybody . This is the permanent terror of different native languages . What I _wanted_ to say is that it is fun to discuss those ideas for me ( and same seems to be true for you as you state later ) because they are easy enough to understand for _me_ . Sorry if I misphrased it too much .

same here

Sure , but so it won't work for another program which doesn't hang out there very well ; this was what I wanted to say .
Yes , there are many GM games with it but this is slightly different I think . I think he wouldn't play it over and over against _me_ and let me score 25% against him
I simulated this by naming myself as a low-rated compi on ICC and crafty still can't escape .
Hmm , let me put it in other words . The problem is when the chessprogram chooses some ok but not optimal move in the opening and plays some low-rated players a few hundred times ( common on ICC for example ) the not so great move will become more and more attractive ; then if you play the big guys you have a somewhat screwed book again ; for example my shredder is currently _extremely_ fond of the 4 knights game and prefers it over the Ruy . It isn't a major problem but winning 100 games against some patzer may change the percentages in your book too much .

Nope . What I mean is : when playing a certain opening or better special opening line over and over and winning and winning ; then later this line has been refuted ; although your percentage is still ok it is time to give up on it , best before you lost too many games .
Yes , I think this idea is nice and sure any learning data will help . I simply think it doesn't matter _too_ much that only few do that as Mr Hyatt himself gets zillions of games from crafty and scrappy so probably a few hundreds more or less won't matter too much .
I think it is interesting to discuss since even the top programs seem to be miles away from even average human chess player learning abilities and it's fun to think about ideas to improve the automated book-learners . I don't care too much about people laughing about my ideas ; on the other hand I usually learn lots from the opinions expressed by others and would also be interested in hearing yours .

[Aaron]

Geschrieben von: / Posted by: Aaron at 01 June 2000 15:29:33:
Als Antwort auf: / As an answer to: Re: Booklearning revisited geschrieben von: / posted by: pete at 31 May 2000 16:29:34:

Yes , see comment on ratings..
Perhaps you should test this out..

The idea with the ratings is nice but it really is a thing only working on >>>an ICS .
Against all crafties using default book since 17.0 you can get a draw as >>>black every time when you are marked as "computer" by simply playing
Shredder4 seems to be able to also recognize refutations now .

Guess where, Crafty operated by Dr Hyatt hangs out at most of the time?
Sure , but so it won't work for another program which doesn't hang out there >>very well ; this was what I wanted to say .
Well that was good enough for Karpov in one game I believe..(He was Black)
Actually it is what i might do..I wouldn't play the sicilian after getting >>thrashed for some 50 or so games..But course who i lost to would also be a >>consideration..
What do you mean? You mean you actually have a refutation for the siclian >>with 2 c3! ???
Not really..The more games crafty plays the faster it learns..Of course, some >>learning will be bad, but statisically speaking a large sample is always >>better..
That's why Dr Hyatt implemented the feature of allowing people to import and >>sent book.lrn files..

I think though this will _still_ spoil your book and probably it would be >better to turn learning off if your opponent is something like 500 points >stronger than you .
Interesting , so the misunderstanding was obviously on my side . Is this >article already also printed on the net somewhere ?
Yes , there are many GM games with it but this is slightly different I think . >I think he wouldn't play it over and over against _me_ and let me score 25% >against him
I simulated this by naming myself as a low-rated compi on ICC and crafty still >can't escape .
Hmm , let me put it in other words . The problem is when the chessprogram >chooses some ok but not optimal move in the opening and plays some low-rated >players a few hundred times ( common on ICC for example ) the not so great >move will become more and more attractive ; then if you play the big guys you >have a somewhat screwed book again ; for example my shredder is currently >_extremely_ fond of the 4 knights game and prefers it over the Ruy . It isn't >a major problem but winning 100 games against some patzer may change the >percentages in your book too much .
Nope . What I mean is : when playing a certain opening or better special >opening line over and over and winning and winning ; then later this line has >been refuted ; although your percentage is still ok it is time to give up on >it , best before you lost too many games .
Yes , I think this idea is nice and sure any learning data will help . I >simply think it doesn't matter _too_ much that only few do that as Mr Hyatt >himself gets zillions of games from crafty and scrappy so probably a few >hundreds more or less won't matter too much .
I think it is interesting to discuss since even the top programs seem to be >miles away from even average human chess player learning abilities and it's >fun to think about ideas to improve the automated book-learners . I don't care >too much about people laughing about my ideas ; on the other hand I >usually learn lots from the opinions expressed by others and would also be >interested in hearing yours .

Or just set a boundary level beyond which learning weights are so low that they have almost no effect..I think this is more a problem when playing stronger players than weaker players.

If I understand correctly, some programs like Crafty only learn from negative data, not postive ones.
I remember Mogan Larsens remarking that Crafty won a game against another engine (cant remember what, but one with no book learning) in the second game, both engines repeated the sames moves, but Crafty was first to vary from the previous winning line and instead lost!
Seems to me, Crafty doesn't learn from postive moves?? Or the randomness factor was set high enough to swamp this learning..
What I'm saying is that the learning seems more geared towards rooting out bad moves that lead to inferior positions than finding the absolutely best move..


Nope.. There was some talk in CCC about putting past issues of ICCA journals online though..


Well, okay..But for all the crafty operators out there who run crafty clones, this is a boon..






In a sense, you are cheating..Regardless of how good crafty is it still relies a lot on the data that is fed to it..GIGO is what they say..
Normally though, I think crafty has certain code that allows it to play more aggressive even swindling mode against weaker players..Similarly the contempt level setting for others..

Probably, what i don't think shredder considers the level of opponent that it is playing against..Another question would be why in the first place does Shredder play the 4 knights game?
2 posibilities
1) The programmers of the orginal handtuned books think it's better suited to Shredder's postional style..After all, who can say the ruy lopez is definitely better? Though GM practice sugguests that of course..
2) If the program is complied from a database of good modern GM games, I suspect ruylopez would be favoured..Playing a Pazter either way it would win..So unless you some how started to score against it when defending the ruy lopez, i don't understand why it would play the four knights..
If you are actually strong enough, for the opening choice to make a difference, you are hardly a patzer and it's learning is valid i guess..
In actualy fact , I suspect it would continue to stick even after learning to ruy lopez if it prefers that in the first place.


So say the percentage is initally 100% ..100 games out of 100..but in the next 10, Shredder loses all 10..so even if the winning percentage is still 90% shredder won't play it?
Risky..This makes playing against a stronger player like say Kasaprov even more risky..I suspect shredder would soon give up 1 d4 after losing 10 games to KID in a row..


zillions? Come on, with all the crafty clones out there, I'm sure if everyone contributes, they can easily dwarf what a single person can do..


Okay I'm quite tired now..Here's one idea..
The idea of taking into account the ratings of players when adjusting the opening book can be extended..
Perhaps, instead of using the same book against all players, crafty can adjust it's play by building up books designed for play against different players depending on their rating.
In a way, Crafty does this by segreting into 2 books , one for Humans and one for computers. But I think this is too broad. Since it assumes that all computer engines are of equal strenght..
My idea is this.. Instead of having one overall book, perhaps it could have various books for players between 1600 and below, 1600-2000 ,2000-2200 2200+ etc..
It could initally start with one uniform book, but learning weights from games would be applied 100% only to the correct books . So a risky gambit that won against a 1400 player would only result in learning that applied to that book..
So Crafty might learn that the whatever gambit works against those between 1600 and below..But it applies less or not at all to the higher classes..Similarly, if it get thrashed using the same gambit against a GM, it will continue to play that against a mere 1400 player.
Or, the learning would continue to apply to other books but to a lesser extent..Hence this will allow more interesting play against various classes of players rather than being too influenced by games against either too weak or too strong players..
I'm not sure how it could figure out your rating though if not playing or servers besides using a crude indication like the average number of moves before you are outplayed..