Computer chess strategy

[Al]

Hello everybody:

I read frequently this newsgroup, but I have never posted to it because
I am not a chess programmer, but in this occasion I have an idea which
could perhaps have some interest. If it has been already said please
accept my apologies.

Let's suppose that Deep Blue (for instance) could be made 18 times
faster than now. The idea is: Instead of using this 18fold times speed
increase to search a couple of plies deeper, why not using it to search
18 times faster? I mean, instead of answering in 3 minutes (180 secs) to
the human being, answering to him in 10 seconds, making it more like a
blitz match. There must be some benefit since the total time taken by
one game is shortened from 4 hours to 2 hours, reducing the amount of
time that the human being is able to think. Perhaps to give him half the
time to think would be worthier that the two plies that are no being
searched.

I understand that this solution can be more practical (ie. leading to a
win) that theoretical (ie. it does not make the program play any
better).

Please forgive my english, it is not my mother tongue.

Alvaro Polo

[graham_douglass]

This strategy, of depriving GK of time to think by playing replies in 10 seconds,
could be DB's best hope.

Since I believe that it has no chance of beating GK under normal circumstances,
maybe it should use this strength. In this battle, it needs every advantage it
can get.

Play the player. Deprive GK of any rest and relaxation. Good thinking.

Graham

In article <5jiouh$t74$1...@lince.lander.es>, Al says...

[Robert Hyatt]

GrahamDouglass wrote:

: This strategy, of depriving GK of time to think by playing replies in 10 seconds,

: could be DB's best hope.

: Since I believe that it has no chance of beating GK under normal circumstances,
: maybe it should use this strength. In this battle, it needs every advantage it
: can get.

: Play the player. Deprive GK of any rest and relaxation. Good thinking.

: Graham

Just to refresh the discussion, I've actually tested this hypothesis and
found it valid. Several years ago, computers used to use the "60 moves
in 5 minutes of cpu time" rule for blitz. I started using the real clock
instead. And the program won more games (Cray Blitz.) Then, via accident,
we were playing IM Mike Valvo a long match of blitz games at an ACM event,
and our high-quality HP terminal got into a mode where the "=" key was not
sending an ascii = character. So I could not change the time control at
all, and had to use the default of 5 secs/move. However, we had 4 cpus
and that turned into 1.25 seconds per move (cumulative CPU time.) When we
played Mike like this, our score went *way* up. Because he was getting
less time to think on CB's time, and had to use more of his own time. This
got my curiosity up for "the experiment" which I ran over a couple of years.

I tweaked the time control allocation in CB as follows: (remember, blitz
chess, 5 mins per side, game played on real chess board using a real chess
clock with my operator time factored in against CB... no special considerations
at all, just a big incentive for me to type like hell and move quickly. :)
for the first minute of the game, CB would use 3 secs/move, and stuck with
this until 1 minute was off of its clock (ie it did not count time spent
thinking on the opponent's time.) The for the next minute, 2 secs per move,
same idea, and then the final fall-back to 1 second per move for the rest of
the game. I could easily reach 100 moves in such games, being a good typist
and experienced chess player, and we lost practically no games from that point
on...

So it might work. It might not... but it is an idea... :)

Bob

: In article <5jiouh$t74$1...@lince.lander.es>, Al says...

[Komputer Korner]

Al wrote:
>
> Hello everybody:
>
> I read frequently this newsgroup, but I have never posted to it because
> I am not a chess programmer, but in this occasion I have an idea which
> could perhaps have some interest. If it has been already said please
> accept my apologies.
>
> Let's suppose that Deep Blue (for instance) could be made 18 times
> faster than now. The idea is: Instead of using this 18fold times speed
> increase to search a couple of plies deeper, why not using it to search
> 18 times faster? I mean, instead of answering in 3 minutes (180 secs) to
> the human being, answering to him in 10 seconds, making it more like a
> blitz match. There must be some benefit since the total time taken by
> one game is shortened from 4 hours to 2 hours, reducing the amount of
> time that the human being is able to think. Perhaps to give him half the
> time to think would be worthier that the two plies that are no being
> searched.
>
> I understand that this solution can be more practical (ie. leading to a
> win) that theoretical (ie. it does not make the program play any
> better).
>
> Please forgive my english, it is not my mother tongue.
>
> Alvaro Polo

No one is questioning Deep Blue's prowess at Blitz chess. At that
time control and perhaps up to 30 minute games, it is the
undisputed chess champion of the world. However the interest is at
40 in 2 chess these days. Next week will answer a lot of questions.
My prediction is Kasparov will win 2 games and draw the rest.
--
Komputer Korner

The inkompetent komputer.

[Komputer Korner]

Graham, Douglass wrote:
>
> This strategy, of depriving GK of time to think by playing replies in 10 seconds,
> could be DB's best hope.
>
> Since I believe that it has no chance of beating GK under normal circumstances,
> maybe it should use this strength. In this battle, it needs every advantage it
> can get.
>
> Play the player. Deprive GK of any rest and relaxation. Good thinking.
>
> Graham
>

snipped

This strategy is flawed for the following 2 reasons.
1) Kasparov will then have longer thinking times than opponent.
2) Deep Blue's P.B. can not match Kasparov's on the move normal thinking
process. Thinking on the opponents time is not the computer programs's
strong area.
Deep Blue will allocate its time according to its
time algorithm which is probably the most efficient/best part of chess
programs.

[Tom C. Kerrigan]

Komputer Korner (kor...@netcom.ca) wrote:

> This strategy is flawed for the following 2 reasons.
> 1) Kasparov will then have longer thinking times than opponent.

Maybe in your universe. In my universe, X units of time is generally just
about as long as X units of time. Dear Lord, please think before
posting...

> 2) Deep Blue's P.B. can not match Kasparov's on the move normal thinking
> process. Thinking on the opponents time is not the computer programs's
> strong area.

This is silly. If DB doesn't think on the opponent's time the same way
every other program does it, I will buy a hat so I can eat it.

Cheers,
Tom

[Komputer Korner]

Tom, the original poster had the idea that Deep Blue should purposefully
move fast the whole game and not allow Kasparov to think on its time
at all. As for the 2nd point, you simply don't read. I said that
thinking
on the opponents time is a vastly inferior area of the game of all
computers when compared to humans. I did not make a distinction between
Deep Blue and any other computer in this area. It is now becoming
tiresome
to have to respond to your attempted refutations of anything I say.
Concentrate on your program and quit wasting time misreading my posts.

[Serge Desmarais]

Tom C. Kerrigan wrote:
>
> Komputer Korner (kor...@netcom.ca) wrote:
>
> > This strategy is flawed for the following 2 reasons.
> > 1) Kasparov will then have longer thinking times than opponent.
>
> Maybe in your universe. In my universe, X units of time is generally just
> about as long as X units of time. Dear Lord, please think before
> posting...
>
> > 2) Deep Blue's P.B. can not match Kasparov's on the move normal thinking
> > process. Thinking on the opponents time is not the computer programs's
> > strong area.
>
> This is silly. If DB doesn't think on the opponent's time the same way
> every other program does it, I will buy a hat so I can eat it.
>
> Cheers,
> Tom

I think that was was intended, here, was that computers only think
about what they identified as the best opponent's answer in playing their
last moves... So, if the opponent comes up with another move, it will be
as if no thinking was done on the opponent's time. But a human, thinking
the same way, would think "Hey! What if he decides to give a check on c4
first?" That wasn't the anticipated answer when he played, but still... U
often see the eval suddenly raise in a computer program, thinking on
the opponent's time... For example, from +1.05 to +2.56... So U think :
it saw a clear win! Nope! It just saw a win IF the opponent plays the
anticipated move! Then, the opponent plays something else ent it comes
back (the eval) to where it was!

