NPS challenge will end after game one...
Ed Schroder
NPS challenge will end after game one...
Today I decided to end the NPS challenge between Crafty and Rebel8 after
game-1 is finished.
Understanding the problems of running a match with an unfinished
product, and since there are so many things that Bob may claim and so
many things that did go wrong with the book, evals, time control,
contempt factor, huge auto232 problems, bugs etc. I consider it a waste
of my time to continue the match.
Besides of this I also want to express my feelings about the following.
I am interested in the scientific relationship between ply depth and an
evaluation function. That's what has started this match. Sofar game-1
has proven my point of view at least in my opinion.
I am *not* interested to read all kind of possible and impossible
explanations that imply to hide the real reason why Crafty will lose
game-1.
For instance, Bob Hyatt wrote:
: I'll post some interesting analysis later, but you might be surprised
: at some of what went on in this match. For example. in at least two
: places, Crafty played a weaker move, simply because it saw, at
: depth=15, something that it thought Rebel would play and it didn't
: like it. In these two cases I know of, I've tested rebel on the
: position and it wouldn't have played the move Crafty feared. So
: Crafty simply avoided something that would not have happened, by
: playing something that was worse. It was one of the common scenarios
: I've seen when two programs play and one outsearches the other by
: a significant amount.
This would mean Crafty / Rebel would come into trouble against the
Mephisto I, Boris etc. because Crafty / Rebel think 4-5 plies deeper.
This is of course total nonsense. Crafty as any other program just
played the best move on a given depth.
I wouldn't dare to mention such an argument if Rebel would have had
the 100 x time advantage!
This is just one example...
I am not in the mood for this kind of explanations for another 9 games
the next 5-6 months. My time is too precious for that.
Rebel8 played this game on an average ply depth of 9/10.
Crafty played on an average ply depth of 13/14.
Still Rebel8's evaluation function was good enough to handle the big
ply gap of 4 plies. This has always been my point and game-1 has
proven it. I believe this also counts for other chess programs like
for instance Hiarcs etc.
- Ed Schroder -
Jack Nerad
I'm sorry to hear this. Does anyone else have the capability (and
desire) to run an experiment similar to this one? I certainly don't
have the capability (although, I do have the desire) to run a match like
this, but am truly interested in the outcome.
Regards,
Jack Nerad
Robert Hyatt
: Today I decided to end the NPS challenge between Crafty and Rebel8 after
: game-1 is finished.
: Understanding the problems of running a match with an unfinished
: product, and since there are so many things that Bob may claim and so
: many things that did go wrong with the book, evals, time control,
: contempt factor, huge auto232 problems, bugs etc. I consider it a waste
: of my time to continue the match.
: Besides of this I also want to express my feelings about the following.
: I am interested in the scientific relationship between ply depth and an
: evaluation function. That's what has started this match. Sofar game-1
: has proven my point of view at least in my opinion.
While I won't argue about game one, one game is not enough. If you
still feel strongly about ply depth, try the alternative experiment
I suggested, which is Rebel vs Rebel at 100:1 time odds. That will
provide better data than Rebel vs Crafty, because the *only* thing
different will be the depth. No differences in eval, no auto232
problems, etc. And the results will be just as interesting...
: I am *not* interested to read all kind of possible and impossible
: explanations that imply to hide the real reason why Crafty will lose
: game-1.
: For instance, Bob Hyatt wrote:
: : I'll post some interesting analysis later, but you might be surprised
: : at some of what went on in this match. For example. in at least two
: : places, Crafty played a weaker move, simply because it saw, at
: : depth=15, something that it thought Rebel would play and it didn't
: : like it. In these two cases I know of, I've tested rebel on the
: : position and it wouldn't have played the move Crafty feared. So
: : Crafty simply avoided something that would not have happened, by
: : playing something that was worse. It was one of the common scenarios
: : I've seen when two programs play and one outsearches the other by
: : a significant amount.
I don't see anything misleading at all in the above. Nor do I see any
"excuse" making either. The thing I pointed out above happened. It has
happened many times in the past, it will continue to happen many times in
the future. And it's a problem that is going to have to be addressed at
some point in time.
But, for the record, have you ever had Rebel play against a strong human,
and saw Rebel throw the game away? Because it saw some deep tactic that
it thought needed looking after, but which cost Rebel a pawn (or more) to
defend against? A tactic that in reality, the human didn't see? If you
haven't seen this, you simply haven't looked closely enough.
A year ago, I posted an ugly game played by Cray Blitz vs a 1900 player
in the 1984 US Open speed chess championship. CB had gone through IM's
and GM's like a knife through butter, and near the end of the first
"section" match (round robin with 16 players I think) we hit a 1900
player, and were doing o.k. Out of the clear, CB sacrificed it's queen
for a knight. Victor Korchnoi was looking over my shoulder and said
"machine make terrible blunder" in his deep, Russian gutteral accent.
After the game, we backed up and I asked him what to play. He suggested
something like RxR which appeared to be hanging. I tried it, and CB
instantly announced a mate in 11 for the opponent. It saw that in the
game and sacrificed the queen to avoid it. At the time, Korchnoi was the
#3 player in the world I believe, and was Karpov's second at the time.
He didn't see the mate. I doubt the 1900 player would have either.
Yet CB, because it just happened to search deeply enough, did see it,
and threw the game away because of it. This was the only game it lost
as I recall, also. That happened in round 1 of the NPS, although it was
not a mate, nor even a material loss... That's not knocking rebel, it's
not making excuses for Crafty losing. It is simply pointing out the fact
that a deeper search doesn't always result in better play. In this game,
Crafty played some poor moves to hang on to the passed E pawn, when it
should not have. And this put it into positions where it eventually had
to give up something somewhere. The deeper search, however, made it think
it had to give up more than it really did.
Sorry you took it as offensive. It wasn't intended to be. I didn't try to
make *any* excuses for Crafty. I didn't mention that in the 12 game
"calibration match" that crafty had black in every game, which would likely
affect the outcome by some amount. I didn't play up the fact that for the
first 7 games, crafty was playing with a gross "contempt factor" because it
thought Rebel had zero time left, and the draw score code sets contempt to
-1.000 in that case to try and avoid drawing when the opponent is about to
lose on time.
In any case, as I mentioned before, the match was on your "nickel" and you
are certainly free to stop whenever you want. I knew it would take a long
time to play out. The calibration match was having problems due to auto232
quirks, but the one NPS game was simply clean so far as I'm concerned. The
book might have affected the game, it might not. However, when Rebel left
book, I didn't consider Crafty's position so inferior that the game should
be thrown out. Rebel played well. I think the most impressive thing I saw
was the fact that it did not seem to make *any* tactical mistakes, anywhere.
Which is surprising. I'd have liked to see more games to see if this could
continue.
: This would mean Crafty / Rebel would come into trouble against the
: Mephisto I, Boris etc. because Crafty / Rebel think 4-5 plies deeper.
: This is of course total nonsense. Crafty as any other program just
: played the best move on a given depth.
I'd bet that Mephisto I would win one game in a lengthy match. And I'd
really bet that once or twice either program would try to dodge some
tactical bullet that Mephisto would have absolutely no chance of firing.
: I wouldn't dare to mention such an argument if Rebel would have had
: the 100 x time advantage!
: This is just one example...
: I am not in the mood for this kind of explanations for another 9 games
: the next 5-6 months. My time is too precious for that.
: Rebel8 played this game on an average ply depth of 9/10.
: Crafty played on an average ply depth of 13/14.
: Still Rebel8's evaluation function was good enough to handle the big
: ply gap of 4 plies. This has always been my point and game-1 has
: proven it. I believe this also counts for other chess programs like
: for instance Hiarcs etc.
: - Ed Schroder -
The only thing you say that I don't agree with is the above. One game doesn't
prove anything. That was the intent for playing 10 games to start with, rather
than a single game.
However, in the interests of seeing how this works, I will see if I can set
up a "correspondence handicap game" with someone that is willing to operate
Rebel on the other end, just to see how this might continue. I can operate
it manually here, but don't want to give anyone the chance to cry foul...
It's still an experiment that needs doing, because it has never been done
before. Everyone agrees that one ply is not a significant advantage at these
depths. But we have absolutely no data about 4-5 plies. I hope you might
consider doing rebel vs rebelX100, as that would generate the data.
ShaktiFire
I guess I can understand Ed's position.
There were alot of glitches with Crafty.
First its big book didn't work. Then its timing
mechanism was screwed up in the even time
match. I would get tired of it too. He just wanted
to set up the auto play and let it run. After all Ed has
other things to worry about.
But I guess, from a marketing standpoint, this
is an excellent time to end the match :-)
Best Wishes
Don Fong
In article <5hbh7j$s91$1...@news1.xs4all.nl>,
Ed Schroder <rebc...@xs4all.nl> wrote:
>Today I decided to end the NPS challenge between Crafty and Rebel8 after
>game-1 is finished.
too bad! maybe YOU feel the point is proved, but to many people,
the point would be better proved by multiple games.
[...]
>I am *not* interested to read all kind of possible and impossible
>explanations that imply to hide the real reason why Crafty will lose
>game-1.
aw cmon. it's fun watching Hyatt squirm. (:-) (:-)
>For instance, Bob Hyatt wrote:
>
>: I'll post some interesting analysis later, but you might be surprised
>: at some of what went on in this match. For example. in at least two
>: places, Crafty played a weaker move, simply because it saw, at
>: depth=15, something that it thought Rebel would play and it didn't
>: like it. In these two cases I know of, I've tested rebel on the
>: position and it wouldn't have played the move Crafty feared. So
>: Crafty simply avoided something that would not have happened, by
>: playing something that was worse. It was one of the common scenarios
>: I've seen when two programs play and one outsearches the other by
>: a significant amount.
i think this is a tremendous argument!! Bob has just proved that
increased search depth is a handicap. (:-) (:-)
*FUTURE NEWSFLASH* Deep Blue XXVII, playing the Black pieces
in a correspondence game, played the stunning move 1...Resigns!
>This would mean Crafty / Rebel would come into trouble against the
>Mephisto I, Boris etc. because Crafty / Rebel think 4-5 plies deeper.
>This is of course total nonsense. Crafty as any other program just
>played the best move on a given depth.
>
>I wouldn't dare to mention such an argument if Rebel would have had
>the 100 x time advantage!
no? i think we must do the experiment!
>This is just one example...
>
>I am not in the mood for this kind of explanations for another 9 games
>the next 5-6 months. My time is too precious for that.
shucks. the public wants more!!
>Rebel8 played this game on an average ply depth of 9/10.
>Crafty played on an average ply depth of 13/14.
>
>Still Rebel8's evaluation function was good enough to handle the big
>ply gap of 4 plies. This has always been my point and game-1 has
>proven it. I believe this also counts for other chess programs like
>for instance Hiarcs etc.
i don't think it is fair to call off the match after one game.
and what if CRAFTY pulls off a miracle? (it aint over til it's over.)
i know i wouldn't be silent if Hyatt called off the match after one game.
this is an historic event. what if Spassky had ended the Reykjavic
match after the first 2 games? (as he probably could have done.)
mclane
Ed Schroder <rebc...@xs4all.nl> wrote:
>NPS challenge will end after game one...
>Today I decided to end the NPS challenge between Crafty and Rebel8 after
>game-1 is finished.
Oh ! What a pity !
>Understanding the problems of running a match with an unfinished
>product, and since there are so many things that Bob may claim and so
>many things that did go wrong with the book, evals, time control,
>contempt factor, huge auto232 problems, bugs etc. I consider it a waste
>of my time to continue the match.
I understand. An experiment must come to A result.
>Besides of this I also want to express my feelings about the following.
>I am interested in the scientific relationship between ply depth and an
>evaluation function. That's what has started this match. Sofar game-1
>has proven my point of view at least in my opinion.
I think, it has. Although the prove is a little prove.
>I am *not* interested to read all kind of possible and impossible
>explanations that imply to hide the real reason why Crafty will lose
>game-1.
It seems programs have the same problems with LOSING a game,
chess-players have :-)
>For instance, Bob Hyatt wrote:
>: I'll post some interesting analysis later, but you might be surprised
>: at some of what went on in this match. For example. in at least two
>: places, Crafty played a weaker move, simply because it saw, at
>: depth=15, something that it thought Rebel would play and it didn't
>: like it. In these two cases I know of, I've tested rebel on the
>: position and it wouldn't have played the move Crafty feared. So
>: Crafty simply avoided something that would not have happened, by
>: playing something that was worse. It was one of the common scenarios
>: I've seen when two programs play and one outsearches the other by
>: a significant amount.
To Bob: If your program is so silly to consider that your opponent
will play always the best move, than that is your programs problem. I
mean: thats really the old problem: you expect the opponent to play a
move that your program has found to be the best. And now you make a
defender move that is a very passive one but helps against the
"THREAD" of the opponent THAT COULD HAPPEN IN PLY 15.
But whose fault is it? Rebels ?! No - Rebel can play whatever it
wants. It's your fault, Bob. A few weeks ago we discussed much about
CSTal and Chris and I told you, that CSTal does NOT calculate accurate
and best play. Because NOBODY knows what BEST-PLAY is.
Bruce read all the stuff I told about how CSTal works, and said he
liked it. The idea that the opponent will move the RIGHT moves is
silly. If your program does not SEE / FEEL that all it's own ideas are
subjective and not objective, it is a programming-bug, or , to say it
not that drastically: it's a lack of recognition of the reality.
Alpha-Beta might be nice, but it does not help much here...
So I understand Ed's reaction. I think ED KNOWS that DIFFERENT ideas
about HOW TO PLAY are just IDEAS. Only fools would post that this
happens because this happens because this and this is forced because
of this and that, in a dicipline where NOTHING is safe.
Computerchess IS a topic where nothing is obvious, maybe a forced mate
is obvious, but anything out of this range is subjective. Search-trees
are very nice, but they do not help to solve the main-problem: the
range is almost infinite, or infinte for search-trees. Therefore Ed
and others trust more evaluations.
>This would mean Crafty / Rebel would come into trouble against the
>Mephisto I, Boris etc. because Crafty / Rebel think 4-5 plies deeper.
!!! :-)
>This is of course total nonsense. Crafty as any other program just
>played the best move on a given depth.
"best move" !!!
Ed is right (IMO): it is totally nonsense. But maybe Bob is not SEEING
this.
Because we (the subscibers of the search-tree effort, and the
subscibers of the evaluation-function-effort) have different point of
views.
>I wouldn't dare to mention such an argument if Rebel would have had
>the 100 x time advantage!
>This is just one example...
But i can understand your feelings, Ed. It is like fighting against
wind-mills. You fight against people who are not capable to UNDERSTAND
you. Maybe if they would understand, they would still disagree. But
this way, it is chanceless.
Both sides are too different. What a pity. How can we ever understand
each other, although speaking the same language.
This was my joke with the GERMAN-language. We speak english, crafty
plays chess against Rebel, we all know the notation and refer to the
same chess-rules, also a game is won or lost or draw, we all refer to
the same paramters. BUT STILL WE TOTALLY DISAGREE.
What happens here ? This is really a paradigm shift ! The subscribers
of the old-age are unable to understand the subscibers of the new-age.
It is a turning-point, IMO.
That causes much of the problems here.
A few months ago Bob called Chris a subsciber of an old-fashioned
idea, somebody who lives and tries an idea that was given up by
university-guys ages ago. I remember Ed said: if Chris is old
fashinoned with his ideas, then my ideas are old-fashioned too,
because I am doing it the same way.
Hm. Ed has shown that his program plays very very strong. I mean: if
they are old-fashioned or senseless, than this is a very nice way of
doing a senseless idea.
>I am not in the mood for this kind of explanations for another 9 games
>the next 5-6 months. My time is too precious for that.
>Rebel8 played this game on an average ply depth of 9/10.
>Crafty played on an average ply depth of 13/14.
>Still Rebel8's evaluation function was good enough to handle the big
>ply gap of 4 plies. This has always been my point and game-1 has
>proven it. I believe this also counts for other chess programs like
>for instance Hiarcs etc.
I subscibe Eds point of view. The ideas of chess-programming are
somehow divided. It looks that the opponent sides will not be able to
have a fruitful exchange of ideas, due to incompatible feelings, point
of views, language and programming efforts.
In my opinon chris could show them (FAST/STUPID-search fans) his
source-code in C, they would not understand anything. They would not
understand the ideas behind, because in their inner mind, they don't
want to understand.
Please Bob, I don't want to offend you with this statements. THIS was
the reason I changed into german-language some weeks before, because I
had the idea, we could talk the same language, and refer to the same
set of rules, still communication would not lead to anything.
This event has maybe proved it again. What a depressing moment.
But still we should not stop communication. I understand Ed. He has to
control a company with employees and has only a small amount of
spare-time for discussions in the Internet. IMO Ed has tried to
demonstrate his ideas here, so Chris has tried. I would be pleased if
we could find a language or way to communicate much easier, with less
problems to understand each other.
I want to start with a sac' on my own, maybe we can only make progress
if we stop behaving idle and arrogant and try to forgive and react
without the fear of losing our face because we have done a mistake
THEREFORE I would like to start with an official public apologize:
As a member of the Anti-Computer-Schach & Spiele-Group
I want to excuse here in public for any offending word or posting I
have done in the past against this big german computer-chess magazine.
I have written many statements and posts about this magazine, because
I felt angry and hurted from their work, from their publications.
Maybe it is time to go one step into another direction. So I will try
to change my style and point-of-views concerning this magazine and
will try to analyse my OWN mistakes and weaknesses over the years more
carfefully and want to see if maybe my own subjective-view has caused
many misunderstandings in this area.
Therefore I want to excuse here in public for any anti-computerschach
& spiele post in the years before.
I will not stop to watch out with open eyes, and will not stop speak
out my ideas uncensored, but I will try to be more tolerant to other
people's point-of-views and statements.
I apologize to Dieter Steinwender and Frederic Friedel. I hope this
message reaches the destinations.
>- Ed Schroder -
I also hope that Ed will still join this newsgroup any further. For
the first time in computerchess we have a forum where all of us have
the choice to communicate. We have to rebuild a new culture and have
to learn that this is a chance but that this forum is also a prove for
us, if we can really live together with all ou different opinons in
peace and not in depressing-fights that nobody profits from.
Thanks for listening.
Please Ed, stay here!
Don't stop talking...
Robert Hyatt
Don Fong (df...@cse.ucsc.edu) wrote:
: In article <5hbh7j$s91$1...@news1.xs4all.nl>,
: Ed Schroder <rebc...@xs4all.nl> wrote:
: >Today I decided to end the NPS challenge between Crafty and Rebel8 after
: >game-1 is finished.
: too bad! maybe YOU feel the point is proved, but to many people,
: the point would be better proved by multiple games.
: [...]
: >I am *not* interested to read all kind of possible and impossible
: >explanations that imply to hide the real reason why Crafty will lose
: >game-1.
: aw cmon. it's fun watching Hyatt squirm. (:-) (:-)
: >For instance, Bob Hyatt wrote:
: >
: >: I'll post some interesting analysis later, but you might be surprised
: >: at some of what went on in this match. For example. in at least two
: >: places, Crafty played a weaker move, simply because it saw, at
: >: depth=15, something that it thought Rebel would play and it didn't
: >: like it. In these two cases I know of, I've tested rebel on the
: >: position and it wouldn't have played the move Crafty feared. So
: >: Crafty simply avoided something that would not have happened, by
: >: playing something that was worse. It was one of the common scenarios
: >: I've seen when two programs play and one outsearches the other by
: >: a significant amount.
: i think this is a tremendous argument!! Bob has just proved that
: increased search depth is a handicap. (:-) (:-)
No... just that it doesn't always work out to the best of circumstances.
Easy to find cases of this... just play a computer against a human
opponent and if you do it enough times, the machine will make a move
that you can see by the expression on the human's face is a complete
surprise. A surprise because it looks bad...
: *FUTURE NEWSFLASH* Deep Blue XXVII, playing the Black pieces
: in a correspondence game, played the stunning move 1...Resigns!
: >This would mean Crafty / Rebel would come into trouble against the
: >Mephisto I, Boris etc. because Crafty / Rebel think 4-5 plies deeper.
: >This is of course total nonsense. Crafty as any other program just
: >played the best move on a given depth.
: >
: >I wouldn't dare to mention such an argument if Rebel would have had
: >the 100 x time advantage!
: no? i think we must do the experiment!
: >This is just one example...
: >
: >I am not in the mood for this kind of explanations for another 9 games
: >the next 5-6 months. My time is too precious for that.
: shucks. the public wants more!!
: >Rebel8 played this game on an average ply depth of 9/10.
: >Crafty played on an average ply depth of 13/14.
: >
: >Still Rebel8's evaluation function was good enough to handle the big
: >ply gap of 4 plies. This has always been my point and game-1 has
: >proven it. I believe this also counts for other chess programs like
: >for instance Hiarcs etc.
: i don't think it is fair to call off the match after one game.
: and what if CRAFTY pulls off a miracle? (it aint over til it's over.)
: i know i wouldn't be silent if Hyatt called off the match after one game.
: this is an historic event. what if Spassky had ended the Reykjavic
: match after the first 2 games? (as he probably could have done.)
As I said, I'd be just as happy with rebel vs rebelX100... it's the depth
I'm interested in, not the personalities, nor the identities of the two
programs. Rebel vs RebelX100 would be a much better experiment anyway,
since there's only one variable...
Robert Hyatt
mclane (mcl...@prima.ruhr.de) wrote:
: To Bob: If your program is so silly to consider that your opponent
: will play always the best move, than that is your programs problem. I
: mean: thats really the old problem: you expect the opponent to play a
: move that your program has found to be the best. And now you make a
: defender move that is a very passive one but helps against the
: "THREAD" of the opponent THAT COULD HAPPEN IN PLY 15.
Of course, *every* program currently playing also makes this same
assumption. You might go read a little about alpha/beta and minimax,
and zero-sum game theory...
: But whose fault is it? Rebels ?! No - Rebel can play whatever it
: wants. It's your fault, Bob. A few weeks ago we discussed much about
: CSTal and Chris and I told you, that CSTal does NOT calculate accurate
: and best play. Because NOBODY knows what BEST-PLAY is.
No, there you are wrong. CSTal *does* play the move that it's search
and evaluation can find. It is still based on alpha/beta/minimax, which
is constrained to do this. Just because its eval likes to go off the
deep end on complications, it still assumes the opponent will make the
best move possible *according to the evaluation/search* in CSTal...
: Bruce read all the stuff I told about how CSTal works, and said he
: liked it. The idea that the opponent will move the RIGHT moves is
: silly. If your program does not SEE / FEEL that all it's own ideas are
: subjective and not objective, it is a programming-bug, or , to say it
: not that drastically: it's a lack of recognition of the reality.
: Alpha-Beta might be nice, but it does not help much here...
: So I understand Ed's reaction. I think ED KNOWS that DIFFERENT ideas
: about HOW TO PLAY are just IDEAS. Only fools would post that this
: happens because this happens because this and this is forced because
: of this and that, in a dicipline where NOTHING is safe.
Sorry, but chess is a lot more precise than you are giving it credit
for. We already know about N-piece endings and perfect play. There
*is* a perfect result available in chess, we just don't know what it
is, and may never know what it is. However, N ply searches produce
information that is useful. N+1 ply searches produce *more* information.
Otherwise, I'll expect everyone to continue using those old P90's for
a long time, because that extra speed on the P6/200 obviously is not
going to help them play any better. That *is* the argument, isn't it?
: Computerchess IS a topic where nothing is obvious, maybe a forced mate
: is obvious, but anything out of this range is subjective. Search-trees
: are very nice, but they do not help to solve the main-problem: the
: range is almost infinite, or infinte for search-trees. Therefore Ed
: and others trust more evaluations.
: >This would mean Crafty / Rebel would come into trouble against the
: >Mephisto I, Boris etc. because Crafty / Rebel think 4-5 plies deeper.
: !!! :-)
: >This is of course total nonsense. Crafty as any other program just
: >played the best move on a given depth.
: "best move" !!!
: Ed is right (IMO): it is totally nonsense. But maybe Bob is not SEEING
: this.
: Because we (the subscibers of the search-tree effort, and the
: subscibers of the evaluation-function-effort) have different point of
: views.
: >I wouldn't dare to mention such an argument if Rebel would have had
: >the 100 x time advantage!
: >This is just one example...
: But i can understand your feelings, Ed. It is like fighting against
: wind-mills. You fight against people who are not capable to UNDERSTAND
: you. Maybe if they would understand, they would still disagree. But
: this way, it is chanceless.
Speaking of wind-mills... I can feel the breeze all the way to
Birmingham... maybe you can tone it down a bit?
: Both sides are too different. What a pity. How can we ever understand
: each other, although speaking the same language.
: This was my joke with the GERMAN-language. We speak english, crafty
: plays chess against Rebel, we all know the notation and refer to the
: same chess-rules, also a game is won or lost or draw, we all refer to
: the same paramters. BUT STILL WE TOTALLY DISAGREE.
: What happens here ? This is really a paradigm shift ! The subscribers
: of the old-age are unable to understand the subscibers of the new-age.
: It is a turning-point, IMO.
: That causes much of the problems here.
: A few months ago Bob called Chris a subsciber of an old-fashioned
: idea, somebody who lives and tries an idea that was given up by
: university-guys ages ago. I remember Ed said: if Chris is old
: fashinoned with his ideas, then my ideas are old-fashioned too,
: because I am doing it the same way.
I think you have your wires crossed. If you back up to crafty last
year, it had a *bunch* of selective things in it. I don't like 'em,
but I continue to experiment with them. However, Deep Blue is a
different animal, and doesn't have to resort to the selective search
stuff to reach reasonable depths...
: Hm. Ed has shown that his program plays very very strong. I mean: if
: they are old-fashioned or senseless, than this is a very nice way of
: doing a senseless idea.
: >I am not in the mood for this kind of explanations for another 9 games
: >the next 5-6 months. My time is too precious for that.
: >Rebel8 played this game on an average ply depth of 9/10.
: >Crafty played on an average ply depth of 13/14.
: >Still Rebel8's evaluation function was good enough to handle the big
: >ply gap of 4 plies. This has always been my point and game-1 has
: >proven it. I believe this also counts for other chess programs like
: >for instance Hiarcs etc.
: I subscibe Eds point of view. The ideas of chess-programming are
: somehow divided. It looks that the opponent sides will not be able to
: have a fruitful exchange of ideas, due to incompatible feelings, point
: of views, language and programming efforts.
Wait a minute. *what* "exchange of ideas" are we talking about here?
I'm thumbing thru my ICCA journals, my ACM SigArt journals, my IEEE
AI journals, and I don't find a *single thing* written by Ed, or by
any other commercial program, excepting perhaps Donniger. So there's
no "exchange of ideas" all right, but it's *not* for the reason you are
postulating. Notice that I've made all of my ideas public. It appears,
from my perspective, to be a one-way information flow. I suppose that
the very concept "exchange of ideas" deserves a <snicker> added to the
end...
: In my opinon chris could show them (FAST/STUPID-search fans) his
Torstein Hall
OH what I disapointment!!
The last days I logged on to the Rebel site to check how the match was
going whenever I had time, so now I'm very disapointed.
I can understand that you feel that it takes up to much valuable time, even
if I think the result of the match would have been very interesting to see
for everyone.
The "excuses" of RH I think you could have overlooked, I'm for one think
its normal to think of someting when in a bad position on the chesstable!
Torstein Hall
tors...@eunet.no
http://login.eunet.no/~torshall/sjakk.html
Torstein Hall
Just to mention it, I put the word excuses inside the " " because I agree
with you, but if E.S. feels it are excuses that somehow degrade the
information from the match, I just whished he overlooked them!
(and Craftys pos. was realy bad at the end! )
But perhaps you could make Crafty autoplay itself under the same conditions
( 100 to 1) and post it here, I think it would be interesting.
Torstein Hall
tors...@eunet.no
http://login.eunet.no/torshall/
Robert Hyatt <hy...@crafty.cis.uab.edu> skrev i artikkelen
<5hbrv5$d...@juniper.cis.uab.edu>...
> Torstein Hall (tors...@eunet.no) wrote:
> : OH what I disapointment!!
>
> : The last days I logged on to the Rebel site to check how the match was
> : going whenever I had time, so now I'm very disapointed.
>
> : I can understand that you feel that it takes up to much valuable time,
even
> : if I think the result of the match would have been very interesting to
see
> : for everyone.
>
> : The "excuses" of RH I think you could have overlooked, I'm for one
think
> : its normal to think of someting when in a bad position on the
chesstable!
>
> Didn't notice any "excuses" that I'd posted here. I posted some
> "explanations" about moves that others asked about, such as why not
> e4 rather than gxh5 for example.
>
> I also mentioned, well before we started, that the depth issue would be
> something interesting to watch. Ken Thompson, Hsu and I had this same
> discussion the first time Cray Blitz played ChipTest, because Ken had
> the same fear of the machine outsearching the opponent and seeing
> something it would like to avoid, not knowing that it really had no
> need to avoid it.
>
> When I get the log file for the one game, I'll show you what I mean in
> a clearly understood form. My explanation was not meant to take anything
> away from Rebel, which played nearly perfectly tactically it seems. It
> was meant to show that in some cases, deeper depth is not always enough.
>
> I'd certainly like to see more games, and will try to arrange something
> to make this happen, because I think the problem this highlights is both
> interesting and important. I'd be willing to bet that if you watch
Crafty
> play on ICC, that in at least one out of every 4 games it "throws the
game
> away" to avoid something that the human wouldn't see. That is a problem
> that needs investigation and a solution. In even computer vs computer
> games, it's not much of a problem, but when one machine is much faster,
> it is going to make an odd mistake here and there due to this problem.
>
> Of course, I also believe that the mistakes are offset a hundred-fold by
> the gains produced by the deeper search. But I'd like to have all of the
> depth gains, with *none* of the depth problems. That'd be a useful thing
> to develop, although with minimax/alpha/beta it might be difficult at
> best.
>
>
Robert Hyatt
Don Fong
In article <5hbrv5$d...@juniper.cis.uab.edu>,
Robert Hyatt <hy...@crafty.cis.uab.edu> wrote:
>When I get the log file for the one game, I'll show you what I mean in
>a clearly understood form.
i think most everyone already understands the concept.
i think you mean, you are searching "deeper" in some sense
than your opponent, you are considering many positions s/he doesn't.
at some point you discover that your opponent has a beautiful winning
line. your "best" defense against it looks lost, but not mated.
this "best" defense tips your hand and makes it "easy" for the opponent
to realize the advantage.
if you are playing a weaker opponent, you can gamble that s/he
won't see the winning moves. but it is still a gamble: you never
know whether your opponent may "luck into" them.
unless you know your opponent's mind to some extent, you cannot
realistically assign probabilities to the various possible lines.
of course, if you REALLY knew your opponent's mind, you could
double your search depth. (:-)
i think the real question is, how did your presumably weaker
opponent manage to put you in such a pickle to begin with? if you
are stronger, then why did it happen that YOU are on the wrong side
of the deep combination, and not your opponent?
the other real question (:-) is how to model your opponent's
mind - how to decide whether to take the gamble.
Kevin James Begley
Robert Hyatt (hy...@crafty.cis.uab.edu) wrote:
<SNIP>
: : For instance, Bob Hyatt wrote:
: : : I'll post some interesting analysis later, but you might be surprised
: : : at some of what went on in this match. For example. in at least two
: : : places, Crafty played a weaker move, simply because it saw, at
: : : depth=15, something that it thought Rebel would play and it didn't
: : : like it. In these two cases I know of, I've tested rebel on the
: : : position and it wouldn't have played the move Crafty feared. So
: : : Crafty simply avoided something that would not have happened, by
: : : playing something that was worse. It was one of the common scenarios
: : : I've seen when two programs play and one outsearches the other by
: : : a significant amount.
: I don't see anything misleading at all in the above. Nor do I see any
: "excuse" making either. The thing I pointed out above happened. It has
: happened many times in the past, it will continue to happen many times in
: the future. And it's a problem that is going to have to be addressed at
: some point in time.
Addressed? We're going to have to address ridiculous claims that a
player made an inferior move only because he was so much stronger to
realize that the superior move was worse?!? How is it then that the
superior player arrived in such a position!? Don't tell me it was
pure luck. Are you absolutely certain that your opponent had nothing
to do with it? If so, please explain how.
Chess is not a negociation. You play the best move you can find, period.
We certainly don't need to address claims of what might have been. Just
assume that you know nothing about your adversary, and let your machine's
superior ability avoid such situations.
Indeed, your claim that "fortune favors the foolish" is the most outlandish
excuse for losing I've ever heard.
: But, for the record, have you ever had Rebel play against a strong human,
: and saw Rebel throw the game away? Because it saw some deep tactic that
: it thought needed looking after, but which cost Rebel a pawn (or more) to
: defend against? A tactic that in reality, the human didn't see? If you
: haven't seen this, you simply haven't looked closely enough.
What's your point? Nobody knows how often this happens! I could just as
easily claim that I lost against a stronger player because I assumed he had
a refutation for the winning move that I failed to play. The result, I'm
afraid, is all that remains. Learn to live with the results.
: A year ago, I posted an ugly game played by Cray Blitz vs a 1900 player
: in the 1984 US Open speed chess championship. CB had gone through IM's
: and GM's like a knife through butter, and near the end of the first
: "section" match (round robin with 16 players I think) we hit a 1900
: player, and were doing o.k.
If you were "doing o.k." how do you explain getting into a situation where
you must sac your queen to avoid mate?!
: Out of the clear, CB sacrificed it's queen
: for a knight. Victor Korchnoi was looking over my shoulder and said
: "machine make terrible blunder" in his deep, Russian gutteral accent.
I love the way you put that -- "out of the clear." As if the bad position
your machine stumbled into had nothing to do with it.
: After the game, we backed up and I asked him what to play. He suggested
: something like RxR which appeared to be hanging. I tried it, and CB
: instantly announced a mate in 11 for the opponent. It saw that in the
: game and sacrificed the queen to avoid it. At the time, Korchnoi was the
: #3 player in the world I believe, and was Karpov's second at the time.
: He didn't see the mate. I doubt the 1900 player would have either.
Unfortunately, you will never know. Why don't you give us an algorithm
that will differentiate between the forced mates that an opponent will
see, and those he will certainly miss. You seem to have a good handle
on it yourself, so it ought to be easy for you to code it.
: Yet CB, because it just happened to search deeply enough, did see it,
: and threw the game away because of it. This was the only game it lost
: as I recall, also. That happened in round 1 of the NPS, although it was
: not a mate, nor even a material loss... That's not knocking rebel, it's
: not making excuses for Crafty losing. It is simply pointing out the fact
: that a deeper search doesn't always result in better play. In this game,
: Crafty played some poor moves to hang on to the passed E pawn, when it
: should not have. And this put it into positions where it eventually had
: to give up something somewhere. The deeper search, however, made it think
: it had to give up more than it really did.
Ha! That's the best line ever -- "...made it think it had to give up more
than it really did." Sounds like you're saying that a deeper search is
actually detrimental, since without it, you might have given up less!
Maybe you should put a governor on how deep your program can search when
it is considering giving up something!!
: Sorry you took it as offensive. It wasn't intended to be. I didn't try to
: make *any* excuses for Crafty. I didn't mention that in the 12 game
: "calibration match" that crafty had black in every game, which would likely
: affect the outcome by some amount. I didn't play up the fact that for the
: first 7 games, crafty was playing with a gross "contempt factor" because it
: thought Rebel had zero time left, and the draw score code sets contempt to
: -1.000 in that case to try and avoid drawing when the opponent is about to
: lose on time.
Super! I didn't make *any* excuses, though here are a whole bunch I might
have made... Sounds like something out of a poor comedy show!
: In any case, as I mentioned before, the match was on your "nickel" and you
: are certainly free to stop whenever you want. I knew it would take a long
: time to play out. The calibration match was having problems due to auto232
: quirks, but the one NPS game was simply clean so far as I'm concerned. The
: book might have affected the game, it might not. However, when Rebel left
: book, I didn't consider Crafty's position so inferior that the game should
: be thrown out. Rebel played well. I think the most impressive thing I saw
: was the fact that it did not seem to make *any* tactical mistakes, anywhere.
: Which is surprising. I'd have liked to see more games to see if this could
: continue.
Excellent analysis above! The opponent suprised you by playing remarkably
well, especially considering its more limited depth.
<SNIP>
: The only thing you say that I don't agree with is the above. One game doesn't
: prove anything. That was the intent for playing 10 games to start with,
: rathe than a single game.
One game is insufficient, I agree. But, one game does show that Rebel can
play remarkably well when suffering from a 4-ply handicap in search depth.
: However, in the interests of seeing how this works, I will see if I can set
: up a "correspondence handicap game" with someone that is willing to operate
: Rebel on the other end, just to see how this might continue. I can operate
: it manually here, but don't want to give anyone the chance to cry foul...
: It's still an experiment that needs doing, because it has never been done
: before. Everyone agrees that one ply is not a significant advantage at these
: depths. But we have absolutely no data about 4-5 plies. I hope you might
: consider doing rebel vs rebelX100, as that would generate the data.
Why not Crafty vs Crafty*100?
I might also suggest, though perhaps it was already, that
you might consider varrying the search depths (while keeping
the differnce in depth constant). For example, 1-ply versus
5-ply, I presume, would be a huge advantage, whereas 10-ply
versus 14-ply would be much less significant. Perhaps you
can approximate the curve, and make some predication about
what will happen at higher depths.
Kevin.
Robert Hyatt
: Robert Hyatt <hy...@crafty.cis.uab.edu> wrote:
: >
: > I'd be willing to bet that if you watch
: Crafty
: > play on ICC, that in at least one out of every 4 games it "throws the
: game
: > away" to avoid something that the human wouldn't see. That is a problem
: > that needs investigation and a solution. In even computer vs computer
: > games, it's not much of a problem, but when one machine is much faster,
: > it is going to make an odd mistake here and there due to this problem.
: >
: >
I was suffering from cosmic rays when I wrote that. :) What I meant,
was that for every 4 games Crafty loses to humans on ICC, one of those
games is lost for no reason, because Crafty saw something the human didn't
see, and avoided it as only a computer can, generally by tossing material
to the 4 winds. Losses to humans on ICC is *rare*. So this is even
rarer... but it's important that it be stopped one day...
Robert Hyatt
Don Fong (df...@cse.ucsc.edu) wrote:
: In article <5hbrv5$d...@juniper.cis.uab.edu>,
Rebel played very well. However, in fairness with this position, I
also tested with genius as well, and the funny part is, from the white
side, given even fairly short times to move, Rebel, Genius *and* Crafty
would follow the same path. White simply ended up with a good position
after the first 10-15 moves, and the rest of the white plan almost played
itself. Of course, as you point out, white had to reach that position,
and black had to allow it. Black was thinking a little too far toward
the endgame, trying to preserve a winning advantage, but, as a result,
didn't make it through the middlegame unscathed.
: the other real question (:-) is how to model your opponent's
: mind - how to decide whether to take the gamble.
Yep. and as several AI well-knowns have said over the years, this is a
crucial step if minimax is going to work. Because at present, minimax doesn't
work exactly like we do in this regard. If you start off with a good position,
the extra depth should guarantee a win. If you start off with an inferior
position (as black does, often enough) then the extra depth might not help,
unless you can figure out what your opponent sees and doesn't see...
as far as your "gamble" comment above, computers are a little more "obvious"
than that. For example, it plays g4 and the pawn is hanging. A human might
wonder "now why does it want my bishop off of that diagonal so badly?" and he
will probably figure out why. And see something he had no inkling was there
before the computer made a fishy move.
Berliner played with a fix for this, but it wasn't good enough, in that it
only tried to fix something that was seen in the very last search iteration.
If you see something bad for 2 or more, it would assume it was obvious and
play the "best" move (which was often an obviously losing move...)
Robert Hyatt
Kevin James Begley (kjbe...@chimi.engr.ucdavis.edu) wrote:
: <SNIP>
Where exactly did you see that "fortune favors the foolish" claim? Your statement
is about the most outlandish interpretation of anything I've seen...
: : But, for the record, have you ever had Rebel play against a strong human,
: : and saw Rebel throw the game away? Because it saw some deep tactic that
: : it thought needed looking after, but which cost Rebel a pawn (or more) to
: : defend against? A tactic that in reality, the human didn't see? If you
: : haven't seen this, you simply haven't looked closely enough.
: What's your point? Nobody knows how often this happens! I could just as
: easily claim that I lost against a stronger player because I assumed he had
: a refutation for the winning move that I failed to play. The result, I'm
: afraid, is all that remains. Learn to live with the results.
I *know* how often it happens, because I have only 80,000 games played on the
Servers. I have over 10,000 games of Crafty vs GM players. I don't go through
every one, but if you'd like I'll find a recent log and show you exactly what
I'm talking about.
: : A year ago, I posted an ugly game played by Cray Blitz vs a 1900 player
: : in the 1984 US Open speed chess championship. CB had gone through IM's
: : and GM's like a knife through butter, and near the end of the first
: : "section" match (round robin with 16 players I think) we hit a 1900
: : player, and were doing o.k.
: If you were "doing o.k." how do you explain getting into a situation where
: you must sac your queen to avoid mate?!
It's called a tactically open position. You enter into a wild tactical
melee, and make a move after a 9 ply search. one ply later you notice that
there is a *very* deep stinger you overlooked before. Obviously it is deep
because the 9 ply search didn't pick it up, but the 10 ply search did. It
was attacking on the kingside, and realized at the end there was one quiet
move that undid everything, about 10 moves down the game tree. And as soon
as it noticed it, it did everything in its power to avoid getting mated.
Have you ever attacked your opponent only to lose yourself? How'd you get
into that position? It happens. Particularly when pieces are hanging all
over the board in a wild attack.
: : Out of the clear, CB sacrificed it's queen
: : for a knight. Victor Korchnoi was looking over my shoulder and said
: : "machine make terrible blunder" in his deep, Russian gutteral accent.
: I love the way you put that -- "out of the clear." As if the bad position
: your machine stumbled into had nothing to do with it.
I love the way you take things. Again, a wild tactical position, not a positional
struggle, but one with pieces attacked, open files, zillions of checks, and it
found what it thought was a brilliant path through all of that that won some sort
of material, yet one ply deeper it saw the truth and tried to evade it... So it
wasn't a bad "position" at all, just a tactical mistake. Guess that's too hard
to grasp???
: : After the game, we backed up and I asked him what to play. He suggested
: : something like RxR which appeared to be hanging. I tried it, and CB
: : instantly announced a mate in 11 for the opponent. It saw that in the
: : game and sacrificed the queen to avoid it. At the time, Korchnoi was the
: : #3 player in the world I believe, and was Karpov's second at the time.
: : He didn't see the mate. I doubt the 1900 player would have either.
: Unfortunately, you will never know. Why don't you give us an algorithm
: that will differentiate between the forced mates that an opponent will
: see, and those he will certainly miss. You seem to have a good handle
: on it yourself, so it ought to be easy for you to code it.
Sorry, but I do know. He asked about why in the world did CB sac the
queen. He thought he was getting mated in fact, and was very skeptical
of taking the queen.
: : Yet CB, because it just happened to search deeply enough, did see it,
: : and threw the game away because of it. This was the only game it lost
: : as I recall, also. That happened in round 1 of the NPS, although it was
: : not a mate, nor even a material loss... That's not knocking rebel, it's
: : not making excuses for Crafty losing. It is simply pointing out the fact
: : that a deeper search doesn't always result in better play. In this game,
: : Crafty played some poor moves to hang on to the passed E pawn, when it
: : should not have. And this put it into positions where it eventually had
: : to give up something somewhere. The deeper search, however, made it think
: : it had to give up more than it really did.
: Ha! That's the best line ever -- "...made it think it had to give up more
: than it really did." Sounds like you're saying that a deeper search is
: actually detrimental, since without it, you might have given up less!
: Maybe you should put a governor on how deep your program can search when
: it is considering giving up something!!
No, but if you'd simply read to understand, rather than reading to nit-pick, the
problem is obvious. I didn't say a deeper search is worse. I said a deeper search
can *sometimes* (can you see the emphasis here) result in playing a worse move.
Simple statement. Easy to prove...
: : Sorry you took it as offensive. It wasn't intended to be. I didn't try to
: : make *any* excuses for Crafty. I didn't mention that in the 12 game
: : "calibration match" that crafty had black in every game, which would likely
: : affect the outcome by some amount. I didn't play up the fact that for the
: : first 7 games, crafty was playing with a gross "contempt factor" because it
: : thought Rebel had zero time left, and the draw score code sets contempt to
: : -1.000 in that case to try and avoid drawing when the opponent is about to
: : lose on time.
: Super! I didn't make *any* excuses, though here are a whole bunch I might
: have made... Sounds like something out of a poor comedy show!
Nope. Ed posted the excuses comment. I simply pointed out that I watched as
Crafty went 9-3 vs Rebel and didn't say a word about any of the above. The
above *are* excuses. They were not posted *until* after Ed chose to call off
the match.
: : In any case, as I mentioned before, the match was on your "nickel" and you
: : are certainly free to stop whenever you want. I knew it would take a long
: : time to play out. The calibration match was having problems due to auto232
: : quirks, but the one NPS game was simply clean so far as I'm concerned. The
: : book might have affected the game, it might not. However, when Rebel left
: : book, I didn't consider Crafty's position so inferior that the game should
: : be thrown out. Rebel played well. I think the most impressive thing I saw
: : was the fact that it did not seem to make *any* tactical mistakes, anywhere.
: : Which is surprising. I'd have liked to see more games to see if this could
: : continue.
: Excellent analysis above! The opponent suprised you by playing remarkably
: well, especially considering its more limited depth.
I never said otherwise, regardless of what you'd like to imply...
: <SNIP>
: : The only thing you say that I don't agree with is the above. One game doesn't
: : prove anything. That was the intent for playing 10 games to start with,
: : rathe than a single game.
: One game is insufficient, I agree. But, one game does show that Rebel can
: play remarkably well when suffering from a 4-ply handicap in search depth.
yes it does.
: : However, in the interests of seeing how this works, I will see if I can set
: : up a "correspondence handicap game" with someone that is willing to operate
: : Rebel on the other end, just to see how this might continue. I can operate
: : it manually here, but don't want to give anyone the chance to cry foul...
: : It's still an experiment that needs doing, because it has never been done
: : before. Everyone agrees that one ply is not a significant advantage at these
: : depths. But we have absolutely no data about 4-5 plies. I hope you might
: : consider doing rebel vs rebelX100, as that would generate the data.
: Why not Crafty vs Crafty*100?
: I might also suggest, though perhaps it was already, that
: you might consider varrying the search depths (while keeping
: the differnce in depth constant). For example, 1-ply versus
: 5-ply, I presume, would be a huge advantage, whereas 10-ply
: versus 14-ply would be much less significant. Perhaps you
: can approximate the curve, and make some predication about
: what will happen at higher depths.
: Kevin.
I don't think there's a point to doing Crafty vs Crafty*100... because we'd only
fill up the newsgroup with endless discussion about "but that's crafty vs crafty
with a handicap, that's doesn't say a thing about Rebel vs Deep Blue or whatever."
As with Ed, I try to use at least some of my time wisely...
chrisw
--
http://www.demon.co.uk/oxford-soft
Robert Hyatt <hy...@crafty.cis.uab.edu> wrote in article
<5hbld2$b...@juniper.cis.uab.edu>...
> Ed Schroder (rebc...@xs4all.nl) wrote:
> : NPS challenge will end after game one...
>
> : Today I decided to end the NPS challenge between Crafty and Rebel8
after
> : game-1 is finished.
>
> : Understanding the problems of running a match with an unfinished
> : product, and since there are so many things that Bob may claim and so
> : many things that did go wrong with the book, evals, time control,
> : contempt factor, huge auto232 problems, bugs etc. I consider it a waste
> : of my time to continue the match.
>
> : Besides of this I also want to express my feelings about the following.
>
> : I am interested in the scientific relationship between ply depth and an
> : evaluation function. That's what has started this match. Sofar game-1
> : has proven my point of view at least in my opinion.
>
> While I won't argue about game one, one game is not enough. If you
> still feel strongly about ply depth, try the alternative experiment
> I suggested, which is Rebel vs Rebel at 100:1 time odds. That will
> provide better data than Rebel vs Crafty, because the *only* thing
> different will be the depth. No differences in eval, no auto232
> problems, etc. And the results will be just as interesting...
>
This is a total red herring. The point of the match was to consider
evaluation v. search depth. That can't be done on Rebel vs. Rebel at 100 to
1 time odds.
No it wouldn't. See above.
Chris Whittington
>
>
Kevin James Begley
Don Fong (df...@cse.ucsc.edu) wrote:
<SNIP>
: i think the real question is, how did your presumably weaker
: opponent manage to put you in such a pickle to begin with? if you
: are stronger, then why did it happen that YOU are on the wrong side
: of the deep combination, and not your opponent?
I would wager it had SOMETHING to do with a superior evaluation
function. Deep search = superior tactics, deep evaluation = superior
position. Certainly, Rebel didn't forsee all the tactics, but it
did manage to obtain a better position with its evaluation.
Perhaps what is needed is a mesh of the two styles -- deep evaluation
to determine good positional candidates in non-tactical situations,
and deep search to verify that some tactic is not missed along the
way. It seems to be a question of which style is in greater demand
today, and after Kasparov vs. Deep-Blue, I would think better eval
would be the obvious answer.
Kevin.
Jack Nerad
Don Fong wrote:
>
> In article <5hbrv5$d...@juniper.cis.uab.edu>,
> Robert Hyatt <hy...@crafty.cis.uab.edu> wrote:
> >When I get the log file for the one game, I'll show you what I mean in
> >a clearly understood form.
>
> i think most everyone already understands the concept.
> i think you mean, you are searching "deeper" in some sense
> than your opponent, you are considering many positions s/he doesn't.
> at some point you discover that your opponent has a beautiful winning
> line. your "best" defense against it looks lost, but not mated.
> this "best" defense tips your hand and makes it "easy" for the opponent
> to realize the advantage.
> if you are playing a weaker opponent, you can gamble that s/he
> won't see the winning moves. but it is still a gamble: you never
> know whether your opponent may "luck into" them.
>
>
> mind - how to decide whether to take the gamble.
I must be missing something here. If the choice is between a move that
loses a vast amount of material and a move that will allow checkmate
somewhere down the line, why not assume that the more immediate material
loss will be more harmful? I am a weak player, I guess, but if I have
the opportunity to take someone's queen or rook apparently to my
advantage, I will take that instead of going for a hazy checkmate
somewhere on the horizon.
If a program is searching 16 plies and sees a checkmate in one of the
branches, or a material loss that is significant, why not merely resign
the game? Or accept the checkmate? Why is it that computers will
accept the material loss that will (given future simplification) lead to
a lost position anyway. instead of resigning?
Regards,
Jack Nerad
brucemo
Ed Schroder wrote:
>
> NPS challenge will end after game one...
>
> Today I decided to end the NPS challenge between Crafty and Rebel8 after
> game-1 is finished.
A shame, especially since it ends in your favor after one game. When you
start a ten-game match, you should play a ten-game match, and not play a
one-game match.
> Understanding the problems of running a match with an unfinished
> product, and since there are so many things that Bob may claim and so
> many things that did go wrong with the book, evals, time control,
> contempt factor, huge auto232 problems, bugs etc. I consider it a waste
> of my time to continue the match.
Both of you guys spent a long time working on this. I got on the CC line
of the emails going back and forth, and there were a lot of them, from each
of you. I have sixty or seventy email messages on this topic from Bob
alone.
These were mostly about bugs in Crafty that surfaced due to the use of
Auto232, and because Crafty doesn't have a debugged install procedure.
Most of the auto232 bugs seem to have been in the "control" match.
On ICC, Crafty is started, plays a game, then is terminated. With auto232,
it plays a game, then is told to start a new game, etc. I just tried to
call Bob to ask him for sure, but he wasn't home. But the bugs that I know
of were, 1) after the first game it thought the opponent was out of time,
so it changed its contempt factor to -1.00 pawns, 2) for each successive
game after the first game, it used less and less time per move in the
secondary time control. It sounds like there were other weird things going
on as well.
I think you were very patient to put up with the auto232 bugs. I am not
sure what caused your patience to run out. Aren't the problems solved?
I asked Bob what was going on when I noticed that the score in the control
match was 6-0, and he told me about the secondary time control problem and
the contempt factor problem.
It is obvious that this could have had an affect upon some games. I asked
Bob if he knew that it DID have an affect upon any of them, and he told me
it looked like Crafty made at least one really stupid move, due to low
time, in an otherwise even position. He may have mentioned another game, I
don't know.
But it wasn't like Bob was whining about this, I had to ask. I didn't see
any mail from Bob that I was CC'd on where he suggested ignoring those six
games. If you ignore them, they can be replayed in a couple of days. If
you keep them, big deal, it's just six games out of 60, and maybe you would
score 4-2 or 5-1 or 6-0 again if you replayed them, so they aren't very
meaningful.
After various bugs were fixed, Rebel and Crafty split the next six games,
3-3.
Crafty blew out of book early in the NPS game because its book was
write-protected on the CD, but you guys both agreed to continue, so who is
to complain?
> Besides of this I also want to express my feelings about the following.
>
> I am interested in the scientific relationship between ply depth and an
> evaluation function. That's what has started this match. Sofar game-1
> has proven my point of view at least in my opinion.
You can't claim victory after one game in a ten-game match. I assume there
was a reason you guys agreed to play a ten-game match. Perhaps Bob would
not have agreed to a one-game match, with black, but that isn't what you
guys agreed to.
Declaring victory and pulling the plug after one game is really bad,
especially after playing white in every control game as well.
> I am *not* interested to read all kind of possible and impossible
> explanations that imply to hide the real reason why Crafty will lose
> game-1.
Crafty lost game one because it didn't play well. Rebel won game one
because it played well.
There can be speculation about why this happened. Some of it may revolve
around mistakes Bob made, rather than your own super-competence.
I think Bob made a serious mistake with his king safety routine, which he
should fix. There is probably more to it than this, certainly, and I'm
sure a lot of the credit goes to Rebel for playing the position well.
Perhaps you could explain WHY it played so well?
With out without this fix, I would like to see game two, especially if
Crafty gets to play white.
> For instance, Bob Hyatt wrote:
>
> : I'll post some interesting analysis later, but you might be surprised
> : at some of what went on in this match. For example. in at least two
> : places, Crafty played a weaker move, simply because it saw, at
> : depth=15, something that it thought Rebel would play and it didn't
> : like it. In these two cases I know of, I've tested rebel on the
> : position and it wouldn't have played the move Crafty feared. So
> : Crafty simply avoided something that would not have happened, by
> : playing something that was worse. It was one of the common scenarios
> : I've seen when two programs play and one outsearches the other by
> : a significant amount.
This isn't an excuse, it is analysis. I'm not sure if he's right or not,
but the idea that a program may kill itself via its own search depth is
something that Bob has been interested in.
Bob of all people has to be able to comment honestly on the games, even if
this may involve questioning the idea that Rebel is the best program on the
planet.
> This would mean Crafty / Rebel would come into trouble against the
> Mephisto I, Boris etc. because Crafty / Rebel think 4-5 plies deeper.
> This is of course total nonsense. Crafty as any other program just
> played the best move on a given depth.
On ICC I played a rated 5 0 game, Ferret vs Spinach. Spinach is Ferret at
50 milliseconds per move, no pondering, on inferior hardware. So this is
more than 100:1 time odds.
I only did this once, as it really scared me to watch Ferret play at -2 for
most of the game. It found a tactic later on and won, but I'd have been
really pissed if it had dropped 32 rating points because of this little
lark.
Why do you think Spinach was up two pawns against Ferret? Do you think it
out-searched it at 100:1 time odds? No, Ferret came out of book bad,
couldn't negotiate the position, found a combination against for the
opponent, and mitigated the problem by dumping material at the root.
That this may have happened doesn't take anything away from Rebel's win in
round one, of ten. As I have stated, Rebel played well and Crafty didn't.
It doesn't do a program credit to get into a situation where it is finding
combinations against itself, assuming that is what happened here.
> I wouldn't dare to mention such an argument if Rebel would have had
> the 100 x time advantage!
It's not a particularly big deal. Self-criticism after losses doesn't take
anything away from the opponent. "He beat me because he's god", doesn't
offer the loser much hope.
> This is just one example...
>
> I am not in the mood for this kind of explanations for another 9 games
> the next 5-6 months. My time is too precious for that.
>
> Rebel8 played this game on an average ply depth of 9/10.
> Crafty played on an average ply depth of 13/14.
>
> Still Rebel8's evaluation function was good enough to handle the big
> ply gap of 4 plies. This has always been my point and game-1 has
> proven it. I believe this also counts for other chess programs like
> for instance Hiarcs etc.
Rather than declaring the match over because you won game one, please come
back and play the rest of the match.
bruce
brucemo
Robert Hyatt wrote:
> However, in the interests of seeing how this works, I will see if I can set
> up a "correspondence handicap game" with someone that is willing to operate
> Rebel on the other end, just to see how this might continue. I can operate
> it manually here, but don't want to give anyone the chance to cry foul...
I have a legal copy of Rebel 8 if you can't find someone who is more
independent or otherwise better suited. I also have Genius 5, Mchess 6, and
Nimzo 3, although I haven't installed them yet, if want to try something
other than Rebel.
By the way, thanks Bert, my order got here in one piece, and quickly too :-)
The machine I have it on is a P5/133 with 32 megabytes, and it's running
Windows '95, so perhaps there'd need to be some additional handicapping.
bruce
brucemo
Herbert Groot Jebbink wrote:
>
> brucemo <bru...@nwlink.com> wrote:
>
> >Ed Schroder wrote:
>
> >> Today I decided to end the NPS challenge between Crafty and Rebel8 after
> >> game-1 is finished.
>
> >A shame, especially since it ends in your favor after one game.
> >Both of you guys spent a long time working on this. I got on the CC line
> >of the emails going back and forth, and there were a lot of them, from each
> >of you. I have sixty or seventy email messages on this topic from Bob
> >alone.
>
>
> Ed has not the time to for that much problems, he must write Rebel 9
> to beat Hiarcs 6 :-)
Of course. It involved a lot of time to get Crafty + auto232 working.
But, in the end, it was working, wasn't it?
bruce
Pete Nielsen
--
Re: "Crafty played a weaker move because ... "
Excuse my ignorance, but this has been puzzeling me ever since it came up
in the DB commentary.
Doesn't the AB search guarentee that ANY move that the computer selects is
at least as good as one it previously considered but now rejects?
brucemo
Robert Hyatt wrote:
> I can be convinced that those extra plies don't make any difference, but
> I have not been yet. I remember Lonnie, last year, posting queries here
> about what machine to buy, because all of a sudden, Crafty and Ferret were
> thrashing him badly because we'd move to the P6/200 and he was still on a
> P5/90 or P5/133, whatever. That was only a factor of 3-4 and it had him
> running around in circles to close the computational gap. Now I'm supposed
> to believe that after one game, we know that if I had a P6/20 gigahergz machine,
> I'd not do any better against the commercial guys than I do now? I wonder
> why I have such a hard time swallowing that?
"Blitz" on ICC is a very vague idea.
3 0 is blitz. This game will be over in 6 minutes of non-lag time, tops.
2 19 is blitz. A 60-move game might take 42 minutes, and a lot of games are longer
than 60 moves.
I do not know what happens in 2 19, Lonnie is free to say if he wishes, but I think
I remember Lonnie being crazy about getting a P6.
Another example might be WChess (the only automatic professional), which is just
death on equal hardware at 5 0, but which seems a lot less imposing with a hardware
advantage.
I like 5 0, the game fits inside my attention span, but the commercial programs on
ICC can't do it, because the human operators would overheat and melt.
I just did a quick match between Ferret and a leading commercial program, in 5 0
blitz, with Ferret on a P6/200 and the commercial program on a P5/133, which in my
opinion is a monstrous hardware advantage for Ferret, perhaps 3.5:1. I didn't use a
clock, but instead let both of them keep time themselves, and I tried to operate
fairly, with no monkey business (I even let the opponent play white half the time),
and the score was +6 -0 =4, with no games that were even close to being duplicates.
I don't know what happens on equal hardware, since I only have unequal machines,
and in any case I have NT on the P6, which limits my choices of commercial software
on that machine.
So yeah, the P6 is a great machine, and a little bit of hardware advantage goes a
long way, at least in blitz.
The question is, do you get this kind of effect in a standard game? I don't know
what the answer is.
bruce
Jay Scott
In article <5hbh7j$s91$1...@news1.xs4all.nl>, Ed Schroder (rebc...@xs4all.nl) wrote:
>Today I decided to end the NPS challenge between Crafty and Rebel8 after
>game-1 is finished.
It's such an expensive experiment, in computer time, that
I can understand stopping it more easily than I can understand
starting it. :-) But it's a shame that we won't find out the answers.
>Still Rebel8's evaluation function was good enough to handle the big
>ply gap of 4 plies. This has always been my point and game-1 has
>proven it.
Alas, one game proves nothing that we did not already know.
Even ten games are barely enough to draw any inference from the
results (though we might learn a lot from analyzing the games).
Jay Scott <j...@forum.swarthmore.edu>
Machine Learning in Games:
http://forum.swarthmore.edu/~jay/learn-game/index.html
Robert Hyatt
chrisw (chr...@cpsoft.demon.co.uk) wrote:
: --
: http://www.demon.co.uk/oxford-soft
Not at all. The question is, once a program reaches some depth the author
thinks is acceptable, will another 3-4-5 plies make a substantial difference
in the outcome? I believe the answer is yes. The reason I posed this
theory in the first place was in comparing Deep Blue to the current flock
of Micro programs.
If we had a Rebel vs Rebel*100 match, and the Rebel*100 wins by a wide
margin, what would be the cause? All it could possibly be would be the
extra depth. I agree that one ply is not a huge difference. Maybe even
two plies is not. But 4-5 is a significant handicap...
What better way to see if 3-5 plies makes a difference, than in a match
where those 3-5 plies is the *only* difference. I never considered this
a match between smart and fast. I considered this a match between slow
and fast. There's nothing to say that DB's eval is not every bit as good
as anything Rebel or anyone else is doing. But they are doing it so very
very fast...
Robert Hyatt
Jack Nerad (JNE...@concentric.net) wrote:
: Don Fong wrote:
: >
: > In article <5hbrv5$d...@juniper.cis.uab.edu>,
: > Robert Hyatt <hy...@crafty.cis.uab.edu> wrote:
: > >When I get the log file for the one game, I'll show you what I mean in
: > >a clearly understood form.
: >
: > i think most everyone already understands the concept.
: > i think you mean, you are searching "deeper" in some sense
: > than your opponent, you are considering many positions s/he doesn't.
: > at some point you discover that your opponent has a beautiful winning
: > line. your "best" defense against it looks lost, but not mated.
: > this "best" defense tips your hand and makes it "easy" for the opponent
: > to realize the advantage.
: > if you are playing a weaker opponent, you can gamble that s/he
: > won't see the winning moves. but it is still a gamble: you never
: > know whether your opponent may "luck into" them.
: >
: >
: > the other real question (:-) is how to model your opponent's
: > mind - how to decide whether to take the gamble.
: I must be missing something here. If the choice is between a move that
: loses a vast amount of material and a move that will allow checkmate
: somewhere down the line, why not assume that the more immediate material
: loss will be more harmful? I am a weak player, I guess, but if I have
: the opportunity to take someone's queen or rook apparently to my
: advantage, I will take that instead of going for a hazy checkmate
: somewhere on the horizon.
This is almost backward. The issue is this: suppose you search very
deeply and find a position you like and you believe you can reach. And
with best play you are going to stand better at the end. Then, just before
you move the piece, you decide to go over it one more time, and this time,
way out at the end, you find a really obscure move your opponent can make
that breaks the position wide open. Now you have two choices: (1) go ahead
and play the move you were going to play, not knowing if your opponent will
find the winning move or not since it is so deep in the tree; or (2) you
now start trying to find ways to thwart that move your opponent "might" play,
but since you are a computer, you assume he will see it, and the only way to
avoid getting busted completely is to give up two pawns.
As a human, I'd think that losing two pawns is probably losing the game,
unless there is some sort of compensation. But if I play the move I had
planned on, there's a good chance my opponent won't see the killer move
until it is too late to play it. I go for the latter. The computer goes
for the former... even if the human has no chance in Hades of seeing the
move (as the mate in 11 in blitz game I mentioned.)
An interesting problem...
: If a program is searching 16 plies and sees a checkmate in one of the
Robert Hyatt
Kevin James Begley (kjbe...@chimi.engr.ucdavis.edu) wrote:
: Robert Hyatt (hy...@crafty.cis.uab.edu) wrote:
: : : Chess is not a negociation. You play the best move you can find, period.
: : : We certainly don't need to address claims of what might have been. Just
: : : assume that you know nothing about your adversary, and let your machine's
: : : superior ability avoid such situations.
: : : Indeed, your claim that "fortune favors the foolish" is the most outlandish
: : : excuse for losing I've ever heard.
: : Where exactly did you see that "fortune favors the foolish" claim? Your
: : statement
: : is about the most outlandish interpretation of anything I've seen...
: Right you are -- guess I just like the sound of that saying. Still, I don't
: forsee any good way to address what might have been, and I doubt you (or any
: other experts) do either.
: Thus, I suggest that we ignore the hypotheticals, and stick with analyzing
: the results. Unfortunately, Ed has pulled-out before we have enough results
: to analyze (and thus, I suppose I understand why you must qualify what took
: place in the single game that Ed is willing to complete in). If he is going
: to leave us with this single game to analyze, perhaps we must look carefully
: into why Crafty lost.
: I must say, I don't think it is very sportsman-like for Ed to agree to a
: 10-game match, only to pull out after winning the first game. He may have
: proven to some that Rebel's eval is suprisingly strong, but in pulling-out,
: I feel he's denied Crafty the oppurtunity to make Bob's point (that such a
: superior depth will overpower the shallower evaluation).
Don't forget that Ed is selling a program. And testing. And working on a new
version that everyone will be chomping at the bit to buy later this year. He
does deserve the right to stop something that is really going to take a long
time to complete. How long was game 1 in progress? And it's a long way from
being over...
We might figure out a way to continue this in a manual correspondence-type
match like the KK Kup, with time odds...
As I've said, on several occasions, (1) I don't make excuses for mistakes that
Crafty makes, any more than I jump up and down when I log on to ICC right now and
find it rated at 2850+. (2) the data provided by this is really worthwhile, since
we don't get to analyze deep blue's output here, Crafty at huge time control
limits is a "sorta-approximation" for fun. It'd be nice to study what Crafty saw,
compare it to what rebel saw, and then do some analysis to see what the extra
depth added to or detracted from the game.
I'd like to know what those extra plies are worth, in reality, because all I
have to go on at present is just "intellectual reckoning" which might or might
not be correct. Hard data is easy to evaluate. Otherwise, we have nothing but
speculation...
I'll try to produce some food for thought and post it. At least it might give
us a *little* insight into what DB is capable of.
One point that Hsu's going to shoot me about... in the Crafty vs CM5000 game
in the Kup, Hsu said that Crafty was going to like the outcome of that sac
*far* more than we knew at the time (we'd barely gotten to the +2 point a few
moves later). He used "baby blue" and found a most impressive variation that
I hope he'll post here one day... to show just what his "baby" machine can
do. And he wasn't giving his "baby blue" machine an hour to search, either,
which is frightening for several reasons...
: Ed, in the interest of good sportsmanship, I suggest you reconsider resigning
: the remaining 9-games. If not for your own sake, do it for RJFischer, who
: must have spent the last 20+ years with profound regrets for not giving
: Karpov the oppurtunity he once had. Do it for Bobby, 'cause if he weren't
: so old, so crazy, and so cowardly, by God, he'd do it for you!
: <SNIP>
: : : What's your point? Nobody knows how often this happens! I could just as
: : : easily claim that I lost against a stronger player because I assumed he had
: : : a refutation for the winning move that I failed to play. The result, I'm
: : : afraid, is all that remains. Learn to live with the results.
: : I *know* how often it happens, because I have only 80,000 games played on the
: : Servers. I have over 10,000 games of Crafty vs GM players. I don't go
: : through
: : every one, but if you'd like I'll find a recent log and show you exactly what
: : I'm talking about.
: Don't bother, Bob. I know exactly what you are talking about
: technically, but I wonder if you appreciate how difficult it would be to
: find every instance in which one player made a concession (regardless of
: how minute) because it feared a deep response possibly unseen by the
: opponent. Surely, you realize this is approaching the limits of impossible,
: so how can we even begin to make sense of the astronomical "what-if's" that
: lie hidden in a chess game?
No... I agree with you. I've been a long-term practitioner of alpha/beta,
so long that it's second nature. And I've seen this problem more times than
I'd care to mention. And it was pointed out to me several times at various
ACM events by the likes of Michie and others.
However, this issue has to be addressed at some point, because otherwise a
machine that is impossibly strong is still going to lose a game here and there
to a rank beginner, something that simply doesn't happen often enough to a GM
to even consider. The computer has got to overcome this...
But with our old friend minimax, it's going to be hard...
: : : : A year ago, I posted an ugly game played by Cray Blitz vs a 1900 player
: : : : in the 1984 US Open speed chess championship. CB had gone through IM's
: : : : and GM's like a knife through butter, and near the end of the first
: : : : "section" match (round robin with 16 players I think) we hit a 1900
: : : : player, and were doing o.k.
: : : If you were "doing o.k." how do you explain getting into a situation where
: : : you must sac your queen to avoid mate?!
: : It's called a tactically open position. You enter into a wild tactical
: : melee, and make a move after a 9 ply search. one ply later you notice that
: : there is a *very* deep stinger you overlooked before. Obviously it is deep
: : because the 9 ply search didn't pick it up, but the 10 ply search did. It
: : was attacking on the kingside, and realized at the end there was one quiet
: : move that undid everything, about 10 moves down the game tree. And as soon
: : as it noticed it, it did everything in its power to avoid getting mated.
: : Have you ever attacked your opponent only to lose yourself? How'd you get
: : into that position? It happens. Particularly when pieces are hanging all
: : over the board in a wild attack.
: Well, the question remains -- how much of this was your own misfortune, and
: how much of this had something to do with how well the 1900-player had been
: playing? How much of Rebel's success can be attributed to it's algorithms
: (especially evaluation function), and how much can we conclude was pure luck?
: I don't think you profess to know -- I certainly can't imagine how you could.
In the 1900 case, it was all Cray Blitz. Nice position, and it unleashes
one of it's famous 25 ply combinations. And then discovered that at about
ply=10, there was a quiet move that broke the combination. Then it had to
scramble to avoid getting mated...
: : : : Out of the clear, CB sacrificed it's queen
: : : : for a knight. Victor Korchnoi was looking over my shoulder and said
: : : : "machine make terrible blunder" in his deep, Russian gutteral accent.
: : : I love the way you put that -- "out of the clear." As if the bad position
: : : your machine stumbled into had nothing to do with it.
: : I love the way you take things. Again, a wild tactical position, not a
: : positional
: : struggle, but one with pieces attacked, open files, zillions of checks, and it
: : found what it thought was a brilliant path through all of that that won some
: : sort
: : of material, yet one ply deeper it saw the truth and tried to evade it...
: : So it
: : wasn't a bad "position" at all, just a tactical mistake. Guess that's too
: : hard to grasp???
: I grasp what you're saying -- simple case of the horizon effect catching you
: in a tactically open position. Still, there's another school of thought --
: one that would claim that you left yourself in a bad position (I said nothing
: of positional struggle, just bad position). Their method of avoiding the
: horizon effect is to better understand how you were positionally weak, and
: by adding this knowledge into their evaluation, they hope to avoid such a
: mistake. I am quite sure you grasp that. So, when you say things like "out
: of the clear" you must understand why I consider that funny -- the other
: school of thought maintains that NOTHING COMES FROM OUT OF THE CLEAR (or
: stated differently, SUPERIOR TACTICS COME FROM A SUPERIOR POSITION).
yes... I agree. But in many of these positions, if you'd look at the
position not knowing who was playing, you'd say "white is winning" and then
discover that white is the computer... and that it simply found some deep
deep tactic that makes a sound position (on the surface) suddenly fall apart...
Not a regular occurence... but it happens too often to dismiss either...
I'll watch for some good examples on ICC. Crafty's been playing IM and GM players
all day today, currently has played over 100 games in last 24 hours against this
class of players... and has lost 3-4 of them. One of those losses probably will
illustrate this... good position, start a combination, then find that it fails
with some deep threat by opponent, bail out of combination and accept material
loss. I had an IM ask me about a game two days ago where this happened. He
couldn't figure out why it didn't recapture a knight. The log showed all and
he was amazed... never saw it... even after looking at it later. Unfortunately
I don't save these logs as there are too many... But I'll grab one that is
interesting, as the problem is interesting...
: : : : After the game, we backed up and I asked him what to play. He suggested
: : : : something like RxR which appeared to be hanging. I tried it, and CB
: : : : instantly announced a mate in 11 for the opponent. It saw that in the
: : : : game and sacrificed the queen to avoid it. At the time, Korchnoi was the
: : : : #3 player in the world I believe, and was Karpov's second at the time.
: : : : He didn't see the mate. I doubt the 1900 player would have either.
: : : Unfortunately, you will never know. Why don't you give us an algorithm
: : : that will differentiate between the forced mates that an opponent will
: : : see, and those he will certainly miss. You seem to have a good handle
: : : on it yourself, so it ought to be easy for you to code it.
: : Sorry, but I do know. He asked about why in the world did CB sac the
: : queen. He thought he was getting mated in fact, and was very skeptical
: : of taking the queen.
: Quite possibly true. It may even be true that, had he stared at a potential
: mate in eleven, he would not have found it under the game conditions (though
: we can never test this for certain). Perhaps he obtained a winning position
: purely out of luck, and perhaps he would have lost (even from having a forced
: mate) had your machine played an inferior move. I don't know if I'd be
: willing to agree that Rebel was as lucky as this individual, though
: perhaps it was somewhat fortunate that your machine has no algorithm which
: can even begin to simulate the intelligence required to address the decision
: of what bad moves it might be able to gamble on. This is not only hard for
: computers -- Soltis alludes to this as partly what distinguishes GM's from
: IM's (I think the book is called _Defense in Chess_).
: You seem to suggest it is time computers begin to consider taking a gamble
: in situations where imminent concessions lead to certain death. I don't
: even have the slightest idea how a computer is going to address these kind
: of decisions, but it certainly gets you thinking...
Yes... and it you think about it, eventually they have to factor this stuff
in to their decision making. Otherwise they are going to cook their own
goose too often. :)
: : : : Yet CB, because it just happened to search deeply enough, did see it,
: : : : and threw the game away because of it. This was the only game it lost
: : : : as I recall, also. That happened in round 1 of the NPS, although it was
: : : : not a mate, nor even a material loss... That's not knocking rebel, it's
: : : : not making excuses for Crafty losing. It is simply pointing out the fact
: : : : that a deeper search doesn't always result in better play. In this game,
: : : : Crafty played some poor moves to hang on to the passed E pawn, when it
: : : : should not have. And this put it into positions where it eventually had
: : : : to give up something somewhere. The deeper search, however, made it think
: : : : it had to give up more than it really did.
: : : Ha! That's the best line ever -- "...made it think it had to give up more
: : : than it really did." Sounds like you're saying that a deeper search is
: : : actually detrimental, since without it, you might have given up less!
: : : Maybe you should put a governor on how deep your program can search when
: : : it is considering giving up something!!
: : No, but if you'd simply read to understand, rather than reading to nit-pick,
: : the
: : problem is obvious. I didn't say a deeper search is worse. I said a deeper
: : search
: : can *sometimes* (can you see the emphasis here) result in playing a worse
: : move.
: : Simple statement. Easy to prove...
: The problem is understood. The solution is beyond my imagination, but as I
: said before, it certainly gets you thinking!
: : : : Sorry you took it as offensive. It wasn't intended to be. I didn't try to
: : : : make *any* excuses for Crafty. I didn't mention that in the 12 game
: : : : "calibration match" that crafty had black in every game, which would likely
: : : : affect the outcome by some amount. I didn't play up the fact that for the
: : : : first 7 games, crafty was playing with a gross "contempt factor" because it
: : : : thought Rebel had zero time left, and the draw score code sets contempt to
: : : : -1.000 in that case to try and avoid drawing when the opponent is about to
: : : : lose on time.
: : : Super! I didn't make *any* excuses, though here are a whole bunch I might
: : : have made... Sounds like something out of a poor comedy show!
: : Nope. Ed posted the excuses comment. I simply pointed out that I watched as
: : Crafty went 9-3 vs Rebel and didn't say a word about any of the above. The
: : above *are* excuses. They were not posted *until* after Ed chose to call off
: : the match.
: Okay, fair enough. I understand that you have your reasons for giving
: such analysis of this game (particularly considering that Ed has refused
: to allow you the opporutinity to complete the match).
"refused" is too strong. Ed ought to be able to decide to withdraw from such
a match if he wants. If we were talking a week, that's one thing, bug we are
really talking most of a year to play this match out. 6 months at a real
minimum. That's a lot of time...
: : : : In any case, as I mentioned before, the match was on your "nickel" and you
: : : : are certainly free to stop whenever you want. I knew it would take a long
: : : : time to play out. The calibration match was having problems due to auto232
: : : : quirks, but the one NPS game was simply clean so far as I'm concerned. The
: : : : book might have affected the game, it might not. However, when Rebel left
: : : : book, I didn't consider Crafty's position so inferior that the game should
: : : : be thrown out. Rebel played well. I think the most impressive thing I saw
: : : : was the fact that it did not seem to make *any* tactical mistakes, anywhere.
: : : : Which is surprising. I'd have liked to see more games to see if this could
: : : : continue.
: : : Excellent analysis above! The opponent suprised you by playing remarkably
: : : well, especially considering its more limited depth.
: : I never said otherwise, regardless of what you'd like to imply...
: I'm not aware that I implied anything other than perhaps you learned
: something interesting from the single game. If you feel I had implied
: that your initial position is somehow refuted by this single game, you
: take me wrong. It takes character to admit when your opponent plays
: suprisingly well, particularly given the circumstances of the match.
: Regardless what the outcome of the match might have been, I doubt very
: much there could be any definitive conclusion drawn, but it certainly
: was useful in providing information on the subject (I happen to think
: there is some truth in what each of you say, and it's really a tragedy
: that a debate has resulted in the two of you being pitted against one
: another in such a match).
Actually, neither Ed nor I take this personally I don't think. I've developed
thick skin over the years. To survive in this newsgroup I have to. To survive
on ICC I also have to. In my mind, about the bravest decision I ever made was
to release the source for Crafty to everyone, because all of a sudden, there is
no "mystery" to what I'm doing. If I'm doing something bad, everyone in the
world knows what it is. It's a very precarious position to put yourself in,
to say the least.
I take the ragging about bad moves here and there, because for every negative
comment I get, I get a couple of positive comments, some suggestions, and, on
occasion, some code contributed. It's been mucho-worth it to me. Even with
all the flak.
: I'll be honest, I never thought Ed would win a single game in this match.
: I think that a 4-ply handicap is TOO much to give up. I never would have
: jumped into this debate, but I was bothered by what seemed hypothetical
: excuses for having lost game one, when I expected an aknowledgement of
: his success in the first game.
I think Rebel played wonderfully, and I said so. I'm still amazed, that at
around move 15-20, when things sort of "opened up" that Rebel didn't make any
sort of tactical mistake at all, which says a lot about the program. I don't
believe it could continue to play that well, but we have no way of knowing at
present. That's not a knock on Rebel either... just that such a time handicap
is significant.
: Still, to claim that he's made his point, and then abort the match is
: IMO cowardly. He's made a point that it is possible to play a game with
: surprisingly good tactics when suffering a 4-ply handicap, and perhaps it
: is true that this is all he wanted to prove. But, you've aknowledged that,
: and you stick to your guns that your machine will win such a match. The
: real question was, can it maintain such form in a 10-game match, and that
: question remains. Again Ed, I hope you reconsider.
me too, or else we'll try this another way. The question is too interesting
to give up on getting an answer. At least this way we can speculate what the
200M nps monster can do, if we can get this match underway in a different way
somewhere...
: : : <SNIP>
: : : : The only thing you say that I don't agree with is the above. One game doesn't
: : : : prove anything. That was the intent for playing 10 games to start with,
: : : : rathe than a single game.
: : : One game is insufficient, I agree. But, one game does show that Rebel can
: : : play remarkably well when suffering from a 4-ply handicap in search depth.
: : yes it does.
: : : : However, in the interests of seeing how this works, I will see if I can set
: : : : up a "correspondence handicap game" with someone that is willing to operate
: : : : Rebel on the other end, just to see how this might continue. I can operate
: : : : it manually here, but don't want to give anyone the chance to cry foul...
: : : : It's still an experiment that needs doing, because it has never been done
: : : : before. Everyone agrees that one ply is not a significant advantage at these
: : : : depths. But we have absolutely no data about 4-5 plies. I hope you might
: : : : consider doing rebel vs rebelX100, as that would generate the data.
: : : Why not Crafty vs Crafty*100?
: : : I might also suggest, though perhaps it was already, that
: : : you might consider varrying the search depths (while keeping
: : : the differnce in depth constant). For example, 1-ply versus
: : : 5-ply, I presume, would be a huge advantage, whereas 10-ply
: : : versus 14-ply would be much less significant. Perhaps you
: : : can approximate the curve, and make some predication about
: : : what will happen at higher depths.
: : I don't think there's a point to doing Crafty vs Crafty*100... because we'd
: : only
: : fill up the newsgroup with endless discussion about "but that's crafty vs
: : crafty
: : with a handicap, that's doesn't say a thing about Rebel vs Deep Blue or
: : whatever."
: Good point.
: : As with Ed, I try to use at least some of my time wisely...
: Kevin.
Kevin James Begley
Robert Hyatt (hy...@crafty.cis.uab.edu) wrote:
: Kevin James Begley (kjbe...@chimi.engr.ucdavis.edu) wrote:
: : Chess is not a negociation. You play the best move you can find, period.
: : We certainly don't need to address claims of what might have been. Just
: : assume that you know nothing about your adversary, and let your machine's
: : superior ability avoid such situations.
: : Indeed, your claim that "fortune favors the foolish" is the most outlandish
: : excuse for losing I've ever heard.
: Where exactly did you see that "fortune favors the foolish" claim? Your
: statement
: is about the most outlandish interpretation of anything I've seen...
Right you are -- guess I just like the sound of that saying. Still, I don't
forsee any good way to address what might have been, and I doubt you (or any
other experts) do either.
Thus, I suggest that we ignore the hypotheticals, and stick with analyzing
the results. Unfortunately, Ed has pulled-out before we have enough results
to analyze (and thus, I suppose I understand why you must qualify what took
place in the single game that Ed is willing to complete in). If he is going
to leave us with this single game to analyze, perhaps we must look carefully
into why Crafty lost.
I must say, I don't think it is very sportsman-like for Ed to agree to a
10-game match, only to pull out after winning the first game. He may have
proven to some that Rebel's eval is suprisingly strong, but in pulling-out,
I feel he's denied Crafty the oppurtunity to make Bob's point (that such a
superior depth will overpower the shallower evaluation).
Ed, in the interest of good sportsmanship, I suggest you reconsider resigning
the remaining 9-games. If not for your own sake, do it for RJFischer, who
must have spent the last 20+ years with profound regrets for not giving
Karpov the oppurtunity he once had. Do it for Bobby, 'cause if he weren't
so old, so crazy, and so cowardly, by God, he'd do it for you!
<SNIP>
: : What's your point? Nobody knows how often this happens! I could just as
: : easily claim that I lost against a stronger player because I assumed he had
: : a refutation for the winning move that I failed to play. The result, I'm
: : afraid, is all that remains. Learn to live with the results.
: I *know* how often it happens, because I have only 80,000 games played on the
: Servers. I have over 10,000 games of Crafty vs GM players. I don't go
: through
: every one, but if you'd like I'll find a recent log and show you exactly what
: I'm talking about.
Don't bother, Bob. I know exactly what you are talking about
technically, but I wonder if you appreciate how difficult it would be to
find every instance in which one player made a concession (regardless of
how minute) because it feared a deep response possibly unseen by the
opponent. Surely, you realize this is approaching the limits of impossible,
so how can we even begin to make sense of the astronomical "what-if's" that
lie hidden in a chess game?
: : : A year ago, I posted an ugly game played by Cray Blitz vs a 1900 player
: : : in the 1984 US Open speed chess championship. CB had gone through IM's
: : : and GM's like a knife through butter, and near the end of the first
: : : "section" match (round robin with 16 players I think) we hit a 1900
: : : player, and were doing o.k.
: : If you were "doing o.k." how do you explain getting into a situation where
: : you must sac your queen to avoid mate?!
: It's called a tactically open position. You enter into a wild tactical
: melee, and make a move after a 9 ply search. one ply later you notice that
: there is a *very* deep stinger you overlooked before. Obviously it is deep
: because the 9 ply search didn't pick it up, but the 10 ply search did. It
: was attacking on the kingside, and realized at the end there was one quiet
: move that undid everything, about 10 moves down the game tree. And as soon
: as it noticed it, it did everything in its power to avoid getting mated.
: Have you ever attacked your opponent only to lose yourself? How'd you get
: into that position? It happens. Particularly when pieces are hanging all
: over the board in a wild attack.
Well, the question remains -- how much of this was your own misfortune, and
how much of this had something to do with how well the 1900-player had been
playing? How much of Rebel's success can be attributed to it's algorithms
(especially evaluation function), and how much can we conclude was pure luck?
I don't think you profess to know -- I certainly can't imagine how you could.
: : : Out of the clear, CB sacrificed it's queen
: : : for a knight. Victor Korchnoi was looking over my shoulder and said
: : : "machine make terrible blunder" in his deep, Russian gutteral accent.
: : I love the way you put that -- "out of the clear." As if the bad position
: : your machine stumbled into had nothing to do with it.
: I love the way you take things. Again, a wild tactical position, not a
: positional
: struggle, but one with pieces attacked, open files, zillions of checks, and it
: found what it thought was a brilliant path through all of that that won some
: sort
: of material, yet one ply deeper it saw the truth and tried to evade it...
: So it
: wasn't a bad "position" at all, just a tactical mistake. Guess that's too
: hard to grasp???
I grasp what you're saying -- simple case of the horizon effect catching you
in a tactically open position. Still, there's another school of thought --
one that would claim that you left yourself in a bad position (I said nothing
of positional struggle, just bad position). Their method of avoiding the
horizon effect is to better understand how you were positionally weak, and
by adding this knowledge into their evaluation, they hope to avoid such a
mistake. I am quite sure you grasp that. So, when you say things like "out
of the clear" you must understand why I consider that funny -- the other
school of thought maintains that NOTHING COMES FROM OUT OF THE CLEAR (or
stated differently, SUPERIOR TACTICS COME FROM A SUPERIOR POSITION).
: : : After the game, we backed up and I asked him what to play. He suggested
: : : something like RxR which appeared to be hanging. I tried it, and CB
: : : instantly announced a mate in 11 for the opponent. It saw that in the
: : : game and sacrificed the queen to avoid it. At the time, Korchnoi was the
: : : #3 player in the world I believe, and was Karpov's second at the time.
: : : He didn't see the mate. I doubt the 1900 player would have either.
: : Unfortunately, you will never know. Why don't you give us an algorithm
: : that will differentiate between the forced mates that an opponent will
: : see, and those he will certainly miss. You seem to have a good handle
: : on it yourself, so it ought to be easy for you to code it.
: Sorry, but I do know. He asked about why in the world did CB sac the
: queen. He thought he was getting mated in fact, and was very skeptical
: of taking the queen.
Quite possibly true. It may even be true that, had he stared at a potential
mate in eleven, he would not have found it under the game conditions (though
we can never test this for certain). Perhaps he obtained a winning position
purely out of luck, and perhaps he would have lost (even from having a forced
mate) had your machine played an inferior move. I don't know if I'd be
willing to agree that Rebel was as lucky as this individual, though
perhaps it was somewhat fortunate that your machine has no algorithm which
can even begin to simulate the intelligence required to address the decision
of what bad moves it might be able to gamble on. This is not only hard for
computers -- Soltis alludes to this as partly what distinguishes GM's from
IM's (I think the book is called _Defense in Chess_).
You seem to suggest it is time computers begin to consider taking a gamble
in situations where imminent concessions lead to certain death. I don't
even have the slightest idea how a computer is going to address these kind
of decisions, but it certainly gets you thinking...
: : : Yet CB, because it just happened to search deeply enough, did see it,
: : : and threw the game away because of it. This was the only game it lost
: : : as I recall, also. That happened in round 1 of the NPS, although it was
: : : not a mate, nor even a material loss... That's not knocking rebel, it's
: : : not making excuses for Crafty losing. It is simply pointing out the fact
: : : that a deeper search doesn't always result in better play. In this game,
: : : Crafty played some poor moves to hang on to the passed E pawn, when it
: : : should not have. And this put it into positions where it eventually had
: : : to give up something somewhere. The deeper search, however, made it think
: : : it had to give up more than it really did.
: : Ha! That's the best line ever -- "...made it think it had to give up more
: : than it really did." Sounds like you're saying that a deeper search is
: : actually detrimental, since without it, you might have given up less!
: : Maybe you should put a governor on how deep your program can search when
: : it is considering giving up something!!
: No, but if you'd simply read to understand, rather than reading to nit-pick,
: the
: problem is obvious. I didn't say a deeper search is worse. I said a deeper
: search
: can *sometimes* (can you see the emphasis here) result in playing a worse
: move.
: Simple statement. Easy to prove...
The problem is understood. The solution is beyond my imagination, but as I
said before, it certainly gets you thinking!
: : : Sorry you took it as offensive. It wasn't intended to be. I didn't try to
: : : make *any* excuses for Crafty. I didn't mention that in the 12 game
: : : "calibration match" that crafty had black in every game, which would likely
: : : affect the outcome by some amount. I didn't play up the fact that for the
: : : first 7 games, crafty was playing with a gross "contempt factor" because it
: : : thought Rebel had zero time left, and the draw score code sets contempt to
: : : -1.000 in that case to try and avoid drawing when the opponent is about to
: : : lose on time.
: : Super! I didn't make *any* excuses, though here are a whole bunch I might
: : have made... Sounds like something out of a poor comedy show!
: Nope. Ed posted the excuses comment. I simply pointed out that I watched as
: Crafty went 9-3 vs Rebel and didn't say a word about any of the above. The
: above *are* excuses. They were not posted *until* after Ed chose to call off
: the match.
Okay, fair enough. I understand that you have your reasons for giving
such analysis of this game (particularly considering that Ed has refused
to allow you the opporutinity to complete the match).
: : : In any case, as I mentioned before, the match was on your "nickel" and you
: : : are certainly free to stop whenever you want. I knew it would take a long
: : : time to play out. The calibration match was having problems due to auto232
: : : quirks, but the one NPS game was simply clean so far as I'm concerned. The
: : : book might have affected the game, it might not. However, when Rebel left
: : : book, I didn't consider Crafty's position so inferior that the game should
: : : be thrown out. Rebel played well. I think the most impressive thing I saw
: : : was the fact that it did not seem to make *any* tactical mistakes, anywhere.
: : : Which is surprising. I'd have liked to see more games to see if this could
: : : continue.
: : Excellent analysis above! The opponent suprised you by playing remarkably
: : well, especially considering its more limited depth.
: I never said otherwise, regardless of what you'd like to imply...
I'm not aware that I implied anything other than perhaps you learned
something interesting from the single game. If you feel I had implied
that your initial position is somehow refuted by this single game, you
take me wrong. It takes character to admit when your opponent plays
suprisingly well, particularly given the circumstances of the match.
Regardless what the outcome of the match might have been, I doubt very
much there could be any definitive conclusion drawn, but it certainly
was useful in providing information on the subject (I happen to think
there is some truth in what each of you say, and it's really a tragedy
that a debate has resulted in the two of you being pitted against one
another in such a match).
I'll be honest, I never thought Ed would win a single game in this match.
I think that a 4-ply handicap is TOO much to give up. I never would have
jumped into this debate, but I was bothered by what seemed hypothetical
excuses for having lost game one, when I expected an aknowledgement of
his success in the first game.
Still, to claim that he's made his point, and then abort the match is
IMO cowardly. He's made a point that it is possible to play a game with
surprisingly good tactics when suffering a 4-ply handicap, and perhaps it
is true that this is all he wanted to prove. But, you've aknowledged that,
and you stick to your guns that your machine will win such a match. The
real question was, can it maintain such form in a 10-game match, and that
question remains. Again Ed, I hope you reconsider.
: : <SNIP>
: : : The only thing you say that I don't agree with is the above. One game doesn't
: : : prove anything. That was the intent for playing 10 games to start with,
: : : rathe than a single game.
: : One game is insufficient, I agree. But, one game does show that Rebel can
: : play remarkably well when suffering from a 4-ply handicap in search depth.
: yes it does.
: : : However, in the interests of seeing how this works, I will see if I can set
: : : up a "correspondence handicap game" with someone that is willing to operate
: : : Rebel on the other end, just to see how this might continue. I can operate
: : : it manually here, but don't want to give anyone the chance to cry foul...
: : : It's still an experiment that needs doing, because it has never been done
: : : before. Everyone agrees that one ply is not a significant advantage at these
: : : depths. But we have absolutely no data about 4-5 plies. I hope you might
: : : consider doing rebel vs rebelX100, as that would generate the data.
: : Why not Crafty vs Crafty*100?
: : I might also suggest, though perhaps it was already, that
: : you might consider varrying the search depths (while keeping
: : the differnce in depth constant). For example, 1-ply versus
: : 5-ply, I presume, would be a huge advantage, whereas 10-ply
: : versus 14-ply would be much less significant. Perhaps you
: : can approximate the curve, and make some predication about
: : what will happen at higher depths.
: I don't think there's a point to doing Crafty vs Crafty*100... because we'd
: only
: fill up the newsgroup with endless discussion about "but that's crafty vs
: crafty
: with a handicap, that's doesn't say a thing about Rebel vs Deep Blue or
: whatever."
Good point.
: As with Ed, I try to use at least some of my time wisely...
Kevin.
Patrick Jagstaidt
brucemo wrote:
>
> Ed Schroder wrote:
> >
> > NPS challenge will end after game one...
> >
> > Today I decided to end the NPS challenge between Crafty and Rebel8 after
> > game-1 is finished.
>
> A shame, especially since it ends in your favor after one game. When you
> start a ten-game match, you should play a ten-game match, and not play a
> one-game match.
Crafty v.11.19 seems to do pretty well on ICC. Why not forget about the
game one, fix the obvious bugs in auto232 and start a new match, maybe
with a less tense coverage by the protagonists (since time is so
precious :)
> Declaring victory and pulling the plug after one game is really bad,
> especially after playing white in every control game as well.
I fully agree. This is bad from a marketing point of view too : the
experiment was very interesting and as you know from your hit counter it
was followed by a lot of people (and potential Rebel 8 buyers !)
Regards,
P. Jagstaidt
Robert Hyatt
brucemo (bru...@nwlink.com) wrote:
: Robert Hyatt wrote:
: > However, in the interests of seeing how this works, I will see if I can set
: > up a "correspondence handicap game" with someone that is willing to operate
: > Rebel on the other end, just to see how this might continue. I can operate
: > it manually here, but don't want to give anyone the chance to cry foul...
: I have a legal copy of Rebel 8 if you can't find someone who is more
: independent or otherwise better suited. I also have Genius 5, Mchess 6, and
: Nimzo 3, although I haven't installed them yet, if want to try something
: other than Rebel.
: By the way, thanks Bert, my order got here in one piece, and quickly too :-)
: The machine I have it on is a P5/133 with 32 megabytes, and it's running
: Windows '95, so perhaps there'd need to be some additional handicapping.
: bruce
I have a copy Ed sent as well.. However, if I do this, and that's a big
"if" indeed, I'd like it to be as impartial as possible. Personally, I'd
be just as happy playing genius, or CM5000, or any commercial program, because
the point's the same... what is 3-4-5 plies worth? I've run tests here
with crafty at time odds, but not anything like this match that Ed and I were
doing... I've tried 5 minutes to 30 seconds as the worst case, and only
a couple of those... But the depth is significant...
I can be convinced that those extra plies don't make any difference, but
I have not been yet. I remember Lonnie, last year, posting queries here
about what machine to buy, because all of a sudden, Crafty and Ferret were
thrashing him badly because we'd move to the P6/200 and he was still on a
P5/90 or P5/133, whatever. That was only a factor of 3-4 and it had him
running around in circles to close the computational gap. Now I'm supposed
to believe that after one game, we know that if I had a P6/20 gigahergz machine,
I'd not do any better against the commercial guys than I do now? I wonder
why I have such a hard time swallowing that?
How about the idea Lonnie? Is speed that important or not? You are about as
impartial a person as can respond here... having been on both sides of the
speed curve...
Chris Mayer
On 26 Mar 1997 17:12:02 GMT, hy...@crafty.cis.uab.edu (Robert Hyatt)
wrote:
>But, for the record, have you ever had Rebel play against a strong human,
>and saw Rebel throw the game away? Because it saw some deep tactic that
>it thought needed looking after, but which cost Rebel a pawn (or more) to
>defend against? A tactic that in reality, the human didn't see? If you
>haven't seen this, you simply haven't looked closely enough.
I once experimented with the following idea. When a computer does say
a 9 ply search, it is implicitly assuming the opponent will make the
best move based on an 8 ply search. What if you know (or want to
assume) that the opponent is only capable of even less? What I did
was to start off with the original 9 ply search, and instead of
returning a single PV, I returned all the top moves within a certain
score from the best. When there was more than 1 to consider, I then
made each move in turn, and looked at the response based not on the
opponent doing 8 ply, but only 6 instead. If the 6 ply response
differed from the 8 ply response, that's the move I chose. The result
is a sound 9 ply search, but also allowing for the opponent to make a
sound 7 ply response that is actually a 9 ply blunder. This played
much better against humans, and did about the same against computer
opponents. This may not be relavant to the above discussion, in that
the requrement for also being a sound 9 ply move may in some cases
hurt.
I never did a thourough or scientifically valid experiment with this,
but has anybody else done something similar?
Any feedback on this idea is welcome.
Chris Mayer
Joe McCaughan
Peter Kappler (pe...@bitsource.com) wrote:
: The termination of the match is very disappointing, and in my opinion,
: borders on poor sportsmanship, especially since you are enjoying a 1-0
: lead. Bob has explained his comments, and even apologized for the
: misunderstanding.
I have to agree. I think maybe Ed realized he had more to lose
and nothing to gain with this match. After a single win it may
have seemed to be a logical termination point.
: I also find the "waste of time" argument very amusing, considering the
: 100+ posts that you and Bob exchanged in the original thread which led
: to the NPS match. Now *that* might have been a waste of time, since
: it was basically speculation by both sides, but the match offers you a
: chance to *prove* your point!
The waste of time arg must have been well thought out before the
match. It was understood that Rebel would be at 40/2 and that
crafty would be at about 100x that. So the waste of time arg seems
watery at best.
Let's face it; Ed has commercial interests at stake, whereas crafty
is freeware. Robert had nothing to lose - Ed has more to lose.
: Please reconsider your decision, or at least continue the match at a
: faster time control. Perhaps a 30:1 time handicap instead of 100:1?
: This should only cost Crafty roughly 1 ply of search depth, and we'd
: still have an interesting experiment.
: One game proves very little...
: --Peter
Ed Schroder
From: hy...@crafty.cis.uab.edu (Robert Hyatt)
[ snip ]
: For instance, Bob Hyatt wrote:
: I'll post some interesting analysis later, but you might be
: surprised at some of what went on in this match. For example.
: in at least two places, Crafty played a weaker move, simply
: because it saw, at depth=15, something that it thought Rebel
: would play and it didn't like it. In these two cases I know of,
: I've tested rebel on the position and it wouldn't have played
: the move Crafty feared. So Crafty simply avoided something that
: would not have happened, by playing something that was worse.
: It was one of the common scenarios I've seen when two programs
: play and one outsearches the other by a significant amount.
> I don't see anything misleading at all in the above. Nor do I
> see any "excuse" making either. The thing I pointed out above
> happened. It has happened many times in the past, it will continue
> to happen many times in the future. And it's a problem that is
> going to have to be addressed at some point in time.
What you described above happened in a *lost* position.
That's what makes the big difference.
What's the point to mention it?
Ink clouds?
Even if Crafty would have played the moves of your choice would
it had helped? No because the black position was already lost at
that time. In Holland we call such statements "smoke screens".
> Sorry you took it as offensive. It wasn't intended to be. I didn't
> try to make *any* excuses for Crafty. I didn't mention that in the
> 12 game "calibration match" that crafty had black in every game,
> which would likely affect the outcome by some amount. I didn't
> play up the fact that for the first 7 games, crafty was playing
> with a gross "contempt factor" because it thought Rebel had zero
> time left, and the draw score code sets contempt to -1.000 in that
> case to try and avoid drawing when the opponent is about to lose
> on time.
Yup, but spelling everything what went wrong here in RGCC implied
the 9-3 doesn't mean anything because of all the Crafty bugs.
That's the difference between you and me, I in your case would have
said nothing here and would hide under my bed and stay there for at
least one week... :))
: Rebel played well. I think the most impressive thing I saw was
: the fact that it did not seem to make *any* tactical mistakes,
: anywhere. Which is surprising.
No it is *not* surprising!
You still miss my point...
After 25.h5 gxh5?? Crafty had a lost position. No 80-90 ply search will
help *AS LONG AS* white makes no mistakes. Just putting the white pieces
on the right squares, that's all there is. I believe it was Capablanca
who said this and his words are burned in my memory.
That's how chess works, that's what I believe in. Searching one ply
deeper for 80% is meant to find better positional moves (this is my
own personal definition). Tactics will come by itself. Game1 (till now)
is a perfect example of this.
If you have build a good position you usually don't have to fear for
any counter tactics since you have all the aces. It's a common rule
in chess. Will we ever agree on this? I wonder... :)))
The Pentium Pro 200 is a fantastic machine. The tactical barrier becomes
smaller and smaller and finally the "best" evaluation function will
decide a comp-comp chess game. And this trend will go on as soon as new
speed monsters will be released by Intel.
[ snip ]
- Ed schroder -
Peter Kappler
On 26 Mar 1997 16:00:51 GMT, Ed Schroder <rebc...@xs4all.nl> wrote:
>NPS challenge will end after game one...
>
>Today I decided to end the NPS challenge between Crafty and Rebel8 after
>game-1 is finished.
>
>Understanding the problems of running a match with an unfinished
>product, and since there are so many things that Bob may claim and so
>many things that did go wrong with the book, evals, time control,
>contempt factor, huge auto232 problems, bugs etc. I consider it a waste
>of my time to continue the match.
>
>Besides of this I also want to express my feelings about the following.
>
>I am interested in the scientific relationship between ply depth and an
>evaluation function. That's what has started this match. Sofar game-1
>has proven my point of view at least in my opinion.
>
>I am *not* interested to read all kind of possible and impossible
>explanations that imply to hide the real reason why Crafty will lose
>game-1.
>
>For instance, Bob Hyatt wrote:
>
>: I'll post some interesting analysis later, but you might be surprised
>: at some of what went on in this match. For example. in at least two
>: places, Crafty played a weaker move, simply because it saw, at
>: depth=15, something that it thought Rebel would play and it didn't
>: like it. In these two cases I know of, I've tested rebel on the
>: position and it wouldn't have played the move Crafty feared. So
>: Crafty simply avoided something that would not have happened, by
>: playing something that was worse. It was one of the common scenarios
>: I've seen when two programs play and one outsearches the other by
>: a significant amount.
>
>
>This would mean Crafty / Rebel would come into trouble against the
>Mephisto I, Boris etc. because Crafty / Rebel think 4-5 plies deeper.
>This is of course total nonsense. Crafty as any other program just
>played the best move on a given depth.
>
>I wouldn't dare to mention such an argument if Rebel would have had
>the 100 x time advantage!
>
>This is just one example...
>
>I am not in the mood for this kind of explanations for another 9 games
>the next 5-6 months. My time is too precious for that.
>
>Rebel8 played this game on an average ply depth of 9/10.
>Crafty played on an average ply depth of 13/14.
>
>Still Rebel8's evaluation function was good enough to handle the big
>ply gap of 4 plies. This has always been my point and game-1 has
>proven it. I believe this also counts for other chess programs like
>for instance Hiarcs etc.
>
>- Ed Schroder -
>
>
The termination of the match is very disappointing, and in my opinion,
borders on poor sportsmanship, especially since you are enjoying a 1-0
lead. Bob has explained his comments, and even apologized for the
misunderstanding.
I also find the "waste of time" argument very amusing, considering the
100+ posts that you and Bob exchanged in the original thread which led
to the NPS match. Now *that* might have been a waste of time, since
it was basically speculation by both sides, but the match offers you a
chance to *prove* your point!
Please reconsider your decision, or at least continue the match at a
Enrique Irazoqui
Kevin James Begley <kjbe...@chimi.engr.ucdavis.edu> escribió en artículo
<5hcn6q$8uh$1...@mark.ucdavis.edu>...
<snip>
> I must say, I don't think it is very sportsman-like for Ed to agree to a
> 10-game match, only to pull out after winning the first game. He may
have
> proven to some that Rebel's eval is suprisingly strong, but in
pulling-out,
> I feel he's denied Crafty the oppurtunity to make Bob's point (that such
a
> superior depth will overpower the shallower evaluation).
This is not a fair statement. One thing is to agree to play an auto232
match, which it simply involves connecting 2 computers, and a very
different one to have to deal with a number of technical problems
concerning the auto232 itself, opening books, compilers, etc. The match
ended up being very time consuming. I have been on the CC email line and I
witnessed all these problems. Ed's decision has nothing to do with
sportsmanship. One has so much time to dedicate to so many things, and
that's that. Like everybody else, I would have liked to see many more
games, but I understand that Ed's days have only 24 hours.
Enrique
Ingo Althoefer
Ed Schroeder wrote:
> NPS challenge will end after game one...
> Today I decided to end the NPS challenge between Crafty and Rebel8 after
> game-1 is finished. ...
Ed Schroeder is a free man, and it is his free decision to end the match.
I accept some of his explanations, although I do not like the decision.
One proposal: It would be more rectilinear to stop the match immediately,
instead of finishing game-1. Especially, noone should exploit
the data of game-1 for any public relation purposes.
Ingo Althoefer.
mclane
hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
>mclane (mcl...@prima.ruhr.de) wrote:
>: To Bob: If your program is so silly to consider that your opponent
>: will play always the best move, than that is your programs problem. I
>: mean: thats really the old problem: you expect the opponent to play a
>: move that your program has found to be the best. And now you make a
>: defender move that is a very passive one but helps against the
>: "THREAD" of the opponent THAT COULD HAPPEN IN PLY 15.
>Of course, *every* program currently playing also makes this same
>assumption. You might go read a little about alpha/beta and minimax,
>and zero-sum game theory...
NOT EVERY !! You forgot that I am working with chris for some years by
now. Of course I could have read many books about alpha/beta and
minimax in the same time, but I though working with chris could be
much more interesting !! There are some animals that are not as equal
as the other animals ! Have you forgotten this ?
All animals are equal ? Not in the british-forrests !
And I know other programs behave the same way like chris' does. There
are efforts arround the world. You have to recognize them, than you
have to study them, than you have to tell the people about. This is
what I do. But whenever I do it, you tell me: it can't be done. It is
wrong. It is old-fashioned.
>: But whose fault is it? Rebels ?! No - Rebel can play whatever it
>: wants. It's your fault, Bob. A few weeks ago we discussed much about
>: CSTal and Chris and I told you, that CSTal does NOT calculate accurate
>: and best play. Because NOBODY knows what BEST-PLAY is.
>No, there you are wrong. CSTal *does* play the move that it's search
>and evaluation can find. It is still based on alpha/beta/minimax, which
>is constrained to do this. Just because its eval likes to go off the
>deep end on complications, it still assumes the opponent will make the
>best move possible *according to the evaluation/search* in CSTal...
AEH ! How do YOU know a program better , I HAVE THE SOURCE CODES and I
HAVE WORKED WITH FOR MANY YEARS BY KNOW.
I have never seen YOU in between chris and my affair !
Sorry, I forgot you are GOD BOB, the first !
Of course under this circumstances you know anything better than
anybody !!!!
>: Bruce read all the stuff I told about how CSTal works, and said he
>: liked it. The idea that the opponent will move the RIGHT moves is
>: silly. If your program does not SEE / FEEL that all it's own ideas are
>: subjective and not objective, it is a programming-bug, or , to say it
>: not that drastically: it's a lack of recognition of the reality.
>: Alpha-Beta might be nice, but it does not help much here...
>: So I understand Ed's reaction. I think ED KNOWS that DIFFERENT ideas
>: about HOW TO PLAY are just IDEAS. Only fools would post that this
>: happens because this happens because this and this is forced because
>: of this and that, in a dicipline where NOTHING is safe.
>Sorry, but chess is a lot more precise than you are giving it credit
>for. We already know about N-piece endings and perfect play. There
>*is* a perfect result available in chess, we just don't know what it
>is, and may never know what it is. However, N ply searches produce
>information that is useful. N+1 ply searches produce *more* information.
>Otherwise, I'll expect everyone to continue using those old P90's for
>a long time, because that extra speed on the P6/200 obviously is not
>going to help them play any better. That *is* the argument, isn't it?
AND MAY NEVER KNOW WHAT IT IS !!!!!!!!
You are arguing against yourself !!
Of course there IS a perfect result. But if WE don't know about, and
we will never know, ....
N ply search produces information. No doubt about this.
N+1 ply search produces further information. No doubt about this too.
We are discussing about the SENSE and the how much this information is
worth.
When I know that I will never know if my effort is right, and I also
know that it would take ages and ages to find out, I should maybe try
in the meantime to change my strategy and try out some strategies that
will come to a better result by using different ways.
If the search-tree effort is no way out, maybe we have to increase the
efforts on the evaluation-topic.
>: Computerchess IS a topic where nothing is obvious, maybe a forced mate
>: is obvious, but anything out of this range is subjective. Search-trees
>: are very nice, but they do not help to solve the main-problem: the
>: range is almost infinite, or infinte for search-trees. Therefore Ed
>: and others trust more evaluations.
mclane
hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
>chrisw (chr...@cpsoft.demon.co.uk) wrote:
>: This is a total red herring. The point of the match was to consider
>: evaluation v. search depth. That can't be done on Rebel vs. Rebel at 100 to
>: 1 time odds.
>Not at all. The question is, once a program reaches some depth the author
>thinks is acceptable, will another 3-4-5 plies make a substantial difference
>in the outcome? I believe the answer is yes. The reason I posed this
>theory in the first place was in comparing Deep Blue to the current flock
>of Micro programs.
BOB, this is totally different with EACH DIFFERENT program.
Maybe the results you would get with Fritz would be totally different
than the results you would get Rebel8.
Again , this "theory" is exactly my argument against your whole:
Deep Blue is stronger-statements.
ANY program could be superior of its own version on a slower machine.
That does not prove the fact that DEEP BLUE is superior. It only shows
that the machines are different.
>If we had a Rebel vs Rebel*100 match, and the Rebel*100 wins by a wide
>margin, what would be the cause? All it could possibly be would be the
>extra depth. I agree that one ply is not a huge difference. Maybe even
>two plies is not. But 4-5 is a significant handicap...
Rebel8 is an intelligent program. Comparing it to a
fast/search-program would be another point.
The programs profit differently from search, because they differently
rely on evaluation or search.
>What better way to see if 3-5 plies makes a difference, than in a match
>where those 3-5 plies is the *only* difference. I never considered this
>a match between smart and fast. I considered this a match between slow
>and fast. There's nothing to say that DB's eval is not every bit as good
>as anything Rebel or anyone else is doing. But they are doing it so very
>very fast...
But WE considered this match as a match between smart and fast !
mclane
kjbe...@chimi.engr.ucdavis.edu (Kevin James Begley) wrote:
>You seem to suggest it is time computers begin to consider taking a gamble
>in situations where imminent concessions lead to certain death. I don't
>even have the slightest idea how a computer is going to address these kind
>of decisions, but it certainly gets you thinking...
Chess System Tal does this. It works. It is human. And it gives nice
games.
Many humans do it too. The game is called chess.
As long as nobody has proved chess, any move is gambling.
>The problem is understood. The solution is beyond my imagination, but as I
>said before, it certainly gets you thinking!
>I'm not aware that I implied anything other than perhaps you learned
>something interesting from the single game. If you feel I had implied
>that your initial position is somehow refuted by this single game, you
>take me wrong. It takes character to admit when your opponent plays
>suprisingly well, particularly given the circumstances of the match.
>Regardless what the outcome of the match might have been, I doubt very
>much there could be any definitive conclusion drawn, but it certainly
>was useful in providing information on the subject (I happen to think
>there is some truth in what each of you say, and it's really a tragedy
>that a debate has resulted in the two of you being pitted against one
>another in such a match).
A chess program should be able to direct the game the way that it
comes into the positions it knows/likes. Otherwise it is just a slave
of the opponent. Some programs have this nice feature. Rebel is one of
them, Genius is NOT able to do this.
>I'll be honest, I never thought Ed would win a single game in this match.
>I think that a 4-ply handicap is TOO much to give up. I never would have
>jumped into this debate, but I was bothered by what seemed hypothetical
>excuses for having lost game one, when I expected an aknowledgement of
>his success in the first game.
>Still, to claim that he's made his point, and then abort the match is
>IMO cowardly.
Don't forget that in the moment Hiarcs6 is trying to overtake Rebel8.
Of course Ed must be busy working on preparations for Rebel9.
Don't forget that Ed has to make money to pay his workers any month.
And also he sometimes wants to travel arround, e.g. next tournament I
hope to see him in AEGON, Den Haag to kill some GMs or Super-GMs !!
> He's made a point that it is possible to play a game with
>surprisingly good tactics when suffering a 4-ply handicap, and perhaps it
>is true that this is all he wanted to prove. But, you've aknowledged that,
>and you stick to your guns that your machine will win such a match. The
>real question was, can it maintain such form in a 10-game match, and that
>question remains. Again Ed, I hope you reconsider.
We can help Ed by overtaking the work on our machines against Crafty
operated by Bob. OK ?
>: : <SNIP>
Robert Hyatt
brucemo (bru...@nwlink.com) wrote:: >
: > >> Today I decided to end the NPS challenge between Crafty and Rebel8 after
: > >> game-1 is finished.
: >
: > >A shame, especially since it ends in your favor after one game.
: > >Both of you guys spent a long time working on this. I got on the CC line
: > >of the emails going back and forth, and there were a lot of them, from each
: > >of you. I have sixty or seventy email messages on this topic from Bob
: > >alone.
: > Seventy emails alone from Bob !!! Don't you see Ed's point?
: >
: > Ed has not the time to for that much problems, he must write Rebel 9
: > to beat Hiarcs 6 :-)
: Of course. It involved a lot of time to get Crafty + auto232 working.
: But, in the end, it was working, wasn't it?
: bruce
I'm really not sure. Several are playing Crafty vs XYZ on auto232, so I'm
waiting on feedback. However I had no "confirmed" success with the Rebel
match so far, so I don't know...
mclane
hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
>: I am interested in the scientific relationship between ply depth and an
>: evaluation function. That's what has started this match. Sofar game-1
>: has proven my point of view at least in my opinion.
>While I won't argue about game one, one game is not enough. If you
>still feel strongly about ply depth, try the alternative experiment
>I suggested, which is Rebel vs Rebel at 100:1 time odds. That will
>provide better data than Rebel vs Crafty, because the *only* thing
>different will be the depth. No differences in eval, no auto232
>problems, etc. And the results will be just as interesting...
Ahhhh ! Hear you now suggest my own ideas, when we discussed about
BEEP BLUE I told you to take one pc-program and to let it play against
itself on another 100 times or more faster machines.
NOW YOU SUGGEST IT. Brilliant.
So you HAVE understood my point that Deep Blue is nothing more than a
fast machine....
>: I am *not* interested to read all kind of possible and impossible
>: explanations that imply to hide the real reason why Crafty will lose
>: game-1.
>: For instance, Bob Hyatt wrote:
>: : I'll post some interesting analysis later, but you might be surprised
>: : at some of what went on in this match. For example. in at least two
>: : places, Crafty played a weaker move, simply because it saw, at
>: : depth=15, something that it thought Rebel would play and it didn't
>: : like it. In these two cases I know of, I've tested rebel on the
>: : position and it wouldn't have played the move Crafty feared. So
>: : Crafty simply avoided something that would not have happened, by
>: : playing something that was worse. It was one of the common scenarios
>: : I've seen when two programs play and one outsearches the other by
>: : a significant amount.
>I don't see anything misleading at all in the above. Nor do I see any
>"excuse" making either. The thing I pointed out above happened. It has
>happened many times in the past, it will continue to happen many times in
>the future. And it's a problem that is going to have to be addressed at
>some point in time.
>But, for the record, have you ever had Rebel play against a strong human,
>and saw Rebel throw the game away? Because it saw some deep tactic that
>it thought needed looking after, but which cost Rebel a pawn (or more) to
>defend against? A tactic that in reality, the human didn't see? If you
>haven't seen this, you simply haven't looked closely enough.
We all know this problem. But there are ways to help the program to
handle this conflict.
Humans have these problems too, when they plan for the future. A mean
person will always think that his enemies will plan also mean-attcks
on him. Thats called Paranoia.
>A year ago, I posted an ugly game played by Cray Blitz vs a 1900 player
>in the 1984 US Open speed chess championship. CB had gone through IM's
>and GM's like a knife through butter, and near the end of the first
>"section" match (round robin with 16 players I think) we hit a 1900
>player, and were doing o.k. Out of the clear, CB sacrificed it's queen
>for a knight. Victor Korchnoi was looking over my shoulder and said
>"machine make terrible blunder" in his deep, Russian gutteral accent.
>After the game, we backed up and I asked him what to play. He suggested
>something like RxR which appeared to be hanging. I tried it, and CB
>instantly announced a mate in 11 for the opponent. It saw that in the
>game and sacrificed the queen to avoid it. At the time, Korchnoi was the
>#3 player in the world I believe, and was Karpov's second at the time.
>He didn't see the mate. I doubt the 1900 player would have either.
>Yet CB, because it just happened to search deeply enough, did see it,
>and threw the game away because of it. This was the only game it lost
>as I recall, also. That happened in round 1 of the NPS, although it was
>not a mate, nor even a material loss... That's not knocking rebel, it's
>not making excuses for Crafty losing. It is simply pointing out the fact
>that a deeper search doesn't always result in better play. In this game,
Thats what WE say for ages now. The only problem in your statement is
the tiny "IN THIS GAME". We would like to see it deleted.
Because then we could make peace....
>Crafty played some poor moves to hang on to the passed E pawn, when it
>should not have. And this put it into positions where it eventually had
>to give up something somewhere. The deeper search, however, made it think
>it had to give up more than it really did.
>Sorry you took it as offensive. It wasn't intended to be. I didn't try to
>make *any* excuses for Crafty. I didn't mention that in the 12 game
>"calibration match" that crafty had black in every game, which would likely
>affect the outcome by some amount. I didn't play up the fact that for the
>first 7 games, crafty was playing with a gross "contempt factor" because it
>thought Rebel had zero time left, and the draw score code sets contempt to
>-1.000 in that case to try and avoid drawing when the opponent is about to
>lose on time.
>In any case, as I mentioned before, the match was on your "nickel" and you
>are certainly free to stop whenever you want. I knew it would take a long
>time to play out. The calibration match was having problems due to auto232
>quirks, but the one NPS game was simply clean so far as I'm concerned. The
>book might have affected the game, it might not. However, when Rebel left
>book, I didn't consider Crafty's position so inferior that the game should
>be thrown out. Rebel played well. I think the most impressive thing I saw
>was the fact that it did not seem to make *any* tactical mistakes, anywhere.
>Which is surprising. I'd have liked to see more games to see if this could
>continue.
That is Rebel8s main strength. Eds Rebel is not really GOOD in
tactics. But you cannot find out about in the GAMES. Because it is
able to control the game, not to allow the opponent to come into
tactics. SO: if one side is able to avoid certain parts of a game, I
would say it plans right.
On the other hand rebel8 itself is able to attack. Unbelievable
contrast that program is. It plays like a ballett-dancer who is also
able to sometimes fight Kung-fu against some criminals in the
south-bronx without breaking his legs or arms, and still does some
classical-ballett-dance afterwards.
>: Still Rebel8's evaluation function was good enough to handle the big
>: ply gap of 4 plies. This has always been my point and game-1 has
>: proven it. I believe this also counts for other chess programs like
>: for instance Hiarcs etc.
>: - Ed Schroder -
>The only thing you say that I don't agree with is the above. One game doesn't
>prove anything. That was the intent for playing 10 games to start with, rather
>than a single game.
>However, in the interests of seeing how this works, I will see if I can set
>up a "correspondence handicap game" with someone that is willing to operate
>Rebel on the other end, just to see how this might continue. I can operate
>it manually here, but don't want to give anyone the chance to cry foul...
>It's still an experiment that needs doing, because it has never been done
>before. Everyone agrees that one ply is not a significant advantage at these
>depths. But we have absolutely no data about 4-5 plies. I hope you might
>consider doing rebel vs rebelX100, as that would generate the data.
We can do it Bob, I have my 486DX4/100 wating for it....
I can give Rebel8 6 minutes to reach the pentium level. Or we can do
benchmarks with Rebel8 to find out ...
I would like to do this with you via email... ok ?!
mclane
df...@cse.ucsc.edu (Don Fong) wrote:
>In article <5hbrv5$d...@juniper.cis.uab.edu>,
>Robert Hyatt <hy...@crafty.cis.uab.edu> wrote:
>>When I get the log file for the one game, I'll show you what I mean in
>>a clearly understood form.
> i think most everyone already understands the concept.
> i think you mean, you are searching "deeper" in some sense
>than your opponent, you are considering many positions s/he doesn't.
>at some point you discover that your opponent has a beautiful winning
>line. your "best" defense against it looks lost, but not mated.
>this "best" defense tips your hand and makes it "easy" for the opponent
>to realize the advantage.
> if you are playing a weaker opponent, you can gamble that s/he
>won't see the winning moves. but it is still a gamble: you never
>know whether your opponent may "luck into" them.
> unless you know your opponent's mind to some extent, you cannot
>realistically assign probabilities to the various possible lines.
>of course, if you REALLY knew your opponent's mind, you could
>double your search depth. (:-)
> i think the real question is, how did your presumably weaker
>opponent manage to put you in such a pickle to begin with? if you
>are stronger, then why did it happen that YOU are on the wrong side
>of the deep combination, and not your opponent?
> the other real question (:-) is how to model your opponent's
>mind - how to decide whether to take the gamble.
Brilliant. Now the whole stuff goes directly into my favourite topic:
Philosophy !!!
Robert Hyatt
: >Ed Schroder wrote:
: >> Today I decided to end the NPS challenge between Crafty and Rebel8 after
: >> game-1 is finished.
: >A shame, especially since it ends in your favor after one game.
: >......
: >Both of you guys spent a long time working on this. I got on the CC line
: >of the emails going back and forth, and there were a lot of them, from each
: >of you. I have sixty or seventy email messages on this topic from Bob
: >alone.
: Seventy emails alone from Bob !!! Don't you see Ed's point?
: Ed has not the time to for that much problems, he must write Rebel 9
: to beat Hiarcs 6 :-)
At least half of those were direct responses to questions by Ed about
various things. For example, about 10 messages were concerned with a
problem Ed found in Rebel with "permanent brain off". He does some
timing analysis while thinking on the opponent's time, and this code
was not being exercised in this match, which was screwing up Rebel's
time allocation somewhat. We had a discussion about what to do, as
Crafty was having a similar problem in using time it thought it would
make up by thinking on the opponent's time, but which it couldn't make
up since it was not thinking on the opponent's time. I simply told Ed
to do whatever it took to make Rebel play correctly. If he needed to
stop/restart every 10 moves, or if he needed to manually adjust the clock
every 10 moves, whatever was needed. We were both on "unseen ground" at
this point. I elected to ignore the problem in crafty, since it had a
big time handicap. I didn't think that was reasonable for Rebel since
it only had about 4.5 minutes per move or so, and any foul-ups would
cause problems.
The problems weren't *all* mine. :) Only most of 'em were... the
high email traffic was generated to try and be sure that we were both
in agreement on solutions for every little snag that came up. so that
the results of the match would be valid.
Whether that took too much of his or my time really wasn't an issue then,
it might have become one later, although I had made the committment to
do what I could to make this happen. It wasn't easy for either of us,
that's for sure...
Robert Hyatt
Peter Herttrich (dh1...@inss1.etec.uni-karlsruhe.de) wrote:
: In rec.games.chess.computer Ed Schroder <rebc...@xs4all.nl> wrote:
: : NPS challenge will end after game one...
: : Understanding the problems of running a match with an unfinished
: : product, and since there are so many things that Bob may claim and so
: : many things that did go wrong with the book, evals, time control,
: : contempt factor, huge auto232 problems, bugs etc. I consider it a waste
: : of my time to continue the match.
: : I am interested in the scientific relationship between ply depth and an
: : evaluation function. That's what has started this match. Sofar game-1
: : has proven my point of view at least in my opinion.
: : For instance, Bob Hyatt wrote:
: : : I'll post some interesting analysis later, but you might be surprised
: : : at some of what went on in this match. For example. in at least two
: : : places, Crafty played a weaker move, simply because it saw, at
: : : depth=15, something that it thought Rebel would play and it didn't
: : : like it. In these two cases I know of, I've tested rebel on the
: : : position and it wouldn't have played the move Crafty feared. So
: : : Crafty simply avoided something that would not have happened, by
: : : playing something that was worse. It was one of the common scenarios
: : : I've seen when two programs play and one outsearches the other by
: : : a significant amount.
: : This would mean Crafty / Rebel would come into trouble against the
: : Mephisto I, Boris etc. because Crafty / Rebel think 4-5 plies deeper.
: : This is of course total nonsense. Crafty as any other program just
: : played the best move on a given depth.
: Ok! If someone spend me an auto232, i use my two 486DX4/100
: to make the challange. I own an Rebel8 and i have enough experience
: in use of Crafty. Both boxes have 32MB RAM so 12 or 13MB hash is
: no problem.
: Bob, if you tune a little bit the rook-problem (my last posting)
: or what u want (Crafty 11.19 or so), i would like to see, who is
: right, the one or the other side.
: By the way, from MY experiences and IMHO a ratio 1:10 in time is
: enough for this experiment.
The rook problem is long since fixed. It was the same passed pawn
values that were incorrect. I have an array that gives the values
of passed pawns rank by rank. I neglected to remember that value[0]
is the first rank where no pawn can ever be. As a result, the values
got higher than they should, quicker than they should. A passer on
the 6th is a serious threat. A passer on the 3rd is not. But if you
forget about 0-origin array subscripting because you've been playing
around with a FORTRAN program, it's easy to confuse 'em. :)
: Bob, another question: Is Crafty able to play under LINUX through
: the RS232 and contact the auto232 at the DOS-machine?
No. The reason is, I did not implement a serial-port level drive in
Crafty. Rather, I use the "PRN" interface from DOS, which is
intercepted by the auto232 TSR that is loaded before a game starts.
As a result, Crafty won't talk to a RS232 port, because the protocol
between chess program and TSR is significantly different from the
protocol auto232 uses to talk to it's "twin" on the other machine
over the RS232 link.
: It's not an absolutely must but it would be better because of the
: memorymanagment in LINUX.
I'd prefer Linux any day of course. That's Crafty's "native"
environment, as well as mine...
:
: I don't comment Ed's posting. But i agree, if someone needs it's
: time to make money, such an experiment cost! money :-)
: cheerio
: Peter
: --
: --
: _____________________________________________________________________________
: Peter Herttrich email: dh1...@inss1.etec.uni-karlsruhe.de
: Universitaet Karlsruhe Tel. +49 721 6083747 FAX +49 721 6086071
: Institut fuer Nachrichtentechnik ..life outside caves is complicated ...
: _____________________________________________________________________________
Robert Hyatt
Patrick Jagstaidt (jags...@iprolink.ch) wrote:
: brucemo wrote:
: >
: > >
: > > NPS challenge will end after game one...
: > >
: > > game-1 is finished.
: >
: > start a ten-game match, you should play a ten-game match, and not play a
: > one-game match.
: Crafty v.11.19 seems to do pretty well on ICC. Why not forget about the
: game one, fix the obvious bugs in auto232 and start a new match, maybe
: with a less tense coverage by the protagonists (since time is so
: precious :)
If it were only so easy. The problem is this:
the "protocol" between Crafty and the auto232 TSR program is not that
difficult. So far as I know that works perfectly, although there is at
least one serious design flaw that might affect a game here and there,
in that promotions *have* to be to a queen. There's no choice. If
Crafty were to try to promote to a knight, the game will hang at that
point.
However, the protocol is not the issue. When the opponent makes a move
and "my" auto232 TSR stuffs the move into the keyboard input buffer so
that Crafty will think the operator typed the move, there is a very
definite "timing window" where crafty has to read this move and echo it
to auto232. This is a pain. You can't echo it too quickly, or it will
hang up. You can't wait too long, or it will hang up. You have to do
it within some as of now unknown timing window. This is not easy to do
in DOS. In unix, I'd set an alarm and go about my business. In DOS,
such signals don't work the same, and most compilers don't even support
such things (DJGPP for example, the compiler we use for DOS).
Therefore, making this work is tough. And when you factor in the different
machines people have, and the speed variance, and couple this with a TSR
that has it's own ideas about timing requirements, the mixture is very
messy. I fix it on a P90, it breaks on a P6/200, I fix it on the P6/200,
now it takes too long and hangs on the P90... etc...
I don't write code like this, because I don't like timing dependent
problems. I'd rather use TCP/IP to sent internet traffic and let the
kernel worry about timing out packets and resending when needed, rather
than using UDP and having to "roll my own" timeout/retransmission code,
which would be just as messy...
: > Declaring victory and pulling the plug after one game is really bad,
Robert Hyatt
Ed Schroder (rebc...@xs4all.nl) wrote:
: From: brucemo <bru...@nwlink.com>
: > Today I decided to end the NPS challenge between Crafty and Rebel8
: > after game-1 is finished.
: : A shame, especially since it ends in your favor after one game.
: : When you start a ten-game match, you should play a ten-game match,
: : and not play a one-game match.
: I explained my reasons...
: > Understanding the problems of running a match with an unfinished
: > product, and since there are so many things that Bob may claim and so
: > many things that did go wrong with the book, evals, time control,
: > contempt factor, huge auto232 problems, bugs etc. I consider it a
: > waste of my time to continue the match.
: : Both of you guys spent a long time working on this. I got on the
: : CC line of the emails going back and forth, and there were a lot
: : of them, from each of you. I have sixty or seventy email messages
: : on this topic from Bob alone.
: : These were mostly about bugs in Crafty that surfaced due to the use
: : of Auto232, and because Crafty doesn't have a debugged install
: : procedure.
: : Most of the auto232 bugs seem to have been in the "control" match.
: : On ICC, Crafty is started, plays a game, then is terminated. With
: : auto232, it plays a game, then is told to start a new game, etc.
: : I just tried to call Bob to ask him for sure, but he wasn't home.
: : But the bugs that I know of were, 1) after the first game it thought
: : the opponent was out of time, so it changed its contempt factor to
: : -1.00 pawns, 2) for each successive game after the first game, it
: : used less and less time per move in the secondary time control.
: : It sounds like there were other weird things going on as well.
: : I think you were very patient to put up with the auto232 bugs. I
: : am not sure what caused your patience to run out. Aren't the problems
: : solved?
: I didn't test Bob's last version, after receiving 5-6 new versions I
: then decided to stop the 50-60 game match. The auto232 problem Bob is
: facing is exactly the same as Fritz3 has. I advised Bob to contact one
: of the programmers of auto232.
: I am used to get 3-4 games a day on 40/2:00 level with auto232. Because
: of all the auto232 communication troubles in Crafty's book phase I ended
: up with one game a day average.
: Auto232 is a strange animal, Genius, Nimzo, Rebel, Hiarcs, Mchess and
: Wchess all run perfectly with auto232. Bob fixed the permanent brain
: bug in Crafty in auto232 mode but for some reason the communication in
: Crafty's book phase is still not good. Bob has given it his best shot.
: It failed, what is a pity.
: I again would advice Bob to contact the Nimzo team who made the auto232
: software. They perhaps could give an instant answer since it's a known
: problem.
The *really* annoying problem here is that this is working better on
slower machines. I've had one person that has played a sack full of
games against Genius 3. The P6/200 is part of the problem, as is the
*very* fast book access mechanism Crafty uses.
The last version I sent introduced a 100ms delay before *anything* is
sent back to the auto232 TSR. The aggravating part of this is that
I have also heard of problems once the endgame is hit and crafty starts
ripping moves instantly from the tablebases. I hope this delay did the
trick. Time will tell...
: : I asked Bob what was going on when I noticed that the score in the
: : control match was 6-0, and he told me about the secondary time
: : control problem and the contempt factor problem.
: : It is obvious that this could have had an affect upon some games.
: : I asked Bob if he knew that it DID have an affect upon any of them,
: : and he told me it looked like Crafty made at least one really stupid
: : move, due to low time, in an otherwise even position. He may have
: : mentioned another game, I don't know.
: : But it wasn't like Bob was whining about this, I had to ask. I
: : didn't see any mail from Bob that I was CC'd on where he suggested
: : ignoring those six games. If you ignore them, they can be replayed
: : in a couple of days. If you keep them, big deal, it's just six games
: : out of 60, and maybe you would score 4-2 or 5-1 or 6-0 again if you
: : replayed them, so they aren't very meaningful.
: It's one of my reasons to stop this whole thing.
: How on earth can I get reliable results like this?
: Obvious Crafty isn't ready for auto232 yet.
: Why continue?
A reasonable comment. I've only worked on this off and on for a year. It
works. It doesn't... at least, recently, I began to get the idea that the
auto232 TSR has some sort of ugly timing requirements built into it that are
causing problems. In the middlegame, crafty was working great with Rebel,
because I don't check for input too often, which introduced a delay that
seemed to satisfy this beast. But in the opening, the book is too fast on
a good P6/200... and with tablebases (not used in the Rebel matches) it is
also too fast in the endgame.
In any case, I accept responsibility for the problems this caused, which is
one reason I never suggested throwing out the first 12 games. Crafty lost
the first game in Jakarta due to an operator error. Tom was pretty upset
about it. I wasn't, having done the same thing myself on more than one
occasion. Since the errors were mine, I thought the 9-3 result should stand.
The games *were* played out OTB, regardless of whether I had a subtle bug or
not....
: I have nothing on results which can be argued and are argued here, no?
: : After various bugs were fixed, Rebel and Crafty split the next
: : six games, 3-3.
: : Crafty blew out of book early in the NPS game because its book was
: : write-protected on the CD, but you guys both agreed to continue, so
: : who is to complain?
: > Besides of this I also want to express my feelings about the following.
: > I am interested in the scientific relationship between ply depth and an
: > evaluation function. That's what has started this match. Sofar game-1
: > has proven my point of view at least in my opinion.
: : You can't claim victory after one game in a ten-game match. I assume
: : there was a reason you guys agreed to play a ten-game match. Perhaps
: : Bob would not have agreed to a one-game match, with black, but that
: : isn't what you guys agreed to.
: : Declaring victory and pulling the plug after one game is really bad,
: : especially after playing white in every control game as well.
: That's up to auto232. Once you started a match the color will remain.
: After the 9-3 I received another version from Bob. In the weekend I
: went to my office installed it and started a new match Crafty having
: the white pieces for the next 12-13 games. Next day I noticed not
: one single game was played due to the auto232 book phase thing in Crafty.
: I accepted Bob's challenge to prove my point. It turned out to be the
: results can be (and are) argued on many points due to serveral bugs
: in Crafty. That's not what I had in mind.
But you are overlooking one point. The original "match" was working
o.k... Rebel played well. Crafty didn't. I could care less about the
"calibration match" as that can be done by anybody, at any time. But
the NPS match would have provided some data that we would likely never
see again, because it takes so long to produce it.
I've just sent a final draft of a paper to the JICCA on the parallel search
in Cray Blitz. almost 8 years after my dissertation was finished, because it
took *so* long to be able to play a complete game using 16 processors at a
real tournament time control, then same game with 8, then with 4, and 2 and
finally 1. Care to guess how long it takes Cray Blitz to move with 1 cpu when
it's trying to search to the same exact depth/etc as it did with 16? One move
with 16 processors took way over 30 minutes. With 1 processor, this blew
out to nearly a day's search. I'll probably *never* have the opportunity to
play a complete game like in this paper again, and play it without changing a
thing except the number of processors. In our NPS match, it is doubtful that
we will ever see this possibility again, unless we can convince Hsu to drop
by on ICC a few times and play a real game there, since he won't need a day
per move to simulate himself. :)
: Sorry, I can't function that way.
: - Ed Schroder -
Robert Hyatt
Ingo Althoefer (alth...@pdec01.uucp) wrote:
: Ed Schroeder wrote:
: > NPS challenge will end after game one...
: > Today I decided to end the NPS challenge between Crafty and Rebel8 after
: > game-1 is finished. ...
: Ed Schroeder is a free man, and it is his free decision to end the match.
: I accept some of his explanations, although I do not like the decision.
: One proposal: It would be more rectilinear to stop the match immediately,
: instead of finishing game-1. Especially, noone should exploit
: the data of game-1 for any public relation purposes.
Doesn't bother me if he puts the game on the front of the Rebel box. Rebel
won it completely fair and square. Whether it would win the next 9 or not
is something for conjecture, but there's no reason to "seal" the result in
a vault. From my perspective, I'm a big boy. I can take my lumps and come
back another day with a bigger stick. Eventually I'll find the right size
stick. :)
Robert Hyatt
mclane (mcl...@prima.ruhr.de) wrote:
: hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
: >: I am interested in the scientific relationship between ply depth and an
: >: evaluation function. That's what has started this match. Sofar game-1
: >: has proven my point of view at least in my opinion.
: >While I won't argue about game one, one game is not enough. If you
: >still feel strongly about ply depth, try the alternative experiment
: >I suggested, which is Rebel vs Rebel at 100:1 time odds. That will
: >provide better data than Rebel vs Crafty, because the *only* thing
: >different will be the depth. No differences in eval, no auto232
: >problems, etc. And the results will be just as interesting...
: Ahhhh ! Hear you now suggest my own ideas, when we discussed about
: BEEP BLUE I told you to take one pc-program and to let it play against
: itself on another 100 times or more faster machines.
: NOW YOU SUGGEST IT. Brilliant.
: So you HAVE understood my point that Deep Blue is nothing more than a
: fast machine....
delete the "is nothing more" and we agree. It is very fast. It is *not*
very stupid. That's where we don't agree at all. Time will prove or
disprove this...
: That is Rebel8s main strength. Eds Rebel is not really GOOD in
: tactics. But you cannot find out about in the GAMES. Because it is
: able to control the game, not to allow the opponent to come into
: tactics. SO: if one side is able to avoid certain parts of a game, I
: would say it plans right.
Email Bruce about his test with Rebel. You might change your mind in
a big hurry...
: We can do it Bob, I have my 486DX4/100 wating for it....
: I can give Rebel8 6 minutes to reach the pentium level. Or we can do
: benchmarks with Rebel8 to find out ...
: I would like to do this with you via email... ok ?!
All I can say at present is "possibly".
mclane
kjbe...@chimi.engr.ucdavis.edu (Kevin James Begley) wrote:
>Don Fong (df...@cse.ucsc.edu) wrote:
> <SNIP>
>: i think the real question is, how did your presumably weaker
>: opponent manage to put you in such a pickle to begin with? if you
>: are stronger, then why did it happen that YOU are on the wrong side
>: of the deep combination, and not your opponent?
>I would wager it had SOMETHING to do with a superior evaluation
>function. Deep search = superior tactics, deep evaluation = superior
>position. Certainly, Rebel didn't forsee all the tactics, but it
>did manage to obtain a better position with its evaluation.
>Perhaps what is needed is a mesh of the two styles -- deep evaluation
>to determine good positional candidates in non-tactical situations,
>and deep search to verify that some tactic is not missed along the
>way. It seems to be a question of which style is in greater demand
>today, and after Kasparov vs. Deep-Blue, I would think better eval
>would be the obvious answer.
> Kevin.
Welcome on our side of the world !
If you want to drive a car, you have to forget about the parts and you
have to drive and drive and after a while, you don't recognize the
parts anymore and drive with your unconsciousness.
Evaluation-function tries to emulate the process of forgetting the
search !! If you can evaluate a position very good, you don't have to
use ANY search at all.
This is the dream of any knowledge-based chess-programmer !
To write a 1 ply chess program.
evaluation function tries to emulate this feature, forward pruning
tries to FORGET the parts of the car, and and and....
mclane
Jack Nerad <JNE...@concentric.net> wrote:
>>
>> the other real question (:-) is how to model your opponent's
>> mind - how to decide whether to take the gamble.
ABOUT GAMBLING:
Computerschach and Spiele, edition 1/1985,
article of Frederic Friedel, page 28 and 29:
Game Human (ELO 2265) vs. Cray Blitz , shown by Ken Thompson, that
Blitz has played in open championships 1982 in Missisipi.
Position:
W Ke3 Ra8 Bf3 a4
B Ke5 Rf7 Be6 a7 b6
Black (Cray Blitz) to move.
47...Rxf3 ! (+4.093 evaluation)48. Kxf3 Bd5+ 49.Ke3 Bxa8
50.a5 ?!
50...Be4 !! (evaluation more than 10 pawns) 51.Kd2 Kd4!! 52.Kc1 b5!!
53.Kb2 Kc4 54.a6 b4 55.Ka2 Kc3 and Blitz says mate in 6.
Frederic mentioned that Cray Blitz played Be4 in the 50th move
because it saw the promotion in this line, and it saw NO promotion in
the other line.
Therefore Cray Blitz played Be4 without *knowing* about the
wrong-color-bishop-endgame.
In Computerschach and Spiele 2/1985 the readers discussed the position
with their dedicated chess computers and most interesting:
Mephisto IIIb from Nitsche/Henne played 50...Kd6 and did not tried to
EAT 50...bxa like Freddy promised,
also other Mephisto III clones did NOT played 50...bxa.
Superconstellation played 50...b5 after 2 hours and ply12.
But the most interesting part about it:
Superconny plays 50.a5 ?! and tries to gamble !
It does this move >level4 > 2 minutes !!!!!
We need Dave Kttinger back into the computer-chess-arena !
ICC and WchessX is nice, but not enough....
>Regards,
>Jack Nerad
mclane
hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
>Rebel played very well. However, in fairness with this position, I
>also tested with genius as well, and the funny part is, from the white
>side, given even fairly short times to move, Rebel, Genius *and* Crafty
>would follow the same path. White simply ended up with a good position
>after the first 10-15 moves, and the rest of the white plan almost played
>itself. Of course, as you point out, white had to reach that position,
>and black had to allow it. Black was thinking a little too far toward
>the endgame, trying to preserve a winning advantage, but, as a result,
>didn't make it through the middlegame unscathed.
Maybe you should not allow it to much too think that early about
ENDGAME.... :-)
>: the other real question (:-) is how to model your opponent's
>: mind - how to decide whether to take the gamble.
>Yep. and as several AI well-knowns have said over the years, this is a
>crucial step if minimax is going to work. Because at present, minimax doesn't
>work exactly like we do in this regard. If you start off with a good position,
>the extra depth should guarantee a win. If you start off with an inferior
>position (as black does, often enough) then the extra depth might not help,
>unless you can figure out what your opponent sees and doesn't see...
Gambling is one main feature of ChessSystemTal.
>as far as your "gamble" comment above, computers are a little more "obvious"
>than that. For example, it plays g4 and the pawn is hanging. A human might
>wonder "now why does it want my bishop off of that diagonal so badly?" and he
>will probably figure out why. And see something he had no inkling was there
>before the computer made a fishy move.
>Berliner played with a fix for this, but it wasn't good enough, in that it
>only tried to fix something that was seen in the very last search iteration.
>If you see something bad for 2 or more, it would assume it was obvious and
>play the "best" move (which was often an obviously losing move...)
Maybe Berliner should have asked Dave Kittinger how to gamble right.
I have seen Superconny GAMBLING very often. It was NOT obvious to see
why it is doing this and that. Maybe you always underestimate the
professional chess-programmers. Of course many things they produce are
not always PLANNED that accurate, but the results are often better
than any idea that is done at the universities.
Superconny was succesful in gambling with less rom, less speed and
less ram in 1984 ! Maybe Berliner never tried good enough ....
mclane
df...@cse.ucsc.edu (Don Fong) wrote:
>In article <5hbh7j$s91$1...@news1.xs4all.nl>,
>Ed Schroder <rebc...@xs4all.nl> wrote:
>>Today I decided to end the NPS challenge between Crafty and Rebel8 after
>>game-1 is finished.
> too bad! maybe YOU feel the point is proved, but to many people,
>the point would be better proved by multiple games.
>[...]
>>I am *not* interested to read all kind of possible and impossible
>>explanations that imply to hide the real reason why Crafty will lose
>>game-1.
> aw cmon. it's fun watching Hyatt squirm. (:-) (:-)
!!!!!! :-)
Schadenfreude ist die beste Freude, as we say in german !
>>For instance, Bob Hyatt wrote:
>>
>>: I'll post some interesting analysis later, but you might be surprised
>>: at some of what went on in this match. For example. in at least two
>>: places, Crafty played a weaker move, simply because it saw, at
>>: depth=15, something that it thought Rebel would play and it didn't
>>: like it. In these two cases I know of, I've tested rebel on the
>>: position and it wouldn't have played the move Crafty feared. So
>>: Crafty simply avoided something that would not have happened, by
>>: playing something that was worse. It was one of the common scenarios
>>: I've seen when two programs play and one outsearches the other by
>>: a significant amount.
> i think this is a tremendous argument!! Bob has just proved that
>increased search depth is a handicap. (:-) (:-)
RIGHT ! He has not recognized it himself, but he is trying to argue
with US against Deep Blue in the moment. Now HE is saying:
Deep Blue is only that strong because of the faster-search.
But - as you have seen - he is not seeing this.
>*FUTURE NEWSFLASH* Deep Blue XXVII, playing the Black pieces
>in a correspondence game, played the stunning move 1...Resigns!
>>This would mean Crafty / Rebel would come into trouble against the
>>Mephisto I, Boris etc. because Crafty / Rebel think 4-5 plies deeper.
>>This is of course total nonsense. Crafty as any other program just
>>played the best move on a given depth.
>>
>>I wouldn't dare to mention such an argument if Rebel would have had
>>the 100 x time advantage!
> no? i think we must do the experiment!
>>This is just one example...
>>
>>I am not in the mood for this kind of explanations for another 9 games
>>the next 5-6 months. My time is too precious for that.
> shucks. the public wants more!!
>>Rebel8 played this game on an average ply depth of 9/10.
>>Crafty played on an average ply depth of 13/14.
>>
>>Still Rebel8's evaluation function was good enough to handle the big
>>ply gap of 4 plies. This has always been my point and game-1 has
>>proven it. I believe this also counts for other chess programs like
>>for instance Hiarcs etc.
> i don't think it is fair to call off the match after one game.
>and what if CRAFTY pulls off a miracle? (it aint over til it's over.)
>i know i wouldn't be silent if Hyatt called off the match after one game.
>this is an historic event. what if Spassky had ended the Reykjavic
>match after the first 2 games? (as he probably could have done.)
Fischer would have been the champ. Fischer is always the champ. He is
the champ per definition and per playing strength. Kasparov is only
champion in the group of players where fischer is not counted in.
So he can call himself the champ of the mankind - Fischer.
Robert Hyatt
mclane (mcl...@prima.ruhr.de) wrote:
: kjbe...@chimi.engr.ucdavis.edu (Kevin James Begley) wrote:
: >You seem to suggest it is time computers begin to consider taking a gamble
: >in situations where imminent concessions lead to certain death. I don't
: >even have the slightest idea how a computer is going to address these kind
: >of decisions, but it certainly gets you thinking...
: Chess System Tal does this. It works. It is human. And it gives nice
: games.
: Many humans do it too. The game is called chess.
: As long as nobody has proved chess, any move is gambling.
: >The problem is understood. The solution is beyond my imagination, but as I
: >said before, it certainly gets you thinking!
: >I'm not aware that I implied anything other than perhaps you learned
: >something interesting from the single game. If you feel I had implied
: >that your initial position is somehow refuted by this single game, you
: >take me wrong. It takes character to admit when your opponent plays
: >suprisingly well, particularly given the circumstances of the match.
: >Regardless what the outcome of the match might have been, I doubt very
: >much there could be any definitive conclusion drawn, but it certainly
: >was useful in providing information on the subject (I happen to think
: >there is some truth in what each of you say, and it's really a tragedy
: >that a debate has resulted in the two of you being pitted against one
: >another in such a match).
: A chess program should be able to direct the game the way that it
: comes into the positions it knows/likes. Otherwise it is just a slave
: of the opponent. Some programs have this nice feature. Rebel is one of
: them, Genius is NOT able to do this.
I agree. In the first NPS game, Crafty did this. It just directed the
game in the wrong direction, due to evaluation problems I introduced. It
was so convinced that the passed pawn was *the* key issue of the game, it
made the game revolve around hanging on to it, which it did. Unfortunately,
it was revolving around the wrong goal, something that was my fault
entirely. I only have about 30 emails in my "to do" queue from people
telling me "you've got to get the passed pawn scoring under control again."
They were obviously correct.
: >I'll be honest, I never thought Ed would win a single game in this match.
: >I think that a 4-ply handicap is TOO much to give up. I never would have
: >jumped into this debate, but I was bothered by what seemed hypothetical
: >excuses for having lost game one, when I expected an aknowledgement of
: >his success in the first game.
: >Still, to claim that he's made his point, and then abort the match is
: >IMO cowardly.
: Don't forget that in the moment Hiarcs6 is trying to overtake Rebel8.
: Of course Ed must be busy working on preparations for Rebel9.
: Don't forget that Ed has to make money to pay his workers any month.
: And also he sometimes wants to travel arround, e.g. next tournament I
: hope to see him in AEGON, Den Haag to kill some GMs or Super-GMs !!
: > He's made a point that it is possible to play a game with
: >surprisingly good tactics when suffering a 4-ply handicap, and perhaps it
: >is true that this is all he wanted to prove. But, you've aknowledged that,
: >and you stick to your guns that your machine will win such a match. The
: >real question was, can it maintain such form in a 10-game match, and that
: >question remains. Again Ed, I hope you reconsider.
: We can help Ed by overtaking the work on our machines against Crafty
Robert Hyatt
: >chrisw (chr...@cpsoft.demon.co.uk) wrote:
: >: This is a total red herring. The point of the match was to consider
: >: evaluation v. search depth. That can't be done on Rebel vs. Rebel at 100 to
: >: 1 time odds.
: >Not at all. The question is, once a program reaches some depth the author
: >thinks is acceptable, will another 3-4-5 plies make a substantial difference
: >in the outcome? I believe the answer is yes. The reason I posed this
: >theory in the first place was in comparing Deep Blue to the current flock
: >of Micro programs.
: BOB, this is totally different with EACH DIFFERENT program.
: Maybe the results you would get with Fritz would be totally different
: than the results you would get Rebel8.
: Again , this "theory" is exactly my argument against your whole:
: Deep Blue is stronger-statements.
: ANY program could be superior of its own version on a slower machine.
: That does not prove the fact that DEEP BLUE is superior. It only shows
: that the machines are different.
: >If we had a Rebel vs Rebel*100 match, and the Rebel*100 wins by a wide
: >margin, what would be the cause? All it could possibly be would be the
: >extra depth. I agree that one ply is not a huge difference. Maybe even
: >two plies is not. But 4-5 is a significant handicap...
: Rebel8 is an intelligent program. Comparing it to a
: fast/search-program would be another point.
: The programs profit differently from search, because they differently
: rely on evaluation or search.
I don't disagree, except in one point. I have *never* seen a program that
didn't play better on faster hardware. Absolutely not one exception that
I'm aware of. But just hold on to that thought. Maybe after Kasparov, we
can get Hsu to drop in here with either the real deep blue, or with baby
blue which seems easier for him to get access to, and we can find out just
how well other programs fare.
And then the discussion can end as it should end..
Robert Hyatt
Ed Schroder (rebc...@xs4all.nl) wrote:
: From: hy...@crafty.cis.uab.edu (Robert Hyatt)
: [ snip ]
: : For instance, Bob Hyatt wrote:
: : I'll post some interesting analysis later, but you might be
: : surprised at some of what went on in this match. For example.
: : in at least two places, Crafty played a weaker move, simply
: : because it saw, at depth=15, something that it thought Rebel
: : would play and it didn't like it. In these two cases I know of,
: : I've tested rebel on the position and it wouldn't have played
: : the move Crafty feared. So Crafty simply avoided something that
: : would not have happened, by playing something that was worse.
: : It was one of the common scenarios I've seen when two programs
: : play and one outsearches the other by a significant amount.
: > I don't see anything misleading at all in the above. Nor do I
: > see any "excuse" making either. The thing I pointed out above
: > happened. It has happened many times in the past, it will continue
: > to happen many times in the future. And it's a problem that is
: > going to have to be addressed at some point in time.
: What you described above happened in a *lost* position.
: That's what makes the big difference.
: What's the point to mention it?
: Ink clouds?
I never tried to make an excuse for how it played, Ed. I pointed out
however, that it made some moves that it would not have made at a
shallower search. The game was *not* lost at move 15.
: Even if Crafty would have played the moves of your choice would
: it had helped? No because the black position was already lost at
: that time. In Holland we call such statements "smoke screens".
: > Sorry you took it as offensive. It wasn't intended to be. I didn't
: > try to make *any* excuses for Crafty. I didn't mention that in the
: > 12 game "calibration match" that crafty had black in every game,
: > which would likely affect the outcome by some amount. I didn't
: > play up the fact that for the first 7 games, crafty was playing
: > with a gross "contempt factor" because it thought Rebel had zero
: > time left, and the draw score code sets contempt to -1.000 in that
: > case to try and avoid drawing when the opponent is about to lose
: > on time.
: Yup, but spelling everything what went wrong here in RGCC implied
: the 9-3 doesn't mean anything because of all the Crafty bugs.
: That's the difference between you and me, I in your case would have
: said nothing here and would hide under my bed and stay there for at
: least one week... :))
I didn't spell everything that went wrong. If you back up and check
the posts, I didn't claim that the bogus time control stuff resulted
in that lopsided score. I didn't claim that 12 blacks in a row did
it either. I did explain what I had done, because I got several
email queries about "Bob, my version of 11.18 won't play that same
bad move at any depth. Is the 11.18 in the match the same or is it
different?" I explained that the programs were the same, but that
the contempt factor was broken. For all I know, playing those 12
games again might result in a 12-0 score.
I didn't object to your "doubles elimination." Which is not something
I agree with, by the way, because I spent about 3 months working on the
book learning stuff. If crafty gets a good position, it would love to
play that opening again, but the double elimination would cut the game
off. That's one way to solve a problem. There is another way, which is
to do what I've done, and what others have done. Otherwise, the results
you get won't match the results the SSDF gets and it won't match the
results that people get when they operate Rebel on a server manually.
If I wanted the best possible result, I'd have insisted that we do the
testing just like the SSDF, so that Rebel has to avoid repeating games
it loses, while Crafty would work like heck to play what wins and avoid
what loses. That wasn't the issue here. The issue was the NPS games
which might have provided some interesting data. The first one did,
for me anyway, because I did see a couple of places where Crafty was
dodging a bullet that wasn't going to be fired. That's the kind of
data I was hoping for, although I'd also like to see it win more games
than it loses too. :)
: : Rebel played well. I think the most impressive thing I saw was
: : the fact that it did not seem to make *any* tactical mistakes,
: : anywhere. Which is surprising.
: No it is *not* surprising!
: You still miss my point...
: After 25.h5 gxh5?? Crafty had a lost position. No 80-90 ply search will
: help *AS LONG AS* white makes no mistakes. Just putting the white pieces
: on the right squares, that's all there is. I believe it was Capablanca
: who said this and his words are burned in my memory.
Tell you what. I'll watch more carefully on the servers, and I'll post
a game or two where Rebel makes some horrible tactical mistakes against
Crafty. Perhaps you'll change your mind. Rebel plays well. It doesn't
play perfect, any more than any program does. BTW, a couple of your
commercial "cousins" also play gxh5... one very highly ranked...
: That's how chess works, that's what I believe in. Searching one ply
: deeper for 80% is meant to find better positional moves (this is my
: own personal definition). Tactics will come by itself. Game1 (till now)
: is a perfect example of this.
: If you have build a good position you usually don't have to fear for
: any counter tactics since you have all the aces. It's a common rule
: in chess. Will we ever agree on this? I wonder... :)))
: The Pentium Pro 200 is a fantastic machine. The tactical barrier becomes
: smaller and smaller and finally the "best" evaluation function will
: decide a comp-comp chess game. And this trend will go on as soon as new
: speed monsters will be released by Intel.
I assume you won't be needing the next P6/400 if you believe that? Since
another ply isn't going to help? I've personally never moved crafty to a
faster machine (or cray blitz for that matter) and not found that it played
better chess. So the depth did something, otherwise why are you testing on
a flock of P6/200 machines, rather than the more affordable P90's you
had? This sword has to cut with both edges. If you get better with a
3x performance increase, so does everyone else. And if you get better with
a 3x performance increase, someone else is going to get a lot better with a
100x performance increase. I don't see how you can argue otherwise...
Robert Hyatt
: > Kevin.
Sorry, but that's wrong. Ask one. To write a one-ply chess program
would be his worst nightmare, not his dream. Because the evaluation would
have to replace that which is done by search. Search + evaluation is the
way humans play chess. I doubt you'll ever find a GM that will play in a
real game without even considering what the opponent might try to do...
: evaluation function tries to emulate this feature, forward pruning
Ed Schroder
From: brucemo <bru...@nwlink.com>
> Today I decided to end the NPS challenge between Crafty and Rebel8
> after game-1 is finished.
: A shame, especially since it ends in your favor after one game.
I explained my reasons...
: I asked Bob what was going on when I noticed that the score in the
: control match was 6-0, and he told me about the secondary time
: control problem and the contempt factor problem.
: It is obvious that this could have had an affect upon some games.
: I asked Bob if he knew that it DID have an affect upon any of them,
: and he told me it looked like Crafty made at least one really stupid
: move, due to low time, in an otherwise even position. He may have
: mentioned another game, I don't know.
: But it wasn't like Bob was whining about this, I had to ask. I
: didn't see any mail from Bob that I was CC'd on where he suggested
: ignoring those six games. If you ignore them, they can be replayed
: in a couple of days. If you keep them, big deal, it's just six games
: out of 60, and maybe you would score 4-2 or 5-1 or 6-0 again if you
: replayed them, so they aren't very meaningful.
It's one of my reasons to stop this whole thing.
How on earth can I get reliable results like this?
Obvious Crafty isn't ready for auto232 yet.
Why continue?
I have nothing on results which can be argued and are argued here, no?
: After various bugs were fixed, Rebel and Crafty split the next
: six games, 3-3.
: Crafty blew out of book early in the NPS game because its book was
: write-protected on the CD, but you guys both agreed to continue, so
: who is to complain?
> Besides of this I also want to express my feelings about the following.
> I am interested in the scientific relationship between ply depth and an
> evaluation function. That's what has started this match. Sofar game-1
> has proven my point of view at least in my opinion.
: You can't claim victory after one game in a ten-game match. I assume
: there was a reason you guys agreed to play a ten-game match. Perhaps
: Bob would not have agreed to a one-game match, with black, but that
: isn't what you guys agreed to.
: Declaring victory and pulling the plug after one game is really bad,
: especially after playing white in every control game as well.
That's up to auto232. Once you started a match the color will remain.
After the 9-3 I received another version from Bob. In the weekend I
went to my office installed it and started a new match Crafty having
the white pieces for the next 12-13 games. Next day I noticed not
one single game was played due to the auto232 book phase thing in Crafty.
I accepted Bob's challenge to prove my point. It turned out to be the
results can be (and are) argued on many points due to serveral bugs
in Crafty. That's not what I had in mind.
Sorry, I can't function that way.
- Ed Schroder -
> I am *not* interested to read all kind of possible and impossible
> explanations that imply to hide the real reason why Crafty will lose
> game-1.
: Crafty lost game one because it didn't play well. Rebel won game one
: because it played well.
: There can be speculation about why this happened. Some of it may
: revolve around mistakes Bob made, rather than your own super-
: competence.
: I think Bob made a serious mistake with his king safety routine, which
: he should fix. There is probably more to it than this, certainly, and
: I'm sure a lot of the credit goes to Rebel for playing the position
: well. Perhaps you could explain WHY it played so well?
: With out without this fix, I would like to see game two, especially
: if Crafty gets to play white.
> For instance, Bob Hyatt wrote:
:: I'll post some interesting analysis later, but you might be surprised
:: at some of what went on in this match. For example. in at least two
:: places, Crafty played a weaker move, simply because it saw, at
:: depth=15, something that it thought Rebel would play and it didn't
:: like it. In these two cases I know of, I've tested rebel on the
:: position and it wouldn't have played the move Crafty feared. So
:: Crafty simply avoided something that would not have happened, by
:: playing something that was worse. It was one of the common scenarios
:: I've seen when two programs play and one outsearches the other by
:: a significant amount.
: This isn't an excuse, it is analysis. I'm not sure if he's right or
: not, but the idea that a program may kill itself via its own search
: depth is something that Bob has been interested in.
: Bob of all people has to be able to comment honestly on the games,
: even if this may involve questioning the idea that Rebel is the best
: program on the planet.
> This would mean Crafty / Rebel would come into trouble against the
> Mephisto I, Boris etc. because Crafty / Rebel think 4-5 plies deeper.
> This is of course total nonsense. Crafty as any other program just
> played the best move on a given depth.
: On ICC I played a rated 5 0 game, Ferret vs Spinach. Spinach is
: Ferret at 50 milliseconds per move, no pondering, on inferior
: hardware. So this is more than 100:1 time odds.
: I only did this once, as it really scared me to watch Ferret play
: at -2 for most of the game. It found a tactic later on and won, but
: I'd have been really pissed if it had dropped 32 rating points because
: of this little lark.
: Why do you think Spinach was up two pawns against Ferret? Do you
: think it out-searched it at 100:1 time odds? No, Ferret came out of
: book bad, couldn't negotiate the position, found a combination against
: for the opponent, and mitigated the problem by dumping material at the
: root.
: That this may have happened doesn't take anything away from Rebel's win
: in round one, of ten. As I have stated, Rebel played well and Crafty
: didn't. It doesn't do a program credit to get into a situation where
: it is finding combinations against itself, assuming that is what
: happened here.
> I wouldn't dare to mention such an argument if Rebel would have had
> the 100 x time advantage!
: It's not a particularly big deal. Self-criticism after losses doesn't
: take anything away from the opponent. "He beat me because he's god",
: doesn't offer the loser much hope.
> This is just one example...
> I am not in the mood for this kind of explanations for another 9 games
> the next 5-6 months. My time is too precious for that.
> Rebel8 played this game on an average ply depth of 9/10.
> Crafty played on an average ply depth of 13/14.
> Still Rebel8's evaluation function was good enough to handle the big
> ply gap of 4 plies. This has always been my point and game-1 has
> proven it. I believe this also counts for other chess programs like
> for instance Hiarcs etc.
: Rather than declaring the match over because you won game one, please
: come back and play the rest of the match.
: bruce
Robert Hyatt
Chris Mayer (cma...@ix.netcom.com) wrote:
: On 26 Mar 1997 17:12:02 GMT, hy...@crafty.cis.uab.edu (Robert Hyatt)
: wrote:
: >But, for the record, have you ever had Rebel play against a strong human,
: >and saw Rebel throw the game away? Because it saw some deep tactic that
: >it thought needed looking after, but which cost Rebel a pawn (or more) to
: >defend against? A tactic that in reality, the human didn't see? If you
: >haven't seen this, you simply haven't looked closely enough.
: I once experimented with the following idea. When a computer does say
: a 9 ply search, it is implicitly assuming the opponent will make the
: best move based on an 8 ply search. What if you know (or want to
: assume) that the opponent is only capable of even less? What I did
: was to start off with the original 9 ply search, and instead of
: returning a single PV, I returned all the top moves within a certain
: score from the best. When there was more than 1 to consider, I then
: made each move in turn, and looked at the response based not on the
: opponent doing 8 ply, but only 6 instead. If the 6 ply response
: differed from the 8 ply response, that's the move I chose. The result
: is a sound 9 ply search, but also allowing for the opponent to make a
: sound 7 ply response that is actually a 9 ply blunder. This played
: much better against humans, and did about the same against computer
: opponents. This may not be relavant to the above discussion, in that
: the requrement for also being a sound 9 ply move may in some cases
: hurt.
: I never did a thourough or scientifically valid experiment with this,
: but has anybody else done something similar?
: Any feedback on this idea is welcome.
: Chris Mayer
If you read HiTech's description, Berliner tried a trick that is
similar.
He noticed that on occasion, after finishing a 9 ply search, that move
X was best, and had appeared best for several iterations. When HiTech
started the 10th (and last) iteration, this move could fail low. Suppose
that at depth=9, the score was +1.2, but at depth=10, this drops to -1.7
or so. And then another move improves this to -1.3... If you look at
the two moves, we know something about the first, but not much about the
second.
For the first move, we know it was best for a long time, and only became
worse after a very deep search. So for reasonable search depths, it looks
like a good move.
For the second move, we only know it looks to be about .4 better, but
it never looked good earlier.
It is possible that the first move loses a pawn to a very deep combination,
while the second just tosses it out and says "here, take it." Berliner
assumed that in the above case, the first move was better because of the
depth issue. I believe he reported this happening in one game and affecting
the outcome, but I could be mistaken. But the idea is pretty logical.
The drawback is that this was only used if the last iteration found the
problem. If depth 10 finds this sort of fail low, but then you continue
to 11 and even 12, the new best move will probably stay best, and now it's
not so clear which is better.
Difficult problem...
Robert Hyatt
: >mclane (mcl...@prima.ruhr.de) wrote:
: >: To Bob: If your program is so silly to consider that your opponent
: >: will play always the best move, than that is your programs problem. I
: >: mean: thats really the old problem: you expect the opponent to play a
: >: move that your program has found to be the best. And now you make a
: >: defender move that is a very passive one but helps against the
: >: "THREAD" of the opponent THAT COULD HAPPEN IN PLY 15.
: >Of course, *every* program currently playing also makes this same
: >assumption. You might go read a little about alpha/beta and minimax,
: >and zero-sum game theory...
: NOT EVERY !! You forgot that I am working with chris for some years by
: now. Of course I could have read many books about alpha/beta and
: minimax in the same time, but I though working with chris could be
: much more interesting !! There are some animals that are not as equal
: as the other animals ! Have you forgotten this ?
: All animals are equal ? Not in the british-forrests !
: And I know other programs behave the same way like chris' does. There
: are efforts arround the world. You have to recognize them, than you
: have to study them, than you have to tell the people about. This is
: what I do. But whenever I do it, you tell me: it can't be done. It is
: wrong. It is old-fashioned.
never said that. However, chris does use alpha/beta and minimax, and
he suffers from the same problem. If his eval + the material score says
the opponent can do "x", the search believes it and will do everything it
can to avoid X, regardless of tal-like evaluations and whatever. That's
the way minimax works... a zero-sum game. were he to switch to some sort
of best-first or other non-minimax approach, this might be different. BUt
when a problem appears within the horizon of the search, the search *must*
respond. It can't work any other way.
: >: But whose fault is it? Rebels ?! No - Rebel can play whatever it
: >: wants. It's your fault, Bob. A few weeks ago we discussed much about
: >: CSTal and Chris and I told you, that CSTal does NOT calculate accurate
: >: and best play. Because NOBODY knows what BEST-PLAY is.
: >No, there you are wrong. CSTal *does* play the move that it's search
: >and evaluation can find. It is still based on alpha/beta/minimax, which
: >is constrained to do this. Just because its eval likes to go off the
: >deep end on complications, it still assumes the opponent will make the
: >best move possible *according to the evaluation/search* in CSTal...
: AEH ! How do YOU know a program better , I HAVE THE SOURCE CODES and I
: HAVE WORKED WITH FOR MANY YEARS BY KNOW.
: I have never seen YOU in between chris and my affair !
Because I *know* that Chris uses alpha/beta/minimax. And there is only
one way *that* works. Just because you have the source doesn't mean anything
of course, you have to be able to understand it. Obviously you don't...
: Sorry, I forgot you are GOD BOB, the first !
: Of course under this circumstances you know anything better than
: anybody !!!!
I know alpha/beta/minimax as *well* as anybody, not necessarily any better,
because it is not a complex algorithm. But it does do certain things that
can't be changed, as mentioned above.
: >: Bruce read all the stuff I told about how CSTal works, and said he
Jack Nerad
Robert Hyatt wrote:
>
> However, the protocol is not the issue. When the opponent makes a move
> and "my" auto232 TSR stuffs the move into the keyboard input buffer so
> that Crafty will think the operator typed the move, there is a very
> definite "timing window" where crafty has to read this move and echo it
> to auto232. This is a pain. You can't echo it too quickly, or it will
> hang up. You can't wait too long, or it will hang up. You have to do
> it within some as of now unknown timing window. This is not easy to do
> in DOS. In unix, I'd set an alarm and go about my business. In DOS,
> such signals don't work the same, and most compilers don't even support
> such things (DJGPP for example, the compiler we use for DOS).
>
> Therefore, making this work is tough. And when you factor in the different
> machines people have, and the speed variance, and couple this with a TSR
> that has it's own ideas about timing requirements, the mixture is very
> messy. I fix it on a P90, it breaks on a P6/200, I fix it on the P6/200,
> now it takes too long and hangs on the P90... etc...
>
> I don't write code like this, because I don't like timing dependent
> problems. I'd rather use TCP/IP to sent internet traffic and let the
> kernel worry about timing out packets and resending when needed, rather
> than using UDP and having to "roll my own" timeout/retransmission code,
> which would be just as messy...
>
How involved a program is auto232? Is is something that one could write
the code for or is it so complex that even dealing with a bug-ridden
version is better than starting fresh?
Regards,
Jack
Nerad
chrisw
--
http://www.demon.co.uk/oxford-soft
Robert Hyatt <hy...@crafty.cis.uab.edu> wrote in article
<5he3lp$a...@juniper.cis.uab.edu>...
> Ingo Althoefer (alth...@pdec01.uucp) wrote:
>
> : Ed Schroeder wrote:
> : > NPS challenge will end after game one...
>
> : > Today I decided to end the NPS challenge between Crafty and Rebel8
after
> : > game-1 is finished. ...
>
> : Ed Schroeder is a free man, and it is his free decision to end the
match.
> : I accept some of his explanations, although I do not like the decision.
>
> : One proposal: It would be more rectilinear to stop the match
immediately,
> : instead of finishing game-1. Especially, noone should
exploit
> : the data of game-1 for any public relation purposes.
>
> Doesn't bother me if he puts the game on the front of the Rebel box.
Rebel
> won it completely fair and square. Whether it would win the next 9 or
not
> is something for conjecture, but there's no reason to "seal" the result
in
> a vault. From my perspective, I'm a big boy. I can take my lumps and
come
> back another day with a bigger stick. Eventually I'll find the right
size
> stick. :)
>
A rhthym stick ? :)
Chris Whittington
>
Robert Hyatt
mclane (mcl...@prima.ruhr.de) wrote:
: Jack Nerad <JNE...@concentric.net> wrote:
: >>
: >> the other real question (:-) is how to model your opponent's
: >> mind - how to decide whether to take the gamble.
: ABOUT GAMBLING:
: Computerschach and Spiele, edition 1/1985,
: article of Frederic Friedel, page 28 and 29:
: Game Human (ELO 2265) vs. Cray Blitz , shown by Ken Thompson, that
: Blitz has played in open championships 1982 in Missisipi.
: Position:
: W Ke3 Ra8 Bf3 a4
: B Ke5 Rf7 Be6 a7 b6
: Black (Cray Blitz) to move.
: 47...Rxf3 ! (+4.093 evaluation)48. Kxf3 Bd5+ 49.Ke3 Bxa8
: 50.a5 ?!
: 50...Be4 !! (evaluation more than 10 pawns) 51.Kd2 Kd4!! 52.Kc1 b5!!
: 53.Kb2 Kc4 54.a6 b4 55.Ka2 Kc3 and Blitz says mate in 6.
: Frederic mentioned that Cray Blitz played Be4 in the 50th move
: because it saw the promotion in this line, and it saw NO promotion in
: the other line.
: Therefore Cray Blitz played Be4 without *knowing* about the
: wrong-color-bishop-endgame.
Frederic was wrong. CB had RP + wrong bishop code then. In fact, ken
thompson ran this same position against belle and sent me a note after
the game that he was impressed that it knew about this, because the
search for belle took forever to find that taking the free pawn and
ending up with two a-pawns was a draw... CB knew and wouldn't have
taken it period. Crafty's code was lifted right out of Cray Blitz for
this case...
depth time score variation (1)
1 0.00 4.128 Be4 axb6
1-> 0.00 4.128 Be4 axb6
2 0.00 ++ Be4!!
2 0.00 6.648 Be4 Kd2 Kd4 <HT>
Not sure why this was written about in 1985... the game was played in
1981 in Jackson Mississippi. The human was Joe Sentef. I believe Ken
said belle took something like 14-16 plies before it decided that taking
the pawn was a bad thing...
: In Computerschach and Spiele 2/1985 the readers discussed the position
: with their dedicated chess computers and most interesting:
: Mephisto IIIb from Nitsche/Henne played 50...Kd6 and did not tried to
: EAT 50...bxa like Freddy promised,
: also other Mephisto III clones did NOT played 50...bxa.
: Superconstellation played 50...b5 after 2 hours and ply12.
: But the most interesting part about it:
: Superconny plays 50.a5 ?! and tries to gamble !
: It does this move >level4 > 2 minutes !!!!!
: We need Dave Kttinger back into the computer-chess-arena !
: ICC and WchessX is nice, but not enough....
He's still here... :)
: >Regards,
: >Jack Nerad
Simon Read
hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
This is disappointing. Crafty losing despite a 100:1 time advantage
is indicative of something very interesting happening, and
makes me want to see more. More games would provide more data
and would lead to stronger conclusions about the relative
utility of Crafty's speed and Rebel's evaluation function.
> I tried it, and CB
>instantly announced a mate in 11 for the opponent. It saw that in the
>game and sacrificed the queen to avoid it. [...]
>Yet CB, because it just happened to search deeply enough, did see it,
>and threw the game away because of it. [...]
> It is simply pointing out the fact
> that a deeper search doesn't always result in better play.
Surely, in your example above, sacrificing the queen was CB's best
move, since it avoided the mate?
Surely you were just unlucky to be on the wrong side of the
mate-in-11 and you would equally often be on the right side
of that combination? This would mean that freak combinations
would, statistically, not affect your results at all.
In the Rebel-Crafty game, Crafty might again have been unlucky,
or it might have been Rebel's evaluation function putting Crafty
in a position where such a combination was likely to pop up. More
games would give us more of a clue as to which of these it was.
Bob, in another posting you discuss methods of getting round the
problem of avoiding things which the opponent wouldn't see. This
all comes under the general heading of opponent modelling which,
I gather, is at least as difficult as the various types of
learning you are working on now.
It's very difficult for a computer to tell how deeply its human
opponent is searching. There are some combinations which look
deep to a computer, but are relatively simple for a human,
being composed of one main idea but with lots of little delaying
moves. So crafty might decide its opponent is searching to a
depth of 6 or 8 ply (beause the opponent lost a pawn) and set traps
appropriately, but then get completely blown away by a 10-ply
combination which it could have avoided.
More subtly, the total task facing crafty is to reproduce what
its opponent would play in any given position. This means
working out what its oppponent's evaluation function is and
its search strategy.
In theory this is impossible, since an
evaluation function can be arbitrarily complex, and the only
information you have to give you a clue is the move made, one
move out of a small set of possible moves. There might be
something lurking in the opponent's evaluation function and
search algorithm which has not yet had any effect on the
move choice. Each time the opponent makes a move, crafty
acquires a small amount of extra information about the
opponent, but does not ever know that it knows all there is
to know about its opponent.
In practice, it's not impossible, just very difficult. :)
Simon s.read at ____ cranfield __.__ ac __.__ uk
Rolf Tueschen
hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
>Rebel played very well. However, in fairness with this position, I
>also tested with genius as well, and the funny part is, from the white
>side, given even fairly short times to move, Rebel, Genius *and* Crafty
>would follow the same path. White simply ended up with a good position
>after the first 10-15 moves, and the rest of the white plan almost played
>itself. Of course, as you point out, white had to reach that position,
>and black had to allow it. Black was thinking a little too far toward
>the endgame, trying to preserve a winning advantage, but, as a result,
>didn't make it through the middlegame unscathed.
------------------------------------------------------------------------------------------------------------
Bob, excuse me, but I'm not on your side in this debate. It makes me really sad
to watch you shifting arguments without requiescence point ...
You overlook on the one side that computer chess is not primarily a question of
best moves. But -- as Ed and Chrilly pointed out -- a matter of omitting too
stupid moves.
On the other side you're occupied with this clearvoyance problem for which I
had the copyright -- at least regarding your heavy attacking me for *that*.
If you misunderstand the *learning* feature as sort of hintsight for the almost
automatical responses of a certain prg you'll fail to catch the problems your
krafti had in game one. (Just to remembering the GK thread: I always told you
GK would try to make the games with as much suspense as possible. But that if
he wanted he could *kill* DB *simply* by playing the expected automatical
weaknesses ... )
I think Ed is right with his non-amusement about your crying game with a lot of
different *unfinished* toys you wished to add to show us the *real* strength of
Crafty (playing factor 100!!). Ed is right, that for anyone who's a bit
aquainted to the game of chess (which is a bit different from pure
C-code-programming) the non-visibility of strength by pure search depth has
become clear throughout the game one. Rebel simply outplayed Crafty
positionnally. The way he went to the middle and then to the weak b5 -- was
high classed. Because a normal 1900-2000 player would have lost his goals after
the *end* of the former kings-attack. By playing 33.Bd3 and NOT Bh6 Rebel went
over the last trap. (Bh6 --- Nf4!).
So instead of crying over krafti -- who surely could be ameliorated for the
task -- we should applaude Rebel for a nice game.
As far as we discuss science I may be allowed to add the following. The
argument of Ed was crystal clear before the event. It hadn't been necessarily
be proved by a ten games match as discribed. Even if Crafty would hold the
score in his favour the former argument still would stand firm. I'm even not
convinced that the match in this form was the best way to test the thesis.
But --- but, after you gave your consent to this, you must deal with what
happened now. Crafty *played* from a standpoint of a real chessplayer NOT on a
level comparable of 2500 ... But more on that of a human Elo of 2000 maximum.
Wheras Rebel played the *200* *more* to keep the opponent safely down ...
Since Crafty played with this outrageous deep time space we could hold the
following results:
Theoretically a deeper search of 4-6 plies should be a terrible advantage.
IF, and only if, if the setting of the evaluation is fitted to the *deeper
chess space*.
For the time being, your prg seems to lack all that. But don't be sad, DB has
the same problem. It is the problem -- as I wrote several times -- of the not
yet prepared thousands of chess patterns that a real grandmaster of >2600
disposes of.
The other problem remains unsolved: how did Ed do his work in his limited deep
search? -- But also for this question you could read some answers in the
steinwender/Friedel book ...
For me this is far too difficult to understand. And I want the CIA to stay out
of this. Secrets are not to be sold! :))
Folks, remember, I love you all. :)
Especially on easter. It's like an automatical reflex somehow.
Rolf <Still weak at computer programming> Tueschen
Robert Hyatt
Pete Nielsen (pnie...@ucla.edu) wrote:
: --
: Re: "Crafty played a weaker move because ... "
: Excuse my ignorance, but this has been puzzeling me ever since it came up
: in the DB commentary.
: Doesn't the AB search guarentee that ANY move that the computer selects is
: at least as good as one it previously considered but now rejects?
yes...
but the problem is.. alpha/beta/minimax is a zero-sum game search strategy,
which assumes that both sides will play their best. If you are out-searching
your opponent badly (hint, if you were to play a brand new chess player that
barely knew the rules, would you try e4, Bc4, Qf3 on him? why?) then you
assume your opponent can see everything you see, which might not be right.
and if you are parrying threats that are very deep, you might give up something
to parry a threat that your opponent had no hope to see... that's the problem
I am interested in, because I see it often enough to know it is a problem...
Don Fong
In article <5he2of$a...@juniper.cis.uab.edu>,
Robert Hyatt <hy...@crafty.cis.uab.edu> wrote:
[...crimes of auto232 deleted...]
>A reasonable comment. I've only worked on this off and on for a year. It
>works. It doesn't... at least, recently, I began to get the idea that the
>auto232 TSR has some sort of ugly timing requirements built into it that are
>causing problems. In the middlegame, crafty was working great with Rebel,
>because I don't check for input too often, which introduced a delay that
>seemed to satisfy this beast. But in the opening, the book is too fast on
>a good P6/200... and with tablebases (not used in the Rebel matches) it is
>also too fast in the endgame.
why can't auto232 be fixed? or better yet, pursue Wuellenweber's
proposal for an improved autoplayer interface? IMHO it is a waste
of time to try to second-guess timing related bugs in 3rd party s/w.
it can also lead to some very embarrassing situations.
mclane
hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
>: >
>I was suffering from cosmic rays when I wrote that. :) What I meant,
>was that for every 4 games Crafty loses to humans on ICC, one of those
>games is lost for no reason, because Crafty saw something the human didn't
>see, and avoided it as only a computer can, generally by tossing material
>to the 4 winds. Losses to humans on ICC is *rare*. So this is even
>rarer... but it's important that it be stopped one day...
That is not only chess, life works the same.
The logically part of the brain extrapolates the future. And the data
to create this image of the future is very rough.
Therefore the future will be totally different.
Rationality is not working in life.
So it is not working in chess.
mclane
pe...@bitsource.com (Peter Kappler) wrote:
>One game proves very little...
One game consists of many positions.
>--Peter
mclane
brucemo <bru...@nwlink.com> wrote:
>> Ed has not the time to for that much problems, he must write Rebel 9
>> to beat Hiarcs 6 :-)
Yes, Ed has to fight against Hiarcs6.........
>Of course. It involved a lot of time to get Crafty + auto232 working.
>But, in the end, it was working, wasn't it?
?!?!?!?!? I don't think it works 100 %. Sometimes it stops.
>bruce
It seems programming the autoplayer device is heavier than programming
a chess-program....
Robert Hyatt
Rolf Tueschen (TUESCHEN.MEDIZ...@t-online.de) wrote:
: hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
who's crying?
: As far as we discuss science I may be allowed to add the following. The
: argument of Ed was crystal clear before the event. It hadn't been necessarily
: be proved by a ten games match as discribed. Even if Crafty would hold the
: score in his favour the former argument still would stand firm. I'm even not
: convinced that the match in this form was the best way to test the thesis.
: But --- but, after you gave your consent to this, you must deal with what
: happened now. Crafty *played* from a standpoint of a real chessplayer NOT on a
: level comparable of 2500 ... But more on that of a human Elo of 2000 maximum.
: Wheras Rebel played the *200* *more* to keep the opponent safely down ...
I don't disagree at all. However, I also don't think that Crafty always will
play quite so badly either. As I pointed out before, in the "Kup" they are
dead even playing at about 1 hour per move on a P6/200... If you think that
Crafty can't play chess, that's your opinion. If you think it can't beat
rebel in even-hardware games, again that's your opinion. You can easily flip
over to ICC and search through the database there. You might be a little
surprised at what you find however... although you won't find any crafty
vs rebel games longer than 60 60, which is roughly 3 minutes per move or
so. Seek and ye shall find. There's plenty of data around... I'm not
embarassed by any of it at all...
: Since Crafty played with this outrageous deep time space we could hold the
: following results:
: Theoretically a deeper search of 4-6 plies should be a terrible advantage.
: IF, and only if, if the setting of the evaluation is fitted to the *deeper
: chess space*.
agreed... always have agreed with this...
: For the time being, your prg seems to lack all that. But don't be sad, DB has
: the same problem. It is the problem -- as I wrote several times -- of the not
: yet prepared thousands of chess patterns that a real grandmaster of >2600
: disposes of.
sorry, but it doesn't. But it will eventually show us just how strong it
is. I suspect you'll be surprised...
: The other problem remains unsolved: how did Ed do his work in his limited deep
Ed Schroder
From: hy...@crafty.cis.uab.edu (Robert Hyatt)
> Seventy emails alone from Bob !!! Don't you see Ed's point?
> Ed has not the time to for that much problems, he must write Rebel 9
> to beat Hiarcs 6 :-)
: At least half of those were direct responses to questions by Ed about
: various things. For example, about 10 messages were concerned with a
: problem Ed found in Rebel with "permanent brain off". He does some
: timing analysis while thinking on the opponent's time, and this code
: was not being exercised in this match, which was screwing up Rebel's
: time allocation somewhat. We had a discussion about what to do, as
: Crafty was having a similar problem in using time it thought it would
: make up by thinking on the opponent's time, but which it couldn't make
: up since it was not thinking on the opponent's time. I simply told Ed
: to do whatever it took to make Rebel play correctly. If he needed to
: stop/restart every 10 moves, or if he needed to manually adjust the
: clock every 10 moves, whatever was needed. We were both on "unseen
: ground" at this point. I elected to ignore the problem in crafty,
: since it had a big time handicap. I didn't think that was reasonable
: for Rebel since it only had about 4.5 minutes per move or so, and
: any foul-ups would cause problems.
Correct, Rebel's time control is based on the thought that the
permanent brain always is ON. Never thought about the possibility
that playing a match with the permanent brain (PB) is OFF would
become an important issue... :)
Without PB=ON Rebel isn't able to build and maintain his favorite
amount of spare time since the spare time will only decrease as the
game continues.
Bob and I talked about possible solutions such as a program change
on our time control in case PB=OFF. We agreed that it would be not
a wise thing to do at that moment since such a change needs tuning
for a very long time and that it was most likely we would make
the time control only worse.
So we agreed on the most easiest solution to do it manually whenever
it was needed for both sides.
: The problems weren't *all* mine. :) Only most of 'em were... the
: high email traffic was generated to try and be sure that we were both
: in agreement on solutions for every little snag that came up. so that
: the results of the match would be valid.
Stop excusing in your postings! It was a pleasure to cooperate with
you! Crafty was never 100% tested for auto232. As you know Murphy's
Law is cruel in these cases. Let's blame Murphy instead... :)
: Whether that took too much of his or my time really wasn't an issue
: then, it might have become one later, although I had made the
: committment to do what I could to make this happen. It wasn't easy
: for either of us, that's for sure...
Right, your tempo was amaaaaaaaazing... :)
- Ed Schroder -
Robert Hyatt
: >Rebel played very well. However, in fairness with this position, I
: >also tested with genius as well, and the funny part is, from the white
: >side, given even fairly short times to move, Rebel, Genius *and* Crafty
: >would follow the same path. White simply ended up with a good position
: >after the first 10-15 moves, and the rest of the white plan almost played
: >itself. Of course, as you point out, white had to reach that position,
: >and black had to allow it. Black was thinking a little too far toward
: >the endgame, trying to preserve a winning advantage, but, as a result,
: >didn't make it through the middlegame unscathed.
: Maybe you should not allow it to much too think that early about
: ENDGAME.... :-)
: >: the other real question (:-) is how to model your opponent's
: >: mind - how to decide whether to take the gamble.
: >Yep. and as several AI well-knowns have said over the years, this is a
: >crucial step if minimax is going to work. Because at present, minimax doesn't
: >work exactly like we do in this regard. If you start off with a good position,
: >the extra depth should guarantee a win. If you start off with an inferior
: >position (as black does, often enough) then the extra depth might not help,
: >unless you can figure out what your opponent sees and doesn't see...
: Gambling is one main feature of ChessSystemTal.
<snip>
: Maybe Berliner should have asked Dave Kittinger how to gamble right.
: I have seen Superconny GAMBLING very often. It was NOT obvious to see
: why it is doing this and that. Maybe you always underestimate the
: professional chess-programmers. Of course many things they produce are
: not always PLANNED that accurate, but the results are often better
: than any idea that is done at the universities.
: Superconny was succesful in gambling with less rom, less speed and
: less ram in 1984 ! Maybe Berliner never tried good enough ....
But it wasn't successful against computers. That's the main point. It played
great sacs against humans. Marty played about as wild a king's gambit line against
Cray Blitz a couple of years ago as has ever been played (with Mchess Pro at an
ACM event.) The result was still predictable. You can gamble when your opponent
is likely to make a mistake. You'd better not when he isn't.
Dave tuned up Wchess a good while back and it started playing moves like the
old supercon on ICC, and it's rating went way up, but against Crafty and Ferret
it wasn't working at all, because they simply don't panic and fold their tent
when you see Bxh7+, Kxh7, Ng5+ Kg6. They could search deeply enough to see
that the king was exposed, but not fatally so. You simply can't afford that
type of gambling against an opponent that doesn't feel pressure.
I need to show you an old newspaper article I have from the 1981 Mississippi
Closed State Championship that Cray Blitz won. The master it beat in round 4
was interviewed after the game and complained about how the machine took all
the "psychology" out of the game. It didn't get disturbed by his attack,
it ignored his slamming the pieces (ala Walter Browne) and banging the clock,
and simply methodically tore him to pieces. So that style of program might
work against humans, and against some programs. Don't try it against Fritz
and programs that are decent searchers however, or else the evaluation had
better be "precise" and not "speculative". Speculative is often a synonym for
loss in computer vs computer games.
Robert Hyatt
Robert Hyatt (hy...@crafty.cis.uab.edu) wrote:
: : >>
: : >> the other real question (:-) is how to model your opponent's
: : >> mind - how to decide whether to take the gamble.
: : ABOUT GAMBLING:
: : Computerschach and Spiele, edition 1/1985,
: : article of Frederic Friedel, page 28 and 29:
: : Game Human (ELO 2265) vs. Cray Blitz , shown by Ken Thompson, that
: : Blitz has played in open championships 1982 in Missisipi.
: : Position:
: : W Ke3 Ra8 Bf3 a4
: : B Ke5 Rf7 Be6 a7 b6
: : Black (Cray Blitz) to move.
: : 47...Rxf3 ! (+4.093 evaluation)48. Kxf3 Bd5+ 49.Ke3 Bxa8
: : 50.a5 ?!
: : 50...Be4 !! (evaluation more than 10 pawns) 51.Kd2 Kd4!! 52.Kc1 b5!!
: : 53.Kb2 Kc4 54.a6 b4 55.Ka2 Kc3 and Blitz says mate in 6.
: : Frederic mentioned that Cray Blitz played Be4 in the 50th move
: : because it saw the promotion in this line, and it saw NO promotion in
: : the other line.
: : Therefore Cray Blitz played Be4 without *knowing* about the
: : wrong-color-bishop-endgame.
: Frederic was wrong. CB had RP + wrong bishop code then. In fact, ken
: thompson ran this same position against belle and sent me a note after
: the game that he was impressed that it knew about this, because the
:
By the way, the "gambling" is not quite what you suspect. I thought
about this and decided to run this position thru crafty just to see what
it would do. here's the result...
Crafty v11.19
+---+---+---+---+---+---+---+---+
8 | R | | | | | | | |
+---+---+---+---+---+---+---+---+
7 | *P| | | | | *R| | |
+---+---+---+---+---+---+---+---+
6 | | *P| | | *B| | | |
+---+---+---+---+---+---+---+---+
5 | | | | | *K| | | |
+---+---+---+---+---+---+---+---+
4 | P | | | | | | | |
+---+---+---+---+---+---+---+---+
3 | | | | | K | B | | |
+---+---+---+---+---+---+---+---+
2 | | | | | | | | |
+---+---+---+---+---+---+---+---+
1 | | | | | | | | |
+---+---+---+---+---+---+---+---+
a b c d e f g h
Black(1): move
end-game phase
clearing transposition table
clearing pawn hash tables
time limit 51.42
depth time score variation (1)
1 0.00 1.266 Bd5 Bxd5 Kxd5
1 0.00 1.290 Bf5
1-> 0.00 1.290 Bf5
2 0.01 1.130 Bf5 Re8+ Be6
2 0.01 1.142 Rc7 Re8
2 0.01 ++ Rxf3+!!
2 0.01 4.964 Rxf3+ Kxf3 Bd5+ Ke3 Bxa8
2-> 0.01 4.964 Rxf3+ Kxf3 Bd5+ Ke3 Bxa8
3 0.02 4.964 Rxf3+ Kxf3 Bd5+ Ke3 Bxa8
3-> 0.09 4.964 Rxf3+ Kxf3 Bd5+ Ke3 Bxa8
4 0.10 4.954 Rxf3+ Kxf3 Bd5+ Ke3 Bxa8 a5
4-> 0.10 4.954 Rxf3+ Kxf3 Bd5+ Ke3 Bxa8 a5
5 0.19 4.814 Rxf3+ Kd2 Rf2+ Kd3 a5 Re8 Rf3+ Kc2
5-> 0.20 4.814 Rxf3+ Kd2 Rf2+ Kd3 a5 Re8 Rf3+ Kc2
6 0.36 4.984 Rxf3+ Kd2 Rf7 Ke3 Rc7 Rd8 Bd5
6-> 0.44 4.984 Rxf3+ Kd2 Rf7 Ke3 Rc7 Rd8 Bd5
7 0.57 4.960 Rxf3+ Kd2 Rf7 Ke3 Rc7 a5 bxa5 Rxa7
7-> 0.65 4.960 Rxf3+ Kd2 Rf7 Ke3 Rc7 a5 bxa5 Rxa7
8 1.42 4.978 Rxf3+ Kd2 Rf7 Ke3 Re7 Rf8 Bd5 Rf4
a6
8-> 1.61 4.978 Rxf3+ Kd2 Rf7 Ke3 Re7 Rf8 Bd5 Rf4
a6
9 5.26 5.128 Rxf3+ Ke2 Bc4+ Kd2 a6 Rc8 Rd3+ Ke1
Kd4 Rc6 Re3+ Kf2 Re2+ Kf3
9-> 5.77 5.128 Rxf3+ Ke2 Bc4+ Kd2 a6 Rc8 Rd3+ Ke1
Kd4 Rc6 Re3+ Kf2 Re2+ Kf3
time: 8.33 cpu:86% mat:4 n:789633 nps:109519
ext-> checks:45140 recaps:778 pawns:393 1rep:15279
nodes full:224143 quiescence:565490 evals:565110
endgame tablebase-> probes done: 1053 successful: 1053
Black(1): Rxf3+
time used: 8.56
White(2): Kxf3
time used: 1.48
time limit 52.00
depth time score variation (1)
1 0.00 -2.444 Bc4 Rxa7
1 0.00 -0.824 a5
1 0.00 4.964 Bd5+ Ke3 Bxa8
1-> 0.00 4.964 Bd5+ Ke3 Bxa8
2 0.00 4.964 Bd5+ Ke3 Bxa8
2-> 0.00 4.964 Bd5+ Ke3 Bxa8
3 0.01 4.954 Bd5+ Ke3 Bxa8 a5
3-> 0.01 4.954 Bd5+ Ke3 Bxa8 a5
4 0.01 ++ Bd5+!!
4 0.02 5.258 Bd5+ Ke3 Bxa8 Kd3 Be4+ Kc4
4-> 0.09 5.258 Bd5+ Ke3 Bxa8 Kd3 Be4+ Kc4
5 0.09 5.258 Bd5+ Ke3 Bxa8 Kf2 Be4 Ke3
5-> 0.10 5.258 Bd5+ Ke3 Bxa8 Kf2 Be4 Ke3
6 0.10 ++ Bd5+!!
6 0.13 6.060 Bd5+ Kg3 Bxa8 Kf2 Bc6 a5 bxa5 Ke3
6-> 0.13 6.060 Bd5+ Kg3 Bxa8 Kf2 Bc6 a5 bxa5 Ke3
7 0.14 6.060 Bd5+ Kg3 Bxa8 Kf2 Bc6 a5 bxa5 Ke3
<HT>
7-> 0.15 6.060 Bd5+ Kg3 Bxa8 Kf2 Bc6 a5 bxa5 Ke3
<HT>
8 0.52 6.220 Bd5+ Ke3 Bxa8 Kd3 Bc6 a5 bxa5 Ke3
a4
8-> 0.60 6.220 Bd5+ Ke3 Bxa8 Kd3 Bc6 a5 bxa5 Ke3
a4
9 0.93 5.976 Bd5+ Ke3 Bxa8 a5 bxa5 Kd3 Be4+ Kc4
Bc2 Kc5 a4
Note that after < 1 second here, it's going to check, pick up the bishop,
and then expects black to try the cheapo a5 "gambling" move you mentioned.
It's not a gamble at all it turns out, although a beginning human player
might take it. Also notice that knowledge is not quite enough here to avoid
problems. Advance the pawns a couple of ranks, and then try this and white
wins... because the bishop and pair of pawns can seal the black king out of
the a1 queening target. Crafty relies on the knowledge that a bishop +
rook pawn(s) on the wrong rook file is a draw, but only if the king can
reach the queening square safely. It uses a search to get the king to the
right place, then the eval says "draw."
I did have to modify it some this morning, because there is a clever way
that will trick most programs into making a mistake, even if they have this
simple rule. It is possible for black to advance the pawns, and then give
up the bishop. If black is not careful, it looks like suddenly the draw is
gone, because the bishop is gone. I simply modified the code to handle
this, something that was needed anyway. It now realizes that K vs Rpawn(s)
is a draw if the king can get to the queening square first, and the opponent
has the wrong bishop (or no bishop at all).
However, there's not much magic in a program playing a5. I can guarantee you
Crafty didn't find it just to "gamble"...
9-> 1.02 5.976 Bd5+ Ke3 Bxa8 a5 bxa5 Kd3 Be4+ Kc4
Bc2 Kc5 a4
10 1.80 6.065 Bd5+ Ke3 Bxa8 a5 Be4 Kd2 Kd4 Kc1 Kc3
a6 Kd3
10-> 2.05 6.065 Bd5+ Ke3 Bxa8 a5 Be4 Kd2 Kd4 Kc1 Kc3
a6 Kd3
11 3.31 6.065 Bd5+ Ke3 Bxa8 a5 Be4 Kd2 Kd4 Kc1 Kc3
a6 Kd3
11-> 3.63 6.065 Bd5+ Ke3 Bxa8 a5 Be4 Kd2 Kd4 Kc1 Kc3
a6 Kd3
12 3.98 ++ Bd5+!!
time: 4.54 cpu:69% mat:4 n:319936 nps:101890
ext-> checks:12005 recaps:2306 pawns:1098 1rep:10739
nodes full:88445 quiescence:231491 evals:197994
endgame tablebase-> probes done: 4381 successful: 4381
Black(2): Bd5+
time used: 4.80
White(3): Ke3
time used: 2.92
time limit 52.70
depth time score variation (1)
1 0.00 4.964 Bxa8
1-> 0.00 4.964 Bxa8
2 0.00 ++ Bxa8!!
2-> 0.00 5.263 Bxa8
3 0.00 ++ Bxa8!!
3 0.01 5.844 Bxa8 a5 bxa5
3-> 0.01 5.844 Bxa8 a5 bxa5
4 0.01 5.844 Bxa8 a5 bxa5 Kd3
4-> 0.08 5.844 Bxa8 a5 bxa5 Kd3
5 0.08 ++ Bxa8!!
5-> 0.09 6.143 Bxa8 a5 bxa5 Kd3
6 0.09 6.138 Bxa8 a5 bxa5 Kd3 Bd5 Ke3
6-> 0.09 6.138 Bxa8 a5 bxa5 Kd3 Bd5 Ke3
7 0.10 6.298 Bxa8 a5 bxa5 Kd3 Bd5 Ke3 a4
7-> 0.10 6.298 Bxa8 a5 bxa5 Kd3 Bd5 Ke3 a4
8 0.18 6.000 Bxa8 a5 bxa5 Kd3 Be4+ Kc4 Bd5+ Kb5
<HT>
8-> 0.27 6.000 Bxa8 a5 bxa5 Kd3 Be4+ Kc4 Bd5+ Kb5
<HT>
9 0.32 ++ Bxa8!!
9-> 0.72 6.299 Bxa8 a5 bxa5 Kd3 Be4+ Kc4 Bd5+ Kb5
<HT>
10 1.54 6.065 Bxa8 a5 Be4 Kd2 Kd4 Kc1 Kc3 a6 Kd3
10-> 1.67 6.065 Bxa8 a5 Be4 Kd2 Kd4 Kc1 Kc3 a6 Kd3
11 1.76 ++ Bxa8!!
11 4.23 6.648 Bxa8 a5 Be4 Kd2 Kd4 Kc1 Kc3 Kd1 bxa5
Ke2 a4
11-> 4.72 6.648 Bxa8 a5 Be4 Kd2 Kd4 Kc1 Kc3 Kd1 bxa5
Ke2 a4
time: 6.00 cpu:80% mat:5 n:505756 nps:105146
ext-> checks:15589 recaps:3323 pawns:3718 1rep:15909
nodes full:138830 quiescence:366926 evals:251430
endgame tablebase-> probes done: 9581 successful: 9581
Black(3): Bxa8
time used: 6.26
White(4): a5
time used: 1.13
depth time score variation (1)
1 0.00 4.128 Be4 axb6
1 0.00 5.844 bxa5
1-> 0.00 5.844 bxa5
2 0.00 5.844 bxa5 Kd3
2 0.00 ++ Be4!!
2 0.00 6.674 Be4 Kd2 Kd4 <HT>
2-> 0.00 6.674 Be4 Kd2 Kd4 <HT>
3 0.00 6.674 Be4 Kd2 Kd4 <HT>
3-> 0.01 6.674 Be4 Kd2 Kd4 <HT>
4 0.01 6.674 Be4 Kd2 Kd4 <HT>
4-> 0.01 6.674 Be4 Kd2 Kd4 <HT>
5 0.01 6.674 Be4 Kd2 Kd4 <HT>
5-> 0.08 6.674 Be4 Kd2 Kd4 <HT>
6 0.08 6.674 Be4 Kd2 Kd4 <HT>
6-> 0.08 6.674 Be4 Kd2 Kd4 <HT>
7 0.08 6.674 Be4 Kd2 Kd4 <HT>
7-> 0.08 6.674 Be4 Kd2 Kd4 <HT>
8 0.09 6.674 Be4 Kd2 Kd4 <HT>
8-> 0.09 6.674 Be4 Kd2 Kd4 <HT>
9 0.09 6.674 Be4 Kd2 Kd4 <HT>
9-> 0.09 6.674 Be4 Kd2 Kd4 <HT>
10 0.09 6.674 Be4 Kd2 Kd4 <HT>
10-> 4.77 6.674 Be4 Kd2 Kd4 <HT>
11 7.20 6.884 Be4 Kd2 Kd4 Kc1 Kc3 Kd1 bxa5 Ke2 a4
Kd1 a3
time: 8.28 cpu:93% mat:5 n:828407 nps:106479
ext-> checks:18394 recaps:5561 pawns:12282 1rep:29380
nodes full:241740 quiescence:586667 evals:395255
endgame tablebase-> probes done: 15345 successful: 15345
Black(4): Be4
time used: 8.60
Robert Hyatt
: >: NPS challenge will end after game one...
: This is disappointing. Crafty losing despite a 100:1 time advantage
: is indicative of something very interesting happening, and
: makes me want to see more. More games would provide more data
: and would lead to stronger conclusions about the relative
: utility of Crafty's speed and Rebel's evaluation function.
: > I tried it, and CB
: >instantly announced a mate in 11 for the opponent. It saw that in the
: >game and sacrificed the queen to avoid it. [...]
: >Yet CB, because it just happened to search deeply enough, did see it,
: >and threw the game away because of it. [...]
: > It is simply pointing out the fact
: > that a deeper search doesn't always result in better play.
: Surely, in your example above, sacrificing the queen was CB's best
: move, since it avoided the mate?
depends on your "best move" definition. Korchnoi didn't see the mate.
The opponent didn't see the mate. In fact, he *almost* didn't take the
queen, because of suspicion after seeing Crafty mince a couple of IM's
and a GM the three games before that one.
It should have simply kept up with its attack, because losing the queen
obviously would lose the game. If the likely happened and the opponent
didn't see the clever variation to make all of the attack break down,
then it would have won easily. If the opponent did, by whatever bit
of luck we would like to imagine, find the right move, we'd have lost
anyway. But making the queen move was a guaranteed loss. *that*
was the problem I didn't like. Taking the obvious loss, as opposed to
one that is so deep that one of the top-5 GM's in the world didn't see
the stinging response...
: Surely you were just unlucky to be on the wrong side of the
: mate-in-11 and you would equally often be on the right side
: of that combination? This would mean that freak combinations
: would, statistically, not affect your results at all.
: In the Rebel-Crafty game, Crafty might again have been unlucky,
: or it might have been Rebel's evaluation function putting Crafty
: in a position where such a combination was likely to pop up. More
: games would give us more of a clue as to which of these it was.
: Bob, in another posting you discuss methods of getting round the
: problem of avoiding things which the opponent wouldn't see. This
: all comes under the general heading of opponent modelling which,
: I gather, is at least as difficult as the various types of
: learning you are working on now.
: It's very difficult for a computer to tell how deeply its human
: opponent is searching. There are some combinations which look
: deep to a computer, but are relatively simple for a human,
: being composed of one main idea but with lots of little delaying
: moves. So crafty might decide its opponent is searching to a
: depth of 6 or 8 ply (beause the opponent lost a pawn) and set traps
: appropriately, but then get completely blown away by a 10-ply
: combination which it could have avoided.
I agree. I'll have more to say on this in a few months, because I've
started working on the endgame database problem finally. The problem
where crafty plays the move leading to the deepest mate (if it is
losing) rather than playing the move that gives the opponent the most
ways to go wrong and draw or lose himself. After I get this done, we
can talk about the idea again, although it is easier knowing that you
have perfect information to make a decision on...
: More subtly, the total task facing crafty is to reproduce what
: its opponent would play in any given position. This means
: working out what its oppponent's evaluation function is and
: its search strategy.
: In theory this is impossible, since an
: evaluation function can be arbitrarily complex, and the only
: information you have to give you a clue is the move made, one
: move out of a small set of possible moves. There might be
: something lurking in the opponent's evaluation function and
: search algorithm which has not yet had any effect on the
: move choice. Each time the opponent makes a move, crafty
: acquires a small amount of extra information about the
: opponent, but does not ever know that it knows all there is
: to know about its opponent.
: In practice, it's not impossible, just very difficult. :)
right. Or perhaps do like Kasparov in game one. Offer a pawn that a
12 ply search can find, but an 11 ply search can't... if he doesn't bite,
offer one that a 11 ply search can find, and so forth.. a couple of games
down the drain could (and *could* is important, not *would*) reveal a lot.
Of course the opponent might recognize this strategy and choose to not take
the pawn to screw up your "feeling out." :)
Pete Nielsen
Robert Hyatt (hy...@crafty.cis.uab.edu) wrote:
>
> This is almost backward. The issue is this: suppose you search very
> deeply and find a position you like and you believe you can reach. And
> with best play you are going to stand better at the end. Then, just before
> you move the piece, you decide to go over it one more time, and this time,
> way out at the end, you find a really obscure move your opponent can make
> that breaks the position wide open. Now you have two choices: (1) go ahead
> and play the move you were going to play, not knowing if your opponent will
> find the winning move or not since it is so deep in the tree; or (2) you
> now start trying to find ways to thwart that move your opponent "might" play,
> but since you are a computer, you assume he will see it, and the only way to
> avoid getting busted completely is to give up two pawns.
>
> As a human, I'd think that losing two pawns is probably losing the game,
I'm still confused by this: "The best move is not the most effective move"
What about a thought experiment. Suppose given the time limits and game
history the program could predict how many plys it's going to be able
to search, would it make sense to weight each successive ply so that it
would have to be some factor that increeased with ply deapth better than
the current favorite in order to replace it. Would that address the problem?
Jack Nerad
>
> > This sword has to cut with both edges. If you get better with a
> > 3x performance increase, so does everyone else. And if you get
> > better with a 3x performance increase, someone else is going to get
> > a lot better with a 100x performance increase. I don't see how you
> > can argue otherwise...
>
> for Rebel. I believe the following is true for Rebel:
> - A doubling in speed is good for a 50-75 ELO gain.
> - P90 to P200 is +50
> - P90 to PP200 is +75
>
> Let's take the average of +60.
> Then I would split the +60 into:
> - +45 for finding better POSITIONAL moves;
> - +15 for a better tactical understanding;
>
> That's why I believe knowledged based chess programs do have a better
> future than speed based chess programs. Of course I can't prove this
> only time will tell.
>
> - Ed Schroder -
If this is the case, I really don't understand why you two argued about
NPS in the first place:
If you take a program and run it on a computer half as fast as another
for twice as long as on the other, you could expect the same result.
Given this, it would imply that total nodes searched is the key.
Therefore:
Time (t) = Rebel 8.0 Elo Strength (Elo)
2t = Elo + 60 (average, see above)
4t = Elo + 120
8t = Elo + 180
16t = Elo + 240
32t = Elo + 300
64t = Elo + 360
100t = Elo + ~393 (linear interpolation).
Just imagine: in 8-9 generations of doubling (How many years is that,
14? 15?) Rebel 8.0 will be able to *Crush* world champions with ease.
How many times faster than a micro is Deep Blue? 1250x? If IBM had the
Rebel 8.0 Engine, it would smear Kasparov -- that's Rebel 8.0 + 600 Elo
points -- at least according to Mr. Schroder's estimation of increased
strength based on speed doubling. 3000 Elo. Wow.
Is this sarcasm? Must be my math.
Regards,
Jack Nerad
Ed Schroder
From: hy...@crafty.cis.uab.edu (Robert Hyatt)
[ snip ]
: What you described above happened in a *lost* position.
: That's what makes the big difference.
: What's the point to mention it?
: Ink clouds?
> I never tried to make an excuse for how it played, Ed. I pointed out
> however, that it made some moves that it would not have made at a
> shallower search.
Ok, let's not argue here any longer...
> The game was *not* lost at move 15.
[ snip ]
: Yup, but spelling everything what went wrong here in RGCC implied
: the 9-3 doesn't mean anything because of all the Crafty bugs.
: That's the difference between you and me, I in your case would have
: said nothing here and would hide under my bed and stay there for at
: least one week... :))
> I didn't spell everything that went wrong. If you back up and check
> the posts, I didn't claim that the bogus time control stuff resulted
> in that lopsided score. I didn't claim that 12 blacks in a row did
> it either. I did explain what I had done, because I got several
> email queries about "Bob, my version of 11.18 won't play that same
> bad move at any depth. Is the 11.18 in the match the same or is it
> different?" I explained that the programs were the same, but that
> the contempt factor was broken. For all I know, playing those 12
> games again might result in a 12-0 score.
We will not agree here too... :)
Let's forget about it.
> I didn't object to your "doubles elimination." Which is not something
> I agree with, by the way, because I spent about 3 months working on the
> book learning stuff. If crafty gets a good position, it would love to
> play that opening again, but the double elimination would cut the game
> off. That's one way to solve a problem. There is another way, which
> is to do what I've done, and what others have done. Otherwise, the
> results you get won't match the results the SSDF gets and it won't
> match the results that people get when they operate Rebel on a
> server manually.
True, but my opinion on SSDF double game policy should be well known
by now.
> If I wanted the best possible result, I'd have insisted that we do the
> testing just like the SSDF, so that Rebel has to avoid repeating games
> it loses, while Crafty would work like heck to play what wins and avoid
> what loses. That wasn't the issue here. The issue was the NPS games
> which might have provided some interesting data. The first one did,
> for me anyway, because I did see a couple of places where Crafty was
> dodging a bullet that wasn't going to be fired. That's the kind of
> data I was hoping for, although I'd also like to see it win more games
> than it loses too. :)
: : Rebel played well. I think the most impressive thing I saw was
: : the fact that it did not seem to make *any* tactical mistakes,
: : anywhere. Which is surprising.
: No it is *not* surprising!
: You still miss my point...
: After 25.h5 gxh5?? Crafty had a lost position. No 80-90 ply search will
: help *AS LONG AS* white makes no mistakes. Just putting the white pieces
: on the right squares, that's all there is. I believe it was Capablanca
: who said this and his words are burned in my memory.
> Tell you what. I'll watch more carefully on the servers, and I'll
> post a game or two where Rebel makes some horrible tactical mistakes
> against Crafty. Perhaps you'll change your mind. Rebel plays well.
> It doesn't play perfect, any more than any program does. BTW, a
> couple of your commercial "cousins" also play gxh5... one very
> highly ranked...
Maybe, for Rebel8 I noticed it needs one minute to play 25.. e4 with
a 0.25 better score than 25.. gxh5 and incredible wild tatics while
analyzing further. Would have been very interesting!
Same interesting moment at 46.. f6
Rebel would have played 46.. Na4 with a 0.50 better score than 46.. f6
Same incredible tactics. To me Rebel can hold the position but...
: That's how chess works, that's what I believe in. Searching one ply
: deeper for 80% is meant to find better positional moves (this is my
: own personal definition). Tactics will come by itself. Game1 (till now)
: is a perfect example of this.
: If you have build a good position you usually don't have to fear for
: any counter tactics since you have all the aces. It's a common rule
: in chess. Will we ever agree on this? I wonder... :)))
: The Pentium Pro 200 is a fantastic machine. The tactical barrier becomes
: smaller and smaller and finally the "best" evaluation function will
: decide a comp-comp chess game. And this trend will go on as soon as new
: speed monsters will be released by Intel.
> I assume you won't be needing the next P6/400 if you believe that?
> Since another ply isn't going to help?
I did not say this, see above. For my opinions about "Chess programs
and computer speed" see the related subject on my home page.
> I've personally never moved crafty to a faster machine (or cray blitz
> for that matter) and not found that it played better chess. So the
> depth did something, otherwise why are you testing on a flock of
> P6/200 machines, rather than the more affordable P90's you had?
Speed does a lot with chess programs but it depends per chess program.
> This sword has to cut with both edges. If you get better with a
> 3x performance increase, so does everyone else. And if you get
> better with a 3x performance increase, someone else is going to get
> a lot better with a 100x performance increase. I don't see how you
> can argue otherwise...
I can't speak for others since I don't have the source code. I can speak
Amir Ban
I thought the object of the NPS challenge is to get everyone to shut up
about this. That's certainly not been achieved.
I don't understand what's happening here. Seems to me that everyone has
an interest in the continuation of this match:
Ed, because he is getting all this exposure, making good so far, and
proving a huge point that he wants proved. C'mon Ed ! We'll buy all
your future versions, Windows-95 bugs notwithstanding.
Bob, because he is still the favorite to win this match, and anyway can
only improve. More importantly, he is getting some excellent lesson on
things that must be worked on and will not disappear by a 2% speed
improvement. More generally, this is a lesson on what chess is really
about. This is what humans tried to tell us for 40 years but many of us
were not listening.
The rest of us, because it's great entertainment.
Just in case your missing my two cents here:
Of course a Rebelx100 - Rebel match or Craftyx100 - Crafty match won't
prove a thing. The x100 program will win big. Doesn't have anything to
do with the evaluation-NPS debate.
On the issue of the deeper search sometimes being a disadvantage: This
is a total red herring to my mind. It's theoretically possible,
practically insignificant. If Crafty is in the habit of losing games
for this reason then it is suffering from a subtle variation of the
horizon effect. Bad evaluation also contribute. This simply does not
happen with the better programs. Best example is probably Genius. It
may run into trouble in a game like the rest of us, but I have never
seen it commit suicide at any time control.
Amir
Rolf Tueschen
brucemo <bru...@nwlink.com> wrote:
>Ed Schroder wrote:
>>
>> NPS challenge will end after game one...
>>
>> Today I decided to end the NPS challenge between Crafty and Rebel8 after
>> game-1 is finished.
>A shame, especially since it ends in your favor after one game. When you
>start a ten-game match, you should play a ten-game match, and not play a
>one-game match.
No, I don't think so. If you go to Nam, and throw millions of tons of napalm
over there you'll always could stop this, no? I mean we could see a lot of
nose/mouth clifts still today not being operated properly. It's a shame.
Sorry, but that's off topic, I must tell you.
>> Understanding the problems of running a match with an unfinished
>> product, and since there are so many things that Bob may claim and so
>> many things that did go wrong with the book, evals, time control,
>> contempt factor, huge auto232 problems, bugs etc. I consider it a waste
>> of my time to continue the match.
>Both of you guys spent a long time working on this. I got on the CC line
>of the emails going back and forth, and there were a lot of them, from each
>of you. I have sixty or seventy email messages on this topic from Bob
>alone.
Hey Whaaats goooing on there? Asked the policeman, holding open his right hand
for the coming bakshish.
Sorry, but that was all off topic, I must seriously tell you.
>These were mostly about bugs in Crafty that surfaced due to the use of
>Auto232, and because Crafty doesn't have a debugged install procedure.
>Most of the auto232 bugs seem to have been in the "control" match.
>On ICC, Crafty is started, plays a game, then is terminated. With auto232,
>it plays a game, then is told to start a new game, etc. I just tried to
>call Bob to ask him for sure, but he wasn't home.
Not possible at all. I just read many postings of Bob. And if he were just one
second out for sniffing a biot of fresh air, that wouldn't been possible. Just
prove this for yourself. It's simple maths.
>But the bugs that I know
>of were, 1) after the first game it thought the opponent was out of time,
>so it changed its contempt factor to -1.00 pawns, 2) for each successive
>game after the first game, it used less and less time per move in the
>secondary time control. It sounds like there were other weird things going
>on as well.
But these are not MY fault, no?
>I think you were very patient to put up with the auto232 bugs. I am not
>sure what caused your patience to run out. Aren't the problems solved?
>I asked Bob what was going on when I noticed that the score in the control
>match was 6-0,
Did you forget the remis score?
>and he told me about the secondary time control problem and
>the contempt factor problem.
>It is obvious that this could have had an affect upon some games. I asked
>Bob
Again. Hey what's going on with Ferret? Could you lend Bob some of your 3500
tools?
>if he knew that it DID have an affect upon any of them, and he told me
>it looked like Crafty made at least one really stupid move, due to low
>time, in an otherwise even position. He may have mentioned another game, I
>don't know.
>But it wasn't like Bob was whining about this,
Thanks, I always wrote the verbium without *h*. :)
>I had to ask.
Ahem. Ok. I see. Perhaps you'll better try to ask ME for more...
Brucimo, you knew long before, that we could easily work together. Just like TC
and CW, no?
> I didn't see
>any mail from Bob that I was CC'd on where he suggested ignoring those six
>games. If you ignore them, they can be replayed in a couple of days. If
>you keep them, big deal, it's just six games out of 60, and maybe you would
>score 4-2 or 5-1 or 6-0 again if you replayed them, so they aren't very
>meaningful.
>After various bugs were fixed, Rebel and Crafty split the next six games,
>3-3.
>Crafty blew out of book early in the NPS game because its book was
>write-protected on the CD, but you guys both agreed to continue, so who is
>to complain?
Try me, you brute.
>> Besides of this I also want to express my feelings about the following.
>>
>> I am interested in the scientific relationship between ply depth and an
>> evaluation function. That's what has started this match. Sofar game-1
>> has proven my point of view at least in my opinion.
>You can't claim victory
Who did claim victory, my young friend brucinello?
Nobody did this as far as my 600 Pentium brain is telling me.
>after one game in a ten-game match. I assume there
>was a reason you guys agreed to play a ten-game match. Perhaps Bob would
>not have agreed to a one-game match, with black, but that isn't what you
>guys agreed to.
No, for this time, don't try one of your famous papa Albrighty diplomatic royal
manouvre. Speak it out what you have on your liver.
>Declaring victory and pulling the plug after one game is really bad,
>especially after playing white in every control game as well.
But remember, the other method. After winning the 200 and 400 meters in Atlanta
your famous athlet fell seriously ill. he couldn't start in the european
meetings. Guess why??? --- I heard rumour he was very ill after setting down
his medical treatment, if you know what I mean. He rn so fast to the toilette
that he couldn't stop in time at the right door. So he had this terrible pipi
problem ...
Excuse me, that was all very, very, very off topic, madame. :)
>> I am *not* interested to read all kind of possible and impossible
>> explanations that imply to hide the real reason why Crafty will lose
>> game-1.
>Crafty lost game one because it didn't play well. Rebel won game one
>because it played well.
>There can be speculation about why this happened. Some of it may revolve
>around mistakes Bob made, rather than your own super-competence.
Heile heile Gänschen, es Kätzchen hat e Schwänzche. Your much too much
agitated, Brucimo. Relax. Breath slowly. Yeah. Do you feel it ....
>I think Bob made a serious mistake with his king safety routine, which he
>should fix.
Helas, yes, I could bet 500 000 $ on that one. But remember, DB surely had the
same problem. Bet accepted??
>There is probably more to it than this, certainly, and I'm
>sure a lot of the credit goes to Rebel for playing the position well.
>Perhaps you could explain WHY it played so well?
You are a naughty boy, you bruce.
>With out without this fix, I would like to see game two, especially if
>Crafty gets to play white.
>> For instance, Bob Hyatt wrote:
>>
>> : I'll post some interesting analysis later, but you might be surprised
>> : at some of what went on in this match. For example. in at least two
>> : places, Crafty played a weaker move, simply because it saw, at
>> : depth=15, something that it thought Rebel would play and it didn't
>> : like it. In these two cases I know of, I've tested rebel on the
>> : position and it wouldn't have played the move Crafty feared. So
>> : Crafty simply avoided something that would not have happened, by
>> : playing something that was worse. It was one of the common scenarios
>> : I've seen when two programs play and one outsearches the other by
>> : a significant amount.
>This isn't an excuse, it is analysis. I'm not sure if he's right or not,
>but the idea that a program may kill itself via its own search depth is
>something that Bob has been interested in.
My God, he will not follow those Sonnentemplers, no? :) Just heard of 49 killed
in Diego ...
>Bob of all people has to be able to comment honestly on the games, even if
>this may involve questioning the idea that Rebel is the best program on the
>planet.
You apparently forgot the universe, including Hale Bobb, eh Bopp.
>> This would mean Crafty / Rebel would come into trouble against the
>> Mephisto I, Boris etc. because Crafty / Rebel think 4-5 plies deeper.
>> This is of course total nonsense. Crafty as any other program just
>> played the best move on a given depth.
>On ICC I played a rated 5 0 game, Ferret vs Spinach. Spinach is Ferret at
>50 milliseconds per move, no pondering, on inferior hardware. So this is
>more than 100:1 time odds.
Before you get crazy about this, may I remind you of the following. We have a
macro and we have a micro world of chess computers. Your experiment dealt with
micro whereas the two other heroes played a macro game. By shortening the
timetable you surely could get a factor 100 ratio but one had to ask -- what
does this mean in terms of chess playing. Soon you'll do something comparable
in the anti-world. And what shall we do if you were lost in one of these black
holes? Bob has no telephone or is out fishing. What should I do without you?
So, please, you'll always carefully prepare parameters in advance.
>I only did this once, as it really scared me to watch Ferret play at -2 for
>most of the game. It found a tactic later on and won, but I'd have been
>really pissed
No, not this again, I mean we have women here around. Please behave like one.
>if it had dropped 32 rating points because of this little
>lark.
But this would clearly be in the statistical marge of error, no?
>Why do you think Spinach was up two pawns against Ferret? Do you think it
>out-searched it at 100:1 time odds? No, Ferret came out of book bad,
>couldn't negotiate the position,
Do you remember the young boy in Grisham's The Client. You write with as much
suspense.
>found a combination against for the
>opponent, and mitigated the problem by dumping material at the root.
>That this may have happened doesn't take anything away from Rebel's win in
>round one, of ten. As I have stated, Rebel played well and Crafty didn't.
> It doesn't do a program credit to get into a situation where it is finding
>combinations against itself, assuming that is what happened here.
>> I wouldn't dare to mention such an argument if Rebel would have had
>> the 100 x time advantage!
>It's not a particularly big deal. Self-criticism after losses doesn't take
>anything away from the opponent. "He beat me because he's god", doesn't
>offer the loser much hope.
You see, now you came back to the more universal stuff.
>> This is just one example...
>>
>> I am not in the mood for this kind of explanations for another 9 games
>> the next 5-6 months. My time is too precious for that.
>>
>> Rebel8 played this game on an average ply depth of 9/10.
>> Crafty played on an average ply depth of 13/14.
>>
>> Still Rebel8's evaluation function was good enough to handle the big
>> ply gap of 4 plies. This has always been my point and game-1 has
>> proven it. I believe this also counts for other chess programs like
>> for instance Hiarcs etc.
>Rather than declaring the match over because you won game one, please come
>back and play the rest of the match.
Look, if Bob only once had written some parts of what you did add now ... But
for me he wrote in all his posts how bad all went technically ... not at all
seeing some implications the event had for the former debate yet!
In another post I tried to explain the game as a good example for Rebel's
strength. And I don't think that I'm known for biased sentimental judgements,
no? It was really a nice game by Rebel. Period. :)
>bruce
Rolf Tueschen
Peter Herttrich <dh1...@inss1.etec.uni-karlsruhe.de> wrote:
>In rec.games.chess.computer Ed Schroder <rebc...@xs4all.nl> wrote:
>: NPS challenge will end after game one...
>: Understanding the problems of running a match with an unfinished
>: product,
snip
>: Sofar game-1
>: has proven my point of view at least in my opinion.
>: For instance, Bob Hyatt wrote:
>: : I'll post some interesting analysis later, but you might be surprised
>: : at some of what went on in this match. For example. in at least two
>: : places, Crafty played a weaker move, simply because it saw, at
>: : depth=15, something that it thought Rebel would play and it didn't
>: : like it. In these two cases I know of, I've tested rebel on the
>: : position and it wouldn't have played the move Crafty feared. So
>: : Crafty simply avoided something that would not have happened, by
>: : playing something that was worse. It was one of the common scenarios
>: : I've seen when two programs play and one outsearches the other by
>: : a significant amount.
>: This would mean Crafty / Rebel would come into trouble against the
>: Mephisto I, Boris etc. because Crafty / Rebel think 4-5 plies deeper.
>: This is of course total nonsense. Crafty as any other program just
>: played the best move on a given depth.
>Ok! If someone spend me an auto232, i use my two 486DX4/100
>to make the challange. I own an Rebel8 and i have enough experience
>in use of Crafty. Both boxes have 32MB RAM so 12 or 13MB hash is
>no problem.
>Bob, if you tune a little bit the rook-problem (my last posting)
>or what u want (Crafty 11.19 or so), i would like to see, who is
>right, the one or the other side.
That was my point. After changing the versions you could test what you like but
not the prosecution of the former match. :)
>By the way, from MY experiences and IMHO a ratio 1:10 in time is
>enough for this experiment.
>Bob, another question: Is Crafty able to play under LINUX through
>the RS232 and contact the auto232 at the DOS-machine?
>It's not an absolutely must but it would be better because of the
>memorymanagment in LINUX.
>
>I don't comment Ed's posting.
Why not? What is your opinion?
>But i agree, if someone needs it's
>time to make money, such an experiment cost! money :-)
>cheerio
>Peter
------------------------------------------------------------------------------
Peter, don't fall from your chair please. Could you explain for me how I could
load the crafty.doc with all the Befehle?
Everytime i try to d/l it system tries to search for *winword* (!) but couldn't
find it because I have Word. Now what could I do?
But I must admit I never tried from your mirror because I don't have the
address. Either give me this address with an easy method to d/l or something
else. I think I have not seen a lot of interesting commands yet ....
Hey I forgot, I need W95. Not Linux. That seems to be even more difficult for
me. :)
Thanks for having done all the work with Crafty.
Rolf Tueschen
>_____________________________________________________________________________
> Peter Herttrich email: dh1...@inss1.etec.uni-karlsruhe.de
> Universitaet Karlsruhe Tel. +49 721 6083747 FAX +49 721 6086071
>Institut fuer Nachrichtentechnik ..life outside caves is complicated ...
>_____________________________________________________________________________
Mark B. Stucky
On 27 Mar 1997 15:34:30 GMT, hy...@crafty.cis.uab.edu (Robert Hyatt)
wrote:
>
>If it were only so easy. The problem is this:
>
>the "protocol" between Crafty and the auto232 TSR program is not that
>difficult. So far as I know that works perfectly, although there is at
>least one serious design flaw that might affect a game here and there,
>in that promotions *have* to be to a queen. There's no choice. If
>Crafty were to try to promote to a knight, the game will hang at that
>point.
>
>However, the protocol is not the issue. When the opponent makes a move
>and "my" auto232 TSR stuffs the move into the keyboard input buffer so
>that Crafty will think the operator typed the move, there is a very
>definite "timing window" where crafty has to read this move and echo it
>to auto232. This is a pain. You can't echo it too quickly, or it will
>hang up. You can't wait too long, or it will hang up. You have to do
>it within some as of now unknown timing window. This is not easy to do
>in DOS. In unix, I'd set an alarm and go about my business. In DOS,
>such signals don't work the same, and most compilers don't even support
>such things (DJGPP for example, the compiler we use for DOS).
What version of DJGPP are you using? If I remember correctly, version
1 did not support the alarm, but it appears to be supported in v2, the
latest version.
I just ran a very simple test of it and it works just fine on my
system. Perhaps my test was too simple (the source is included below)
for what you need to do. If it was then ignore this message. If not
and you're using version 1, then you might consider upgrading to v2.
--Mark
mst...@tiac.net
#include <stdio.h>
#include <unistd.h>
#include <signal.h>
void myalarm()
{
static int count=0;
fprintf(stdout,"myalarm called : count = %d\n",++count);
alarm(1);
}
int main(int argc, char **argv)
{
int i;
fprintf(stdout,"Running a test alarm routine...\n");
signal(SIGALRM,myalarm);
alarm(1);
fprintf(stdout,"...going to sleep...\n");
sleep(30);
fprintf(stdout,"Done.\n");
}
Robert Hyatt
Ed Schroder (rebc...@xs4all.nl) wrote:
: : The problems weren't *all* mine. :) Only most of 'em were... the
: : high email traffic was generated to try and be sure that we were both
: : in agreement on solutions for every little snag that came up. so that
: : the results of the match would be valid.
: Stop excusing in your postings! It was a pleasure to cooperate with
: you! Crafty was never 100% tested for auto232. As you know Murphy's
: Law is cruel in these cases. Let's blame Murphy instead... :)
Damn murphy to hell then.. :) I can honestly say that I've had easier
operating system bugs to find than this auto232 thing. I've always hated
timing problems. This has done *nothing* at all to change that one bit. :)
: : Whether that took too much of his or my time really wasn't an issue
: : then, it might have become one later, although I had made the
: : committment to do what I could to make this happen. It wasn't easy
: : for either of us, that's for sure...
: Right, your tempo was amaaaaaaaazing... :)
I might not be too bright, but I can certainly change things quickly, with the
best of 'em... even if the changes don't do anything. My wife thought I was
nuts. Going to bed every nite at 3am-4am, getting up at 6am and looking like
hell all day long. :)
I frequently heard the magic "chessshit" obscenity around the house... :)
With some mumbled adjectives I was afraid to even ask about....
Robert Hyatt
Pete Nielsen (pnie...@ucla.edu) wrote:
: Robert Hyatt (hy...@crafty.cis.uab.edu) wrote:
: >
: > This is almost backward. The issue is this: suppose you search very
: > deeply and find a position you like and you believe you can reach. And
: > with best play you are going to stand better at the end. Then, just before
: > you move the piece, you decide to go over it one more time, and this time,
: > way out at the end, you find a really obscure move your opponent can make
: > that breaks the position wide open. Now you have two choices: (1) go ahead
: > and play the move you were going to play, not knowing if your opponent will
: > now start trying to find ways to thwart that move your opponent "might" play,
: > but since you are a computer, you assume he will see it, and the only way to
: > avoid getting busted completely is to give up two pawns.
: >
: > As a human, I'd think that losing two pawns is probably losing the game,
: I'm still confused by this: "The best move is not the most effective move"
: What about a thought experiment. Suppose given the time limits and game
: history the program could predict how many plys it's going to be able
: to search, would it make sense to weight each successive ply so that it
: would have to be some factor that increeased with ply deapth better than
: the current favorite in order to replace it. Would that address the problem?
That's one idea. Problem with it is that the solution is not only a "depth"
issue. Some deep lines are move obvious than some shallow lines due to the odd
moves in the shallow line. Michie called this the "discernability problem"
Robert Hyatt
Mark B. Stucky (mst...@tiac.net) wrote:
: On 27 Mar 1997 15:34:30 GMT, hy...@crafty.cis.uab.edu (Robert Hyatt)
: --Mark
: mst...@tiac.net
: signal(SIGALRM,myalarm);
: alarm(1);
: fprintf(stdout,"Done.\n");
: }
Actually I had heard that V2 supported this. However, even in unix, you
don't *ever* use timing code of any sort that wants to time something to
the millisecond accuracy range. I believe the magic number I was looking
for, according to feedback from someone that has fooled with auto232, was
50 ms... That's real close to the standard PC RTC frequency...
Komputer Korner
Robert Hyatt wrote:
>
snipped
>
> The rook problem is long since fixed. It was the same passed pawn
> values that were incorrect. I have an array that gives the values
> of passed pawns rank by rank. I neglected to remember that value[0]
> is the first rank where no pawn can ever be. As a result, the values
> got higher than they should, quicker than they should. A passer on
> the 6th is a serious threat. A passer on the 3rd is not. But if you
> forget about 0-origin array subscripting because you've been playing
> around with a FORTRAN program, it's easy to confuse 'em. :)
>
> : Bob, another question: Is Crafty able to play under LINUX through
> : the RS232 and contact the auto232 at the DOS-machine?
>
> Crafty. Rather, I use the "PRN" interface from DOS, which is
> intercepted by the auto232 TSR that is loaded before a game starts.
>
> As a result, Crafty won't talk to a RS232 port, because the protocol
> between chess program and TSR is significantly different from the
> protocol auto232 uses to talk to it's "twin" on the other machine
> over the RS232 link.
>
Bob, can you explain what it will take to get Crafty working with the
autoplayer and why don't you run the match?
--
Komputer Korner
The inkompetent komputer.
Komputer Korner
Ed Schroder wrote:
>
snipped
> I can't speak for others since I don't have the source code. I can speak
> for Rebel. I believe the following is true for Rebel:
> - A doubling in speed is good for a 50-75 ELO gain.
> - P90 to P200 is +50
> - P90 to PP200 is +75
>
> Let's take the average of +60.
> Then I would split the +60 into:
> - +45 for finding better POSITIONAL moves;
> - +15 for a better tactical understanding;
>
> That's why I believe knowledged based chess programs do have a better
> future than speed based chess programs. Of course I can't prove this
> only time will tell.
>
> - Ed Schroder -
Ed, assuming that you think that Hiarcs 6 and Rebel 8 have the 2 best
evaluation functions,even better than Deep Blue, then how fast does
Intel have to build an IC so that you micros will be able to defeat
Deep Blue in a match? Or do you think that Deep Blue would lose a
match right now?
Komputer Korner
> Komputer Korner
>
> The inkompetent komputer.
Okay you don't have to answer this. You answered it on other posts.
Ed Schroder
From: brucemo <bru...@nwlink.com>
> Seventy emails alone from Bob !!! Don't you see Ed's point?
> Ed has not the time to for that much problems, he must write Rebel 9
> to beat Hiarcs 6 :-)
: Of course. It involved a lot of time to get Crafty + auto232 working.
: But, in the end, it was working, wasn't it?
No it was not... :(((
Please reread Bob's 70 emails to me... :)))
- Ed Schroder -
: bruce
jmc...@aol.com
OK now for the important issue!
I bet my brother a case of beer (of his choice) that rebel would *not* lose by more than 2 points. Now my Brother is saying that if Rebel forfeits then the match result is:
Rebel:1 Crafty:9.
Will this in fact be the result of the match *if* Ed does what he has indicated? Is there any argument from chess custom that would allow me to say that I win or at least Draw?
Ed Schroder
From: Jack Nerad <JNE...@concentric.net>
> Speed does a lot with chess programs but it depends per chess program.
>> This sword has to cut with both edges. If you get better with a
>> 3x performance increase, so does everyone else. And if you get
>> better with a 3x performance increase, someone else is going to get
>> a lot better with a 100x performance increase. I don't see how you
>> can argue otherwise...
> I can't speak for others since I don't have the source code. I can speak
> for Rebel. I believe the following is true for Rebel:
> - A doubling in speed is good for a 50-75 ELO gain.
> - P90 to P200 is +50
> - P90 to PP200 is +75
> Let's take the average of +60.
> Then I would split the +60 into:
> - +45 for finding better POSITIONAL moves;
> - +15 for a better tactical understanding;
> That's why I believe knowledged based chess programs do have a better
> future than speed based chess programs. Of course I can't prove this
> only time will tell.
> - Ed Schroder -
: If this is the case, I really don't understand why you two argued
: about NPS in the first place:
: If you take a program and run it on a computer half as fast as another
: for twice as long as on the other, you could expect the same result.
: Given this, it would imply that total nodes searched is the key.
: Therefore:
: Time (t) = Rebel 8.0 Elo Strength (Elo)
: 2t = Elo + 60 (average, see above)
: 4t = Elo + 120
: 8t = Elo + 180
: 16t = Elo + 240
: 32t = Elo + 300
: 64t = Elo + 360
: 100t = Elo + ~393 (linear interpolation).
Wish this was true!
Buth the truth unfortunately isn't linear... :)
: Just imagine: in 8-9 generations of doubling (How many years is that,
: 14? 15?) Rebel 8.0 will be able to *Crush* world champions with ease.
: How many times faster than a micro is Deep Blue? 1250x? If IBM
: had the Rebel 8.0 Engine, it would smear Kasparov -- that's Rebel 8.0
: + 600 Elo points -- at least according to Mr. Schroder's estimation
: of increased strength based on speed doubling. 3000 Elo. Wow.
: Is this sarcasm? Must be my math.
Your math is ok Jack but your conclusion is wromg. It's an old subject
discussed here many times. In the past I posted a kind of scale which
for Rebel I estimate (guess) would look as follows:
P90 => P200 +50
P200 => PP200 +25
PP200 => PP400 +50
PP400 => PP800 +40
PP800 => PP1600 +30
PP1600 => PP3200 +20
And so on...
It's my opinion there comes a time that a doubling in speed (without
changing the program) will simply bring nothing.
Uff...
Feel free to flame me... :))
With my:
Let's take the average of +60.
Then I would split the +60 into:
- +45 for finding better POSITIONAL moves;
- +15 for a better tactical understanding;
statement I simply explained my thoughts and views again about the
relationship between ply depth and positional play which is in my
opinion true for Rebel8 and is certainly completely different for
*any* other chess program!
Example for chess program X...
- P90 => PP200 +30
- +10 for finding better POSITIONAL moves;
- +20 for a better tactical understanding;
Example for chess program Y...
- P90 => PP200 +80
- +40 for finding better POSITIONAL moves;
- +40 for a better tactical understanding;
And so on...
Also the *scale* for X and Y (P90 => PP3200) would look completely
different.
Quite a mess he?
- Ed Schroder -
: Regards,
: Jack Nerad
Komputer Korner
Robert Hyatt wrote:
>
snipped
>
> agreed... always have agreed with this...
>
> : For the time being, your prg seems to lack all that. But don't be sad, DB has
> : the same problem. It is the problem -- as I wrote several times -- of the not
> : yet prepared thousands of chess patterns that a real grandmaster of >2600
> : disposes of.
>
> sorry, but it doesn't. But it will eventually show us just how strong it
> is. I suspect you'll be surprised...
>
> : The other problem remains unsolved: how did Ed do his work in his limited deep
> : search? -- But also for this question you could read some answers in the
> : steinwender/Friedel book ...
> : For me this is far too difficult to understand. And I want the CIA to stay out
> : of this. Secrets are not to be sold! :))
>
> : Folks, remember, I love you all. :)
> : Especially on easter. It's like an automatical reflex somehow.
>
> : Rolf <Still weak at computer programming> Tueschen
Bob the mosquito has attacked you now. First, it was Ed, and now you.
Perhaps I should look over my shoulder and get some bug repellant
in preparation.
Dan Kirkland
In article <5hbh7j$s91$1...@news1.xs4all.nl> Ed Schroder,
>game-1 is finished.
Ed, you wimp! :)
Oh well, Ed forfeits, crafty WINS!!!
I would also like to add...
Even IF you won this match (you wouldn't!), you could reverse
the time odds against Deep Blue (rebel get 100 to DB's 1) and
you would still get CRUSHED!
Okay, I don't really KNOW that this is true...
But from looking at what evidence that there is...
What I can't figure out...
Ed Schroder has written a very fine chess program. He has also
been in the business for quite a while and should have a very
good idea what is going on. Yet he makes these silly statements
about Deep Blue...?
It makes me wonder if maybe this Ed Schroder making these
statements on this newsgroup and the author of Rebel are two
different people...? Will the REAL author of Rebel please step
forward? :)
Or maybe...
Just maybe..., Ed is a bit jealous of Deep Blue? Hmmmm...
Well Ed, maybe you can get a bunch of money together and then
challenge Deep Blue to a match? Then you could maybe show
that you are right (ya right!) and that I am wrong!
Of course you would need a BUNCH of money. Seeing that you
are known to wimp out of matches after the first game! This
way, when you wimp out after the first game you will forfeit
the match and the money!
dan
Torstein Hall
--
Torstein Hall
tors...@eunet.no
http://login.eunet.no/torshall/
Robert Hyatt <hy...@crafty.cis.uab.edu> skrev i artikkelen
<5heho7$h...@juniper.cis.uab.edu>...
> Pete Nielsen (pnie...@ucla.edu) wrote:
>
> : --
>
> : Re: "Crafty played a weaker move because ... "
>
> : Excuse my ignorance, but this has been puzzeling me ever since it came
up
> : in the DB commentary.
>
> : Doesn't the AB search guarentee that ANY move that the computer selects
is
> : at least as good as one it previously considered but now rejects?
>
> yes...
>
> but the problem is.. alpha/beta/minimax is a zero-sum game search
strategy,
> which assumes that both sides will play their best. If you are
out-searching
> your opponent badly (hint, if you were to play a brand new chess player
that
> barely knew the rules, would you try e4, Bc4, Qf3 on him? why?) then you
> assume your opponent can see everything you see, which might not be
right.
>
> and if you are parrying threats that are very deep, you might give up
something
> to parry a threat that your opponent had no hope to see... that's the
problem
> I am interested in, because I see it often enough to know it is a
problem...
>
If you keep on seeing problems way ahead of your opponent, I promise you
that in the end you will win the game!
And normaly, if you can figure it out at ply 11-12 that a line will lead
into trouble, then the other player ( program ) should normaly be able to
see it at ply 6 or 7 at least?!
So then you better avoid that line!?
I belive that a program trying to fix a "problem" like this one will result
in a program play weaker and more speculative chess. (or perhaps more human
chess.)
On the other hand a problem I can see is something like this:
A line leading to a eval -6 at ply 15 is thrown away for a line leading to
-5 at ply 3.
I belive that when you are at a big negative score, trying to keep your
position together as long as posible is what is called tough resistance! Is
this dificult to fix?
--
Torstein Hall
tors...@eunet.no
http://login.eunet.no/torshall/sjakk.html
Robert Hyatt
jmc...@aol.com wrote:
: OK now for the important issue!
: I bet my brother a case of beer (of his choice) that rebel would *not* lose by more than 2 points. Now my Brother is saying that if Rebel forfeits then the match result is:
: Rebel:1 Crafty:9.
: Will this in fact be the result of the match *if* Ed does what he has indicated? Is there any argument from chess custom that would allow me to say that I win or at least Draw?
:
Of course not. It simply ends at 1:0... I didn't see nor expect to see the
word "forfeit" anywhere. I think he used the word "stop"...
Robert Hyatt
Komputer Korner (kor...@netcom.ca) wrote:
: Robert Hyatt wrote:
: >
: snipped
: >
: > The rook problem is long since fixed. It was the same passed pawn
: > values that were incorrect. I have an array that gives the values
: > of passed pawns rank by rank. I neglected to remember that value[0]
: > is the first rank where no pawn can ever be. As a result, the values
: > got higher than they should, quicker than they should. A passer on
: > the 6th is a serious threat. A passer on the 3rd is not. But if you
: > forget about 0-origin array subscripting because you've been playing
: > around with a FORTRAN program, it's easy to confuse 'em. :)
: >
: > : Bob, another question: Is Crafty able to play under LINUX through
: > : the RS232 and contact the auto232 at the DOS-machine?
: >
: > No. The reason is, I did not implement a serial-port level drive in
: > Crafty. Rather, I use the "PRN" interface from DOS, which is
: > intercepted by the auto232 TSR that is loaded before a game starts.
: >
: > As a result, Crafty won't talk to a RS232 port, because the protocol
: > between chess program and TSR is significantly different from the
: > protocol auto232 uses to talk to it's "twin" on the other machine
: > over the RS232 link.
: >
: Bob, can you explain what it will take to get Crafty working with the
: autoplayer and why don't you run the match?
: --
If I could answer that, it'd already be fixed. :) The problem I see is
that there are a bunch of different auto232 drivers, one for each different
commercial program, plus a "no-name generic interface" for others. As I
get input, part of it is for the no-name interface, part of it is for the
program-specific interfaces. This makes life miserable... :)
Also, I don't have two machines, although in this match one machine is all
that is needed since both are never thinking at the same time. Big problem is
that my machines are all unix. And not many commercial programs will run under
that system. :)
We'll find a solution, however...
Robert Hyatt
Ed Schroder (rebc...@xs4all.nl) wrote:
: Wish this was true!
: Buth the truth unfortunately isn't linear... :)
Monty Newborn has asked me to run some tests for an upcoming ICCA Journal
article he's going to write. It is an extension of the article he wrote
back in 1985 or so, with the idea of searching a very large set of problem
positions to extreme depths, and then looking at the output for each one to
see if going from (say) 10 to 11 produces different (and better) moves or
not and how often. Then from 11 to 12, and from 12-13, and so forth. This
gives insight into what exactly another ply means. Once it hits the JICCA,
I'll see if we can get a summary of the results posted here. RIght now I don't
have an answer. This also might be useful to compare depth=10 to depth=15,
which is a huge difference computationally. Be nice to know what it means
playing skill wise too...
Robert Hyatt
<WHACK> <splat>
problem solved...
Robert Hyatt
Torstein Hall (tors...@eunet.no) wrote:
: --
yes, with minimax. Because the -5 looks better. to recap your example:
move X produces a score of -5 at ply=3,4,5,6,... N
move Y produces a score of +2 at ply=3,4,5,6,... N-1, but at ply=n it
drops to-6.
minimax will play X if the search goes to N. It would play Y if the search
only reached N-1. That can be a problem.
Ed Schroder
From: hy...@crafty.cis.uab.edu (Robert Hyatt)
: Stop excusing in your postings! It was a pleasure to cooperate with
: you! Crafty was never 100% tested for auto232. As you know Murphy's
: Law is cruel in these cases. Let's blame Murphy instead... :)
> Damn murphy to hell then.. :) I can honestly say that I've had
> easier operating system bugs to find than this auto232 thing. I've
> always hated timing problems. This has done *nothing* at all to
> change that one bit. :)
: : Whether that took too much of his or my time really wasn't an issue
: : then, it might have become one later, although I had made the
: : committment to do what I could to make this happen. It wasn't easy
: : for either of us, that's for sure...
: Right, your tempo was amaaaaaaaazing... :)
> I might not be too bright, but I can certainly change things quickly,
> with the best of 'em... even if the changes don't do anything. My
> wife thought I was nuts. Going to bed every nite at 3am-4am, getting
> up at 6am and looking like hell all day long. :)
> I frequently heard the magic "chessshit" obscenity around the
> house... :) With some mumbled adjectives I was afraid to even ask
> about....
Chessshit, how nice...
Well here it goes different, "Daddy when can I call?"
As you see my daugthers already have given up... :)))
- Ed Schroder -