Serge

[Tom C. Kerrigan]

Komputer Korner (kor...@netcom.ca) wrote:

> Tom, the original poster had the idea that Deep Blue should purposefully
> move fast the whole game and not allow Kasparov to think on its time

If DB is thinking on the opponent's time, and the opponent is Kasparov,
then DB is thinking just as long as Kasparov (assuming DB is moving
instantly). Makes sense, huh?

> at all. As for the 2nd point, you simply don't read. I said that
> thinking
> on the opponents time is a vastly inferior area of the game of all
> computers when compared to humans. I did not make a distinction between
> Deep Blue and any other computer in this area. It is now becoming

I simply don't read? Perhaps you simply don't remember. Let's see what you
actually said:

> 2) Deep Blue's P.B. can not match Kasparov's on the move normal thinking
> process.

You could have just said, "DB isn't as good as Kasparov."

> Thinking on the opponents time is not the computer programs's
> strong area.

And it isn't the weak area, either. I have no idea why you would think
pondering was any better or worse than thinking while the clock is
ticking, unless you don't quite understand how it works, in which case you
shouldn't be posting about it in the first place.

Cheers,
Tom

[mclane]

Serge Desmarais <psy...@total.net> wrote:

> I think that was was intended, here, was that computers only think
>about what they identified as the best opponent's answer in playing their
>last moves... So, if the opponent comes up with another move, it will be
>as if no thinking was done on the opponent's time. But a human, thinking
>the same way, would think "Hey! What if he decides to give a check on c4
>first?" That wasn't the anticipated answer when he played, but still... U
>often see the eval suddenly raise in a computer program, thinking on
>the opponent's time... For example, from +1.05 to +2.56... So U think :
>it saw a clear win! Nope! It just saw a win IF the opponent plays the
>anticipated move! Then, the opponent plays something else ent it comes
>back (the eval) to where it was!

>Serge

Serge you are right. This is something, we discussed over and over in
our circles. If you only cosider about the best move, and the opponent
is human, or a stupid chess program, you would have to reconsider each
time again. Over the years almost any chess program works the way that
they take the next move from the deepest - search - main-line, and
consider about this move in permanent-brain. Of course this is not
very human like, and very optimistic. AND it is of course wrong: we
know from life that there is no best move at all. So the chances that
the opponent plays this move has to do with the similarity between the
two system. Is the opponent another chess program, maybe a
chess-program using the same ideology, the chances are good.

If the circumstances are different, the chances for the PB -move
generate different.

I guess all programmers made their peace with this easy idea of :
maybe it works - or not algorithm.

They think: if opponent plays another move, than this move is per
definition a weaker move anyway. So I can later find good moves
against a weaker move. The problem here again is:
A best move in a position might be a weak move in a game.

Against Kosahvilil chess system tal sacced the bishop on h3.
All chess programs believe that 17.f4 is the best.
Therefore they do permanent brain about f4.
Maybe f4 is the best move in the position from a materialistic
computerish point of view. From a human point of view, from the point
of a strong human, a "weaker" move, giving the rook back against the
bishop (in german we say, giving the quality back) is THE BETTER MOVE
FOR THE GAME.

Do you see the difference and the problems here ? The better move for
the game has something to do with quality, planning and intelligence.
Finding the best move in a position has somehting to do with quantity
(NPS, search-depth, tactics) and no planning and stupidity.
All chess programs think about f4 !!!!!!
But if you input 17.Re5 instead, also ALLLLLLL chess programs see that
this move increases the white evaluation !!!!
So alpha-beta is not able to handle this conflict.
The conflict that the opponent is not forced to play the "best-move".

The conflict between "best-move in a position" and "Better move for
the whole game" has the same big grave like the difference between
dead and living materia has.
You have to breath spirit into materia, to make it living.
You have to breath KNOWLEDGE into chess-programs, to build a bridge to
go over this grave from "best-move" Alpha-beta thinking to
17.Re5 - thinking.

I have tried to explain this in this newsgroup in several posts now.
But nobody listens. If they would put the positions in their
chess-programs, they would see that their programs are far away from
beating humans anyway.

The whole idea of building chess-programs with alpha-beta and relying
on all these many stupid axioms is coming into the dead-end-street
now. If they don't shift their paradigm or ideology in the next years,
they can stop work. Because they will not make any progress. Better
start by sero and begin again, but with algorithms that try to
understand about chess instead of waiting until quantity generates
quality.

[Dan Thies]

Since DB is twice as fast, they could *sort of* do this.
If they use GK's time to search all possible GK moves, instead of the
usual trick of just searching along the PV, they would frequently have
a move ready to play when GK made a move.

They'd get pretty close to the same tactical depth that they got last
year, with the added advantage of a better evaluation (well, it might
be worse for all anyone knows) and less ponder time for the opponent.

I do not think that this is what they *should* do, but it is possible.

Dan

[Robert Hyatt]

mclane (mcl...@prima.ruhr.de) wrote:
: Serge Desmarais <psy...@total.net> wrote:

: > I think that was was intended, here, was that computers only think
: >about what they identified as the best opponent's answer in playing their
: >last moves... So, if the opponent comes up with another move, it will be
: >as if no thinking was done on the opponent's time. But a human, thinking
: >the same way, would think "Hey! What if he decides to give a check on c4
: >first?" That wasn't the anticipated answer when he played, but still... U
: >often see the eval suddenly raise in a computer program, thinking on
: >the opponent's time... For example, from +1.05 to +2.56... So U think :
: >it saw a clear win! Nope! It just saw a win IF the opponent plays the
: >anticipated move! Then, the opponent plays something else ent it comes
: >back (the eval) to where it was!

: >Serge

: Serge you are right. This is something, we discussed over and over in
: our circles. If you only cosider about the best move, and the opponent
: is human, or a stupid chess program, you would have to reconsider each
: time again. Over the years almost any chess program works the way that
: they take the next move from the deepest - search - main-line, and
: consider about this move in permanent-brain. Of course this is not
: very human like, and very optimistic. AND it is of course wrong: we
: know from life that there is no best move at all. So the chances that
: the opponent plays this move has to do with the similarity between the
: two system. Is the opponent another chess program, maybe a
: chess-program using the same ideology, the chances are good.

It's not "wrong"... if you'd like the simple mathematical explanation
of why it is the correct way to search, let me know and I'll post it.
It's been covered before, about once per year...

[Komputer Korner]

Tom C. Kerrigan wrote:
snipped

>
> You could have just said, "DB isn't as good as Kasparov."
>
> > Thinking on the opponents time is not the computer programs's
> > strong area.
>
> And it isn't the weak area, either. I have no idea why you would think
> pondering was any better or worse than thinking while the clock is
> ticking, unless you don't quite understand how it works, in which case you
> shouldn't be posting about it in the first place.
>
> Cheers,
> Tom

Is your program's pondering algorithm different than all the rest and do
you know if Deep Blue's pondering is different than all the rest of
the programs? As I understand pondering, if the opponent does not
play the 1 move that the computer is thinking about, then all the
thinking on the opponents time is wasted. Since this happens about
50% of the time, then 50 % of the pondering or Permanent Brain time is
wasted thinking. Thus it is much better to have the computer use its
sophisticated time allocation algorithm and not worry about the
opponent. As a programmer you should know this.

[Komputer Korner]

Robert Hyatt wrote:
snipped

> It's not "wrong"... if you'd like the simple mathematical explanation
> of why it is the correct way to search, let me know and I'll post it.
> It's been covered before, about once per year...

Bob, isn't it true only for full width programs and those that
evaluate at the leaf nodes. For programs that are completely selective
search and/ or evaluate at the root wouldn't the program's search
benefit from a variable pondering algorithm? Also it might be
enlightening to show the alpha beta proof of 1 move pondering
analysis. A lot of us haven't seen it including me even though I
believe you.

[Komputer Korner]

mclane wrote:
snipped

Bob's and Tom's point is that they assume that the computer guesses
on average correctly 50 % of the time. This then is counterbalanced
by the 50% on average time (Hyatt's and Althofer's numbers) that is
taken
on the principal variation when the computer is on its own time. The
flaw
in the ointment is that Hyatt's number of 50% guesses is a bit high.
Kaufman
had figures of 33 % about 7 years ago. If the computers have improved
enough to 50 % guesses, I would be doubtful, but I guess we need
some experiments. Hyatt says Crafty reaches 50% which is quite good,
but Bob is it really 50%? If it is less, say 45%, then since that is
less
that it is for the 50 % figure for time taken for principal
variation thinking time, then it is better to think on your own time.
Even if it is 50%, a strategy of instant moving just so you can think on
the opponents time is rolling the dice. If Deep Blue can get up to 12
ply almost instantaneously, the 4 ply less that it loses (than if it
thought for the full 3 minutes) would cost it. Let us take this to the
logical extreme. If Deep Blue was able to guess Kasparov's moves 100% of
the time, would it be a correct strategy to move instantly all the time?
This is in practice what happens when most computer programs guess
correctly the opponents move ( not all the time, but very frequently
they move instantly).Assuming a guess rate of 50% by Deep Blue, the best
effective time advantage that Gary would realize would never be more
than
a 2 to 1 ratio. This sounds beneficial until you realize that to input a
strategy of always moving instantly for Deep Blue would cost it on
average
4 ply every 2nd move, or 2 ply per move (assuming a 50% guess rate).
Thus
you would be negating Deep Blue's speed increases the last 3 years at a
possible benefit of preventing Kasparov of thinking on Deep Blue's time.
This policy would be beneficial only if Kasparov got into time trouble
and I don't think it could force Kasparov into time trouble by itself
once
Gary realized what was happening. All this assumes that Gary in playing
to the maximum 4 hours per side wouldn't run out of moves in sudden
death.

[chrisw]

--
http://www.demon.co.uk/oxford-soft

Komputer Korner <kor...@netcom.ca> wrote in article
<336AD7...@netcom.ca>...

>
> Tom C. Kerrigan wrote:
> >
> > Komputer Korner (kor...@netcom.ca) wrote:
> >
> > > This strategy is flawed for the following 2 reasons.
> > > 1) Kasparov will then have longer thinking times than opponent.
> >
> > Maybe in your universe. In my universe, X units of time is generally
just
> > about as long as X units of time. Dear Lord, please think before
> > posting...
> >

> > > 2) Deep Blue's P.B. can not match Kasparov's on the move normal
thinking

> > > process. Thinking on the opponents time is not the computer
programs's
> > > strong area.
> >

> > This is silly. If DB doesn't think on the opponent's time the same way
> > every other program does it, I will buy a hat so I can eat it.
> >
> > Cheers,
> > Tom
>

> Tom, the original poster had the idea that Deep Blue should purposefully
> move fast the whole game and not allow Kasparov to think on its time

> at all. As for the 2nd point, you simply don't read. I said that
> thinking
> on the opponents time is a vastly inferior area of the game of all
> computers when compared to humans. I did not make a distinction between
> Deep Blue and any other computer in this area. It is now becoming

> tiresome
> to have to respond to your attempted refutations of anything I say.
> Concentrate on your program and quit wasting time misreading my posts.

Sorry KK, we forgot to tell you. I resigned from the 'Keeping KK under
Kontrol' job, and the committee advertised, and was very impressed with Tom
Kerrigan's application and interview, so he was appointed. His salary is by
results. Only you can save us the money now :)

Chris Whittington

[Robert Hyatt]

Komputer Korner (kor...@netcom.ca) wrote:

: Tom C. Kerrigan wrote:
: snipped
: >
: > You could have just said, "DB isn't as good as Kasparov."

: >
: > > Thinking on the opponents time is not the computer programs's
: > > strong area.
: >
: > And it isn't the weak area, either. I have no idea why you would think

: > pondering was any better or worse than thinking while the clock is
: > ticking, unless you don't quite understand how it works, in which case you
: > shouldn't be posting about it in the first place.
: >
: > Cheers,
: > Tom

: Is your program's pondering algorithm different than all the rest and do
: you know if Deep Blue's pondering is different than all the rest of
: the programs? As I understand pondering, if the opponent does not
: play the 1 move that the computer is thinking about, then all the

: thinking on the opponents time is wasted. Since this happens about

: 50% of the time, then 50 % of the pondering or Permanent Brain time is
: wasted thinking. Thus it is much better to have the computer use its
: sophisticated time allocation algorithm and not worry about the
: opponent. As a programmer you should know this.

: --
: Komputer Korner

: The inkompetent komputer.

First, the time spent is not wasted, thanks to the transposition table. Even
if the opponent makes a *different* move than the one predicted, it is likely
to be in the PV's that get computed when the program predicts the wrong move,
because in many cases, it is an issue of the orderin which moves are played,
not an issue of which moves are played. And all that searching seeds the
transposition table with good scores, and move ordering information, and seeds
the history move ordering heuristic with a good idea of which moves are good,
and so forth.

Second, time allocation is pretty much dependent on pondering, at least in
every program I've worked on, because you generally want to use some of the
time you gain by pondering, at a point earlier in the game when the position
is more complicated and often more critical. Turning it off certainly makes
Crafty play worse, although I have modified the time allocation code in newer
versions so that ponder=off turns off the "anticipation" that time will be
saved later on...

[Robert Hyatt]

Komputer Korner (kor...@netcom.ca) wrote:

: --
: Komputer Korner

: The inkompetent komputer.

Here's the simple idea about pondering: Let's assume a 50% hit rate,
which is actually quite low for todays programs. I've seen Cray Blitz
run at 75-80% most of the time, in correctly predicting the opponent's
move. But I'll stick with 50%.

1. Crafty correctly predicts the opponent's move 50% of the time, and
only considers that move in pondering. So roughly 50% of the time, Crafty
is right, and after the opponent moves, Crafty replies instantly, which
saves exactly 50% of the search time. Which makes a 40/2 game turn into
a 40/3 game...

2. Crafty spends it's time searching the best "n" opponent moves, somehow
obtained by doing an inefficient but shallow search to get them. Here, I'll
use "n" as the number of moves search. Then, after the opponent moves, let's
assume that the move he plays is *always* one of the chosen "n" moves (an
assumption that is *very* optimistic unless n=# of legal moves for the
opponent of course.) So after the opponent moves (and again I'll assume a
straight 3 minutes per move in a 40/2 game) Crafty has spent only 180/n
seconds searching the *right move. It has to continue searching for 180-180/n
seconds to spend the full 180 seconds on the *right* move. How does this
compare? If n=2, it is exactly the same as (1) above of course, because it
would then save 90 seconds per move, for 40 moves, which is 1 hour. And it
again gets 40/3 for the time control.

If n > 2, it fails to improve. For n=3, it saves only 60 seconds per move,
which is worse than averaging 90 seconds per move saved. for n>3 it only
gets worse...

*BUT* we aren't yet done.

If you analyze alpha/beta, searching the position below *one* move is about
twice as efficient as searching the positions below *two* moves, because if
you ponder one move, you only have to produce *one* PV, which is very
efficient. If you ponder two different moves, you have to produce two
different PV's, which is about 1/2 as efficient as only producing one.

Therefore, searching (pondering) two moves by the opponent is a total loss
in terms of time saved, when you compare everything...

If that's not clear enough, feel free to ask for more details...

[Robert Hyatt]

Komputer Korner (kor...@netcom.ca) wrote:
: mclane wrote:
: snipped

Against computers it is *much* higher at tournament time controls. Cray
Blitz was generally predicting at about 70% at the last ACM event we played
at. Crafty is nearly never below 50% on the servers playing against good
players...

: less

: that it is for the 50 % figure for time taken for principal
: variation thinking time, then it is better to think on your own time.

: Even if it is 50%, a strategy of instant moving just so you can think on
: the opponents time is rolling the dice. If Deep Blue can get up to 12

: ply almost instantaneously, the 4 ply less that it loses (than if it
: thought for the full 3 minutes) would cost it. Let us take this to the
: logical extreme. If Deep Blue was able to guess Kasparov's moves 100% of
: the time, would it be a correct strategy to move instantly all the time?
: This is in practice what happens when most computer programs guess
: correctly the opponents move ( not all the time, but very frequently
: they move instantly).Assuming a guess rate of 50% by Deep Blue, the best
: effective time advantage that Gary would realize would never be more
: than
: a 2 to 1 ratio. This sounds beneficial until you realize that to input a
: strategy of always moving instantly for Deep Blue would cost it on
: average
: 4 ply every 2nd move, or 2 ply per move (assuming a 50% guess rate).
: Thus
: you would be negating Deep Blue's speed increases the last 3 years at a
: possible benefit of preventing Kasparov of thinking on Deep Blue's time.
: This policy would be beneficial only if Kasparov got into time trouble
: and I don't think it could force Kasparov into time trouble by itself
: once
: Gary realized what was happening. All this assumes that Gary in playing
: to the maximum 4 hours per side wouldn't run out of moves in sudden
: death.

:

See previous post for why searching two moves is *much* worse than searching
one. There's a subtle alpha/point you are overlooking...

:
:

[Don Fong]

In article <336D67...@netcom.ca>,

Komputer Korner <kor...@netcom.ca> wrote:
>Bob's and Tom's point is that they assume that the computer guesses
>on average correctly 50 % of the time. This then is counterbalanced
>by the 50% on average time (Hyatt's and Althofer's numbers) that is
>taken on the principal variation when the computer is on its own
>time. The flaw in the ointment is that Hyatt's number of 50% guesses
>is a bit high. Kaufman had figures of 33 % about 7 years ago.

PMFJI, but what about the hashtable? even if it doesn't "guess"
the right move, isn't it true that "pondering" could speed up future
searches somewhat by putting some important positions in the hashtable?

[Komputer Korner]

Robert Hyatt wrote:
snipped The following are the 2 alternatives of which No. 1 is the
real world.
. 1. Crafty correctly predicts the opponent's move 50% of the time, and
. only considers that move in pondering. So roughly 50% of the time,
Crafty
. is right, and after the opponent moves, Crafty replies instantly,
which
. saves exactly 50% of the search time. Which makes a 40/2 game turn
into
. a 40/3 game...
.
. 2. Crafty spends it's time searching the best "n" opponent moves,
somehow
. obtained by doing an inefficient but shallow search to get them.
Here, I'll
. use "n" as the number of moves search. Then, after the opponent
moves, let's
. assume that the move he plays is *always* one of the chosen "n" moves
(an
. assumption that is *very* optimistic unless n=# of legal moves for the
. opponent of course.) So after the opponent moves (and again I'll
assume a
. straight 3 minutes per move in a 40/2 game) Crafty has spent only
180/n
. seconds searching the *right move. It has to continue searching for
180-180/n
. seconds to spend the full 180 seconds on the *right* move. How does
this
. compare? If n=2, it is exactly the same as (1) above of course,
because it
. would then save 90 seconds per move, for 40 moves, which is 1 hour.

And it
> again gets 40/3 for the time control.
>
> If n > 2, it fails to improve. For n=3, it saves only 60 seconds per move,
> which is worse than averaging 90 seconds per move saved. for n>3 it only
> gets worse...
>
> *BUT* we aren't yet done.
>
> If you analyze alpha/beta, searching the position below *one* move is about
> twice as efficient as searching the positions below *two* moves, because if
> you ponder one move, you only have to produce *one* PV, which is very
> efficient. If you ponder two different moves, you have to produce two
> different PV's, which is about 1/2 as efficient as only producing one.
>
> Therefore, searching (pondering) two moves by the opponent is a total loss
> in terms of time saved, when you compare everything...
>
> If that's not clear enough, feel free to ask for more details...

Ingo Althofer has produced stats showing that the k variation mode has
been improved so that it produces efficiencies so that the total time
taken
is less than twice the 1 variation mode even when K is large. However
your
first point is VERY true. There is no advantage in pondering more than 1
move assuming that the program can think normally if it guesses wrongly
on
the previous pondering think. However, The original poster postulated a
Deep Blue strategy of ALWAYS moving instantly. This would result in some
situations of the program having to make a move after a 12 ply instant
search instead of its 16 ply average after thinking for 3 minutes if it
guessed wrongly. 12 ply is not enough to defeat Kasparov in certain
critical
middlegame decisions.

[Komputer Korner]

Yes, the hash table will definitely help both cases, but that is the
point. Whether you ponder or not hash table hits will help the search.
But we are comparing the pondering to thinking on its own time.
Because of the way alpha beta works, Bob is correct and has shown the
proof that one move pondering is optimal. However, the original
poster wanted Deep Blue to instantly move until and even after Kasparov
got into time trouble. This would result in a situation of when the
computer has guessed wrong, the next computer move would be based on
a 12 ply search instead of 16 ply, one out every 2 moves if the guess
rate
is 50% against a human like Kasparov . 12 ply searches are not enough
to defeat Kasparov.

[Tom C. Kerrigan]

Serge Desmarais (psy...@total.net) wrote:

> I think that was was intended, here, was that computers only think
> about what they identified as the best opponent's answer in playing their
> last moves... So, if the opponent comes up with another move, it will be
> as if no thinking was done on the opponent's time. But a human, thinking

Ah hah! Funny how you have to explain what a professional writer is trying
to say, huh?

If memory serves, this thread was created to propose the idea of pondering
several moves, so the opponent never "comes up with another move." I guess
KK totally missed this part.

Cheers,
Tom

[Komputer Korner]

Tom, the argument here has 2 threads.
a) Is pondering or Permanent Brain less efficient overall than normal
thinking time? The answer is NO, as Bob Hyatt has proved in another
post.
b) However the 2nd thread is the original poster's proposal of Deep
Blue ALWAYS moving instantly throughout the whole game. This is
flawed unless Deep Blue'e guess rate is 100%. Flawed because by
moving instantly Deep Blue sacrifices 4 ply from its 16 ply full
width average during the middlegame ( just a guess here but hopefully
the Deep Blue team will confirm or deny this with the real average.)
12 ply is not enough to defeat Kasparov at certain critical positions.
If Deep Blue was able to get up to 16 or 18 ply almost instantly, then
the instant strategy might work, but only if Kasparov lets himself get
into time trouble.

[mclane]

Komputer Korner <kor...@netcom.ca> wrote:

> A lot of us haven't seen it including me even though I
>believe you.

>--
>Komputer Korner

>The inkompetent komputer.

I don't know about what you discuss here. I lost the point. But I can
tell you that there were dedicated-chess-computers from saitek, mainly
low-costs-machines, that
did the permanent-brain different.
If you are interested I name you the machine, the way it workde
precisely and the results of this technique.

When I remember it, the computer computed ALL legal moves in the move
list. Of course the time he spent on each move was ordered.
arround 60 % was spent on the first branches,
20 % on the second branch. 10 % on the 3rd branch and and and.

The machine was a Saitek machine, Saitek Turbo 16 K with a Julio
Kaplan program. I guess some years ago.

It was quite interesting. Blitz-games vs. this machine were not easy,
because whatever you played, it came almost instantly with a nice
answer. Human was not able to COMPUTE itself on the computer's time
because the computer moved always very fast.

I have not forgotten this. I mean: this machine was a SLOW 8-Bit
single-chip CPU, only 16 KByte Rom and maybe less Ram.
And although this is all not much ---- it worked.

I guess PC-programs could use (because of HASHING, as Bob mentioned)
this method with some clever routines much better.

If Bob is against this method, I will tell Chris to implement it in
the next CSTal version, just to show you THAT IT MAKES SENSE to use a
more senseful HUMAN way than the 50:50 chances we use today.

Thanks.

[mclane]

Komputer Korner <kor...@netcom.ca> wrote:

>mclane wrote:
>snipped

>Gary realized what was happening. All this assumes that Gary in playing
>to the maximum 4 hours per side wouldn't run out of moves in sudden
>death.
>

Instead of writing comments you should have tried out the Bxh3
position, than you would see that all the claims from Hyatt and
fellows are just over-the-years-established-prejudices of some
old-fashioned guys, thinking that the way THEY do it is so heavily the
ONLY way, that they behave like nuts whenever anybody is telling them:
i don't believe what you say. Instead of giving answers they shoot
with empty beer-cans.

[mclane]

df...@cse.ucsc.edu (Don Fong) wrote:

> PMFJI, but what about the hashtable? even if it doesn't "guess"
>the right move, isn't it true that "pondering" could speed up future
>searches somewhat by putting some important positions in the hashtable?

Don , do you really believe this is a dialogue ? I mean, you ask, and
KK asks much, and Bob answers. But the answers were written down
before somebody has invented them.

I guess these guys have a macro in their word-prg for each possible
request.

In my opinion the whole discussion is just idle fighting about nothing
than: who is right and who is more wrong.

They would discuss with you that the earth is flat. And they would
show you evidence for it. And they would show you statistics,
photographs and all that stuff, showing that the earth is a disc.

And than they would tell you: you don't need to believe it, but we
know it better because we have the source-codes of your program.

More and more this whole usenet remembers me on the home-magazin of
the vatican and - as you know - pope johannes 2nd is not making any
mistake whatever he talks about, these speeches to earth he holds are
not written to start a discussion, they are held to stop discussion
and show anybody: here is the one and only way, and this way leads to
heaven.

[mclane]

hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

Sorry Bob, but IMO this totally wrong, the only argument you have is
that almost all pc-programs are doing it the way crafty is doing it
(50:50 chance).

Humans do it - of course - different.
If I have to extrapolate behaviour of an unknown opponent, I will OF
COURSE NOT ONLY THIN ABOUT

O N E

possibility. I would be a fool only to consider about ONE possible
way. It is also not very intelligent, therefore no humans are doing
it.

A human (i guess) is thinking about a few answers his opponent could
make, and he has an aswer for each branch, or the human is not clever.
If he would not have something in petto, he would oversee many threads
and moves against him.

>If you analyze alpha/beta, searching the position below *one* move is about
>twice as efficient as searching the positions below *two* moves, because if
>you ponder one move, you only have to produce *one* PV, which is very
>efficient. If you ponder two different moves, you have to produce two
>different PV's, which is about 1/2 as efficient as only producing one.

>Therefore, searching (pondering) two moves by the opponent is a total loss
>in terms of time saved, when you compare everything...

I do neither understand your strange sentences above about searching 2
moves should be less efficient, nor do I think that you EXPLAIN
anything. The only thing you do is to spread bubbles of not filled
claims arround , instead of telling anything.

Your remarks about this topic remember me on the Flintstones. Of
course for them it would be much easier only to save 1 pv, because it
is so difficult to get the pv into stone.

We are not living in this time.

>If that's not clear enough, feel free to ask for more details...

How nice of you. As if you would know anyhing better. Oh - sorry, I
forgot you have Ed's and Chris' source-codes, what a pity you don't
have my sources...

[Robert Hyatt]

mclane (mcl...@prima.ruhr.de) wrote:
: Komputer Korner <kor...@netcom.ca> wrote:

: > A lot of us haven't seen it including me even though I
: >believe you.

: >--
: >Komputer Korner

: >The inkompetent komputer.

: I don't know about what you discuss here. I lost the point. But I can

: tell you that there were dedicated-chess-computers from saitek, mainly
: low-costs-machines, that
: did the permanent-brain different.
: If you are interested I name you the machine, the way it workde
: precisely and the results of this technique.

: When I remember it, the computer computed ALL legal moves in the move
: list. Of course the time he spent on each move was ordered.
: arround 60 % was spent on the first branches,
: 20 % on the second branch. 10 % on the 3rd branch and and and.

: The machine was a Saitek machine, Saitek Turbo 16 K with a Julio
: Kaplan program. I guess some years ago.

: It was quite interesting. Blitz-games vs. this machine were not easy,
: because whatever you played, it came almost instantly with a nice
: answer. Human was not able to COMPUTE itself on the computer's time
: because the computer moved always very fast.

: I have not forgotten this. I mean: this machine was a SLOW 8-Bit
: single-chip CPU, only 16 KByte Rom and maybe less Ram.
: And although this is all not much ---- it worked.

: I guess PC-programs could use (because of HASHING, as Bob mentioned)
: this method with some clever routines much better.

: If Bob is against this method, I will tell Chris to implement it in
: the next CSTal version, just to show you THAT IT MAKES SENSE to use a
: more senseful HUMAN way than the 50:50 chances we use today.

: Thanks.

Suit yourself of course. But Chris can do the math and determine that any
other method is inferior, unless the prediction rate is down in the 10-20%
range, something I've not seen from any program...

[Robert Hyatt]

mclane (mcl...@prima.ruhr.de) wrote:
: df...@cse.ucsc.edu (Don Fong) wrote:

: > PMFJI, but what about the hashtable? even if it doesn't "guess"
: >the right move, isn't it true that "pondering" could speed up future
: >searches somewhat by putting some important positions in the hashtable?

I'll answer Don't legit question, and ignore the bullshit rantings that
follow.

If you ponder and predict incorrectly, you are correct that the hash
table contains things that will make the re-search more efficient. *but*
you can't really tell that this happens... and it would be questionable
to cut the search way short, just because you think you got a lot of hash
hits. If you were only interested in searching to a fixed depth, then you
would hit this depth faster after an incorrect ponder, than if you had just
"sat" while waiting on the opponent, even if he played the wrong move, because
of the hash hits you'd find. But we really don't target a specific depth,
we target a specific time limit, so that after a wrong ponder, we use the
normal time, but we probably get extra depth for our wasted ponder, which
is *still* a win of course...

: Don , do you really believe this is a dialogue ? I mean, you ask, and

: KK asks much, and Bob answers. But the answers were written down
: before somebody has invented them.

: I guess these guys have a macro in their word-prg for each possible
: request.

: In my opinion the whole discussion is just idle fighting about nothing
: than: who is right and who is more wrong.

: They would discuss with you that the earth is flat. And they would
: show you evidence for it. And they would show you statistics,
: photographs and all that stuff, showing that the earth is a disc.

Please first write a program, and get up-to-speed on the algorithms that
are involved, before posting such drivel. No one's saying the earth is
flat. But I can certainly prove that the sqrt(2) = 1.4, or that
PI=3.14159265358979323846264...., and in the case of alpha/beta, I've
studied the thing long enough to know exactly how it behaves... something
you obviously don't understand.

Every time you disagree, you go from the point at hand, to something like the
earth is flat, to the next level of hyperbole... And the content (meat) of
the comments varies inversely in proportion to the hyperbole... Again,
shoot my math down, or pipe down... take your choice... but don't start
again with the name-calling bullshit...

: And than they would tell you: you don't need to believe it, but we

[mclane]

hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

>: In my opinion the whole discussion is just idle fighting about nothing
>: than: who is right and who is more wrong.

>: They would discuss with you that the earth is flat. And they would
>: show you evidence for it. And they would show you statistics,
>: photographs and all that stuff, showing that the earth is a disc.

>Please first write a program, and get up-to-speed on the algorithms that
>are involved, before posting such drivel. No one's saying the earth is
>flat. But I can certainly prove that the sqrt(2) = 1.4, or that
>PI=3.14159265358979323846264...., and in the case of alpha/beta, I've
>studied the thing long enough to know exactly how it behaves... something
>you obviously don't understand.

Unbelievable.
You know exactly what ?
That you permanent-brain on a move that will not come ?
And that you try to convince me that it is very effective to compute
on a move that will not come. Are you nuts ?

>Every time you disagree, you go from the point at hand, to something like the
>earth is flat, to the next level of hyperbole... And the content (meat) of
>the comments varies inversely in proportion to the hyperbole... Again,
>shoot my math down, or pipe down... take your choice... but don't start
>again with the name-calling bullshit...

I don't have to write a chess program to find out that somebody else
is an idiot.
Why don't you rely on examples, e.g. a game of chess ?
Why do you need to argue that you know better than I instead of
commenting on the chess-notation ?
Why do you claim that you know better from theory that what Ed does is
not possible instead of let him explain because he has written the
code and knows better than you ???

I can tell you: because you are unable to refer on facts.
As soon as you lose the discussion, you refer on strange weird ideas
like crafty lost against rebel8 because crafty has seen too much.
Briliant. Than make crafty more stupid next time and it will again
lose against rebel8.

I got an email from KK . He says that: you are making a fool of
yourself with this one. Read up
on alpha- beta first.

Dear KK: If I would count all the examples where you have made a fool
out of yourself, I would have to count a long long time.

Thats typical for you. Please do believe in his words, but let me
alone with my ideas. I have not said what I think about your postings
because I thought: ough - he has too learn much. 89 % of what he is
posting is wrong. So I watched how the others corrected you.

[mclane]

hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

>Suit yourself of course. But Chris can do the math and determine that any
>other method is inferior, unless the prediction rate is down in the 10-20%
>range, something I've not seen from any program...

A B-strategy program does not have to rely on optimisez backward or
forward-pruning algorithms. It has the chance to follow (sometimes)
more and sometimes less branches than any of the normal programs. If
this helps to find moves, the others have pruned away, than this is a
better way.

Again : How can somebody be so stupid only to expect that the opponent
would do the best move.

Only stupid computers behave this way.

Because in a chess-position there is not always THE BEST MOVE your
whole idea makes no sense. In most of the chess positions I know,
there is no best move. But your programs rely on ONLY one move,
because godfather Hyatt knows better than any chess-expert in the
world.

If there is no best move, how you your program using its second move
of the main-line prediction find the right move ????

[mclane]

hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

>We aren't living in the same world, because you don't have a clue about
>alpha/beta. The above is perfectly clear to *anyone* that knows beans
>about alpha/beta minimax. If you don't understand AB, then you sure aren't
>going to follow the above discussion... Therefore I elect to save my
>breath...

I really don't know what alpha/beta might have to do with the question
how my strategy should be in PB.

You have not answered to my question WHY all your programs play e.g.
Bxh3 and expect later a draw by repetition and plan f4 for Kosashvili
because THEY think f4 is the best move, and than Yona plays better Re5
instead, and suddenly the evaluation of all programs drop, although
Re5 should be a weaker move.

You always behave - here you show the same behaviour again -
as if OTHERS have not understand some main law you know for 40 years,
and therefore the others (in the PB-discussion me) cannot follow you.
How do you know that I don't know about AB algorithm ?
How do you know better about Ed Schroeders Source Codes?

If a program guesses a move XYZ is the best one. And therefore creates
a tree and a main-line, and it does a PB relying on the 2nd move of
the main-line, but the move is not coming because the opponent plays
his own move ABC, then I would expect that the move ABC is losing or
is a weaker move, because the tree was build in AB-backtracking.

But the move ABC isn't weaker.

No I have 2 possibilities:

a ) the evaluation function is shit
b) the search-algorithms prune important stuff (MOVE ABC) away and
oversee something.

In my example the programs expected f4. And evaluated with draw.
f4 was replaced by human with Re5 and what would you expect ?
That the computers' evaluation increases >0.00
But the evaluation drops.

Because ALL chess-programs react this way (try crafty yourself)
you can throw them all in the trash.

No - I am not interested which algorithm is in charge. If they
consider about moves that are stupid, and when the better moves come,
they drop in their evaluations, then something global is wrong...

Instead of claiming your: I have known about alpha-beta when you were
not born prejudice you should concentrate on what I said and not what
you wanted to read out of my words.

If your computers would have used the PB time, and instead of assuming
the shit f4 would come, computed on whatever strange moves, would have
computed anyhow different , maybe not only looking for the best moves,
they would MAYBE have seen Re5. If they would have seen it, they could
have better used PB-time instead of been shocked about the move, and
then dropped dropped and lost the game.

When your alpha-beta is so stupid that it cannot handle easy cases of
chess, than throw it away and create an algorithm that handles these
cases.

The best way to spare your breath would be to die. This would spare
much breath to me either. Please... do it or stop this :
I know something better than you do - behaviour. I am discussing about
points I can give you examples in chess-notation, games, names and
whatever. You always fight these : "I have known about TOPIC X when
you - my young fellow - have made in you clothes" - discussions.
Do them with KK, he will follow you and believe you magic words. I do
not believe you. I would never believe somebody who claims he is god
or a computer or has invented chess-programming-universe.

I can only hope that there comes the day somebody writes an
intelligent parallel program, and this - with the help of
zillion-cpu's kicks all of you stupid-fast-.fanatics in the ass and
send you to the back-side of the moon. There you can play with you
deep-blue and whatever stupid/hardware-monsters until the last day of
the universe.

Maybe the time comes earlier than you would expect it. From my
observations it does not take much time until CILK-Chess will show
your MONSTER-machines that chess has not only to do with hardware ,
and that arrogance will not help you to hide (=HYATT, maybe the best
synonym for hiding Artificial Stupidy behind the foggy words) that
these software running on fast hardware could only survive because
there was no program near, that was more intelligent.

I don't think that they would ever beat the champion. Maybe Kasparov.
(For me Kasparov is no world-champion because he wanted to make his
own deal in PCA. I don't want him, and I am happy that he left our
organisation.) You can buy him. But you can't buy others.

Maybe Cilk-chess could beat the champ. But Deep Blue ? Ha - they shall
play g5 - because it is forced and you prefer it.

[Robert Hyatt]

mclane (mcl...@prima.ruhr.de) wrote:
: hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

: Sorry Bob, but IMO this totally wrong, the only argument you have is
: that almost all pc-programs are doing it the way crafty is doing it
: (50:50 chance).

: Humans do it - of course - different.
: If I have to extrapolate behaviour of an unknown opponent, I will OF
: COURSE NOT ONLY THIN ABOUT

: O N E

: possibility. I would be a fool only to consider about ONE possible
: way. It is also not very intelligent, therefore no humans are doing
: it.

: A human (i guess) is thinking about a few answers his opponent could
: make, and he has an aswer for each branch, or the human is not clever.
: If he would not have something in petto, he would oversee many threads
: and moves against him.

But you aren't a computer using alpha/beta either. You can compare two
moves quite easily, and not figure out just which is better, but by how
much it is better. And that has nothing to do with this discussion. I
gave the simple math... take it apart, or refute it, but please don't
change the subject with wild hand-waving and gesticulation... You don't
like 50%? Pick a number... I'll revise my math using that number. But
for my case, 50% is low... I have data to support this...

: >If you analyze alpha/beta, searching the position below *one* move is about

: >twice as efficient as searching the positions below *two* moves, because if
: >you ponder one move, you only have to produce *one* PV, which is very
: >efficient. If you ponder two different moves, you have to produce two
: >different PV's, which is about 1/2 as efficient as only producing one.

: >Therefore, searching (pondering) two moves by the opponent is a total loss
: >in terms of time saved, when you compare everything...

: I do neither understand your strange sentences above about searching 2
: moves should be less efficient, nor do I think that you EXPLAIN
: anything. The only thing you do is to spread bubbles of not filled
: claims arround , instead of telling anything.

: Your remarks about this topic remember me on the Flintstones. Of
: course for them it would be much easier only to save 1 pv, because it
: is so difficult to get the pv into stone.

We aren't living in the same world, because you don't have a clue about

alpha/beta. The above is perfectly clear to *anyone* that knows beans
about alpha/beta minimax. If you don't understand AB, then you sure aren't
going to follow the above discussion... Therefore I elect to save my
breath...

: We are not living in this time.

: >If that's not clear enough, feel free to ask for more details...

: How nice of you. As if you would know anyhing better. Oh - sorry, I
: forgot you have Ed's and Chris' source-codes, what a pity you don't
: have my sources...

I have lots of blank pages... In the case of real computer chess internals,
that seems to be your source code...

[chrisw]

Robert Hyatt <hy...@crafty.cis.uab.edu> wrote in article
<5km4mg$3...@juniper.cis.uab.edu>...

Stand by to be flamed Thorsten.

You don't TELL me to do anything, right ?

I'm TELLING you to get your arse over here to Burford and do some work,
instead of lounging around on german unemployment pay ........... :)

Chris Whittington

>: to implement it in

> : the next CSTal version, just to show you THAT IT MAKES SENSE to use a
> : more senseful HUMAN way than the 50:50 chances we use today.
>
> : Thanks.
>
>

[Robert Hyatt]

mclane (mcl...@prima.ruhr.de) wrote:
: hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

: >We aren't living in the same world, because you don't have a clue about

: >alpha/beta. The above is perfectly clear to *anyone* that knows beans
: >about alpha/beta minimax. If you don't understand AB, then you sure aren't
: >going to follow the above discussion... Therefore I elect to save my
: >breath...

: I really don't know what alpha/beta might have to do with the question

: how my strategy should be in PB.

: You have not answered to my question WHY all your programs play e.g.
: Bxh3 and expect later a draw by repetition and plan f4 for Kosashvili
: because THEY think f4 is the best move, and than Yona plays better Re5
: instead, and suddenly the evaluation of all programs drop, although
: Re5 should be a weaker move.

If you look at the title of this thread, and you read back through "most"
of the postings, this has been about "what is the correct way to think
on the opponents time?" and why... It doesn't have a thing to do with
any specific game of chess that was played, that was some off-topic
rambling you introduced. If you want to start a thread on this game,
please post the pgn, specific moves you are interested in, and I'll be
happy to join in with whatever I can offer. However the definition of
"thread" is "postings that are about a common topic."..

: You always behave - here you show the same behaviour again -

: as if OTHERS have not understand some main law you know for 40 years,
: and therefore the others (in the PB-discussion me) cannot follow you.

: How do you know that I don't know about AB algorithm ?

: How do you know better about Ed Schroeders Source Codes?

(1) I know that you don't know beans about alpha/beta by reading your
posts here. It is "intuitively obvious to the casual observer" based
on nothing more than your comments and obvious misconceptions. But one
example: saying that some program or another would spend 60% of it's time
thinking about moveA, 10% about move B, etc. Sorry, but that "ain't possible
using alpha/beta" and I am aware of *no* program that doesn't use alpha/beta.
Therefore, the only other explanation is that the claim is incorrect...

(2) You might as well drop the horsecrap about Ed's source codes. I don't
have any of his source. However, there are some things that are certainly
being done by him, and alpha/beta is one of them. There are others. And they
cen be discovered *without* source code, by anyone with enough sense to take
the time. This is true for any program...

: If a program guesses a move XYZ is the best one. And therefore creates

: a tree and a main-line, and it does a PB relying on the 2nd move of
: the main-line, but the move is not coming because the opponent plays
: his own move ABC, then I would expect that the move ABC is losing or
: is a weaker move, because the tree was build in AB-backtracking.

: But the move ABC isn't weaker.

: No I have 2 possibilities:

: a ) the evaluation function is shit
: b) the search-algorithms prune important stuff (MOVE ABC) away and
: oversee something.

: In my example the programs expected f4. And evaluated with draw.
: f4 was replaced by human with Re5 and what would you expect ?
: That the computers' evaluation increases >0.00
: But the evaluation drops.

: Because ALL chess-programs react this way (try crafty yourself)
: you can throw them all in the trash.

: No - I am not interested which algorithm is in charge. If they
: consider about moves that are stupid, and when the better moves come,
: they drop in their evaluations, then something global is wrong...

Yes, but it is *not* a flaw with "PB"... it's a flaw most likely
caused by a lack of depth... you know how to fix that I assume...

: Instead of claiming your: I have known about alpha-beta when you were

: not born prejudice you should concentrate on what I said and not what
: you wanted to read out of my words.

: If your computers would have used the PB time, and instead of assuming
: the shit f4 would come, computed on whatever strange moves, would have
: computed anyhow different , maybe not only looking for the best moves,
: they would MAYBE have seen Re5. If they would have seen it, they could
: have better used PB-time instead of been shocked about the move, and
: then dropped dropped and lost the game.

nonsense... utter nonsense... and again you overlook alpha/beta. If I
were to spread the search time over multiple moves, I would lose probably
two plies of search depth in the same time limit. That's insane...

: When your alpha-beta is so stupid that it cannot handle easy cases of

: chess, than throw it away and create an algorithm that handles these
: cases.

big words. Exactly what do you propose for all of us to use instead of
alpha/beta?

: The best way to spare your breath would be to die. This would spare

: much breath to me either. Please... do it or stop this :

Here's what I'll do. Since you only post drivel much of the time (you
do post interesting stuff at times of course) I'm simply going to stop
reading *anything* you write, then I won't respond to *anything* from you
and you won't have to deal with my comments. That will at least solve my
problem... caio...

: I know something better than you do - behaviour. I am discussing about

: points I can give you examples in chess-notation, games, names and
: whatever. You always fight these : "I have known about TOPIC X when
: you - my young fellow - have made in you clothes" - discussions.
: Do them with KK, he will follow you and believe you magic words. I do
: not believe you. I would never believe somebody who claims he is god
: or a computer or has invented chess-programming-universe.

I have been quite clear about what I've invented, which has not been very
much (rotated bitmaps, time allocation for fail-lows, etc.) However, you,
on the other hand, have such an incredibly high bullshit to information
ratio that skipping your posts really won't be a big loss...

: I can only hope that there comes the day somebody writes an

: intelligent parallel program, and this - with the help of
: zillion-cpu's kicks all of you stupid-fast-.fanatics in the ass and
: send you to the back-side of the moon. There you can play with you
: deep-blue and whatever stupid/hardware-monsters until the last day of
: the universe.

That's what I'd expect from you, "I hope someone else..." Most of us are
trying to do it ourselves... but of course we at least have ideas to try
and understand what has been done in the past...

: Maybe the time comes earlier than you would expect it. From my

: observations it does not take much time until CILK-Chess will show
: your MONSTER-machines that chess has not only to do with hardware ,
: and that arrogance will not help you to hide (=HYATT, maybe the best
: synonym for hiding Artificial Stupidy behind the foggy words) that
: these software running on fast hardware could only survive because
: there was no program near, that was more intelligent.

You really are ignorant. you need to ask Don Dailey about the engine.
I won't say any more. Ask Don. And then report back. And maybe, just
maybe, by then you'll have figured out just how stupid some of your
comments really are. *and* just how *wrong*...

: I don't think that they would ever beat the champion. Maybe Kasparov.

[Robert Hyatt]

mclane (mcl...@prima.ruhr.de) wrote:
: hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

: >: In my opinion the whole discussion is just idle fighting about nothing

: >: than: who is right and who is more wrong.

: >: They would discuss with you that the earth is flat. And they would
: >: show you evidence for it. And they would show you statistics,
: >: photographs and all that stuff, showing that the earth is a disc.

: >Please first write a program, and get up-to-speed on the algorithms that
: >are involved, before posting such drivel. No one's saying the earth is
: >flat. But I can certainly prove that the sqrt(2) = 1.4, or that
: >PI=3.14159265358979323846264...., and in the case of alpha/beta, I've
: >studied the thing long enough to know exactly how it behaves... something
: >you obviously don't understand.

: Unbelievable.
: You know exactly what ?
: That you permanent-brain on a move that will not come ?
: And that you try to convince me that it is very effective to compute
: on a move that will not come. Are you nuts ?

: >Every time you disagree, you go from the point at hand, to something like the
: >earth is flat, to the next level of hyperbole... And the content (meat) of
: >the comments varies inversely in proportion to the hyperbole... Again,
: >shoot my math down, or pipe down... take your choice... but don't start
: >again with the name-calling bullshit...

: I don't have to write a chess program to find out that somebody else
: is an idiot.

I don't believe you are smart enough to find out someone is an idiot.
You have to be smarter than that person. I can't think of even one you
could make that claim about...

[Robert Hyatt]

Komputer Korner (kor...@netcom.ca) wrote:

: >

: > If you analyze alpha/beta, searching the position below *one* move is about
: > twice as efficient as searching the positions below *two* moves, because if
: > you ponder one move, you only have to produce *one* PV, which is very
: > efficient. If you ponder two different moves, you have to produce two
: > different PV's, which is about 1/2 as efficient as only producing one.
: >
: > Therefore, searching (pondering) two moves by the opponent is a total loss
: > in terms of time saved, when you compare everything...
: >

: > If that's not clear enough, feel free to ask for more details...

: Ingo Althofer has produced stats showing that the k variation mode has

: been improved so that it produces efficiencies so that the total time
: taken
: is less than twice the 1 variation mode even when K is large. However
: your
: first point is VERY true. There is no advantage in pondering more than 1
: move assuming that the program can think normally if it guesses wrongly
: on
: the previous pondering think. However, The original poster postulated a
: Deep Blue strategy of ALWAYS moving instantly. This would result in some
: situations of the program having to make a move after a 12 ply instant
: search instead of its 16 ply average after thinking for 3 minutes if it

: guessed wrongly. 12 ply is not enough to defeat Kasparov in certain
: critical
: middlegame decisions.

There's no way, using alpha/beta, to search two moves and get a PV and
score for each, without each of these two moves taking roughly exactly
the same amount of time as the other move. The hash table might figure
in, or it might not. For K very large, the math won't work... here's
the short explanation:

For an alpha/beta search of a position where the score is not known
(the case we always have of course) the number of terminal positions
is as follows:

N=W^floor(D/2)+W^ceil(D/2)

floor(n) = integer division round *down*, while ceil(n) = integer
division, round *up*.

For D even, floor(D/2) == ceil(D/2) so the formula can be collapsed
to the typical shortened form

N=2*W^(D/2) although I am going to use the correct formula for the
rest of this, the function given earlier for N.

Now for the math:

For searching the first branch, we use this formula recursively,
because for the first successor to the first root move we search,
we still don't know beans about the score. So we get this

NFB=W^floor((D-1)/2)+w^ceil((D-1)/2)

(NFB=Number of terminal nodes on the First Branch)

Now that we know how many tip positions for the first branch only, and we
also know how many tip positions for the complete tree, we can compute

NRB=(N-NFB)/(W-1)

for the following reason. N is the total nodes to be searched for a normal
alpha/beta search to depth D, with a branching factor of W, which is typically
around 38 in chess. NFB is the total nodes for the subtree below the first
root move. If we subtract these, and then divide by the number of moves
that are left after searching the first root move (W-1) we can compute the
size of the tree for the remaining moves at the root.

Compare the following two numbers:

NFB to NRB. Lets pick a depth, say 11 plies. And lets stick with the
traditional value for W of 38... my calculator then produces

NFB=2*38^5=300,000,000 nodes (roughly)

NRB=38*5+38*5/(37)-300,000,000=300,000,000/37=8,100,000 nodes for each
of the remaining branches. Compare the two numbers now...

300M compared to 8M is a factor of 37, which is interesting. This means
that it takes *exactly* as long to find a PV for the first move at the root
as it does to search *all* of the remaining W-1 (37) moves *together*. So,
if you want to find a PV for the second move, you are going to increase
the total nodes searched by 50%, since you will have two first branches, and
36 remaining branches. If you find the PV's for the first 3, you just doubled
the size of the tree. There's not any way to work around that problem. Each
extra PV you find has a huge cost. If you ponder two moves for the opponent,
you now can see what happens. It will cost you 50% more time to reach the
same depth on *both* moves as it would cost you to reach the same depth on
only one move.

But it is *much* worse. Because after the first move, we are now free to
factor in null move. Let's take a sample case from crafty, using the usual
R=2 null move, to compare it to the above, if you'd rather have some
empirical data rather than the fairly simple math above (which comes from two
papers, BTW, "An analysis of alpha-beta pruning" by Knuth and Moore, 1975, and
"A parallel alpha/beta tree searching algorithm", Parallel Computing 10, Hyatt,
Nelson and Suter.)

10-> 35.11 -0.513 Qd8 f3 Be7 axb5 axb5 Qf2 Rxa3 Rxa3
Rb8 Ra7 Bg5 Bxg5 Nxg5
11 2:42 -0.494 Qd8 f3 Be7 Qf2 Bh4 axb5 axb5 Ra6 Nf6
Ra7 Rxa7 Rxa7 Ra8
11-> 2:51 -0.494 Qd8 f3 Be7 Qf2 Bh4 axb5 axb5 Ra6 Nf6
Ra7 Rxa7 Rxa7 Ra8

The above came from one of the positions in the second DB/Kasparov match.

Notice that Crafty finished the 10 ply search at 35.11 seconds. (I picked
this position since the depth matches the example from above for an 11 ply
search.) Notice that the 11 ply PV takes 2 minutes and 7 seconds of additional
time to get the 11 ply PV. And then notice that the remaining moves took only
9 seconds *total* to dismiss as hopeless, thanks to the rather significant
help null-move gives.

In this position, assume you want to produce the PV for two different moves,
not just Qd8, but Qd8 + another move. That is not just bad, it is *very*
bad. Qd8 takes 2 minutes to produce the PV, the second move takes 2 minutes
to produce a PV, and the remaining W-2 moves will take a few seconds. You
just *doubled* the search time. If you want the PV for 3 moves, you triple
the search time. And there's not any way to get around this when using
alpha/beta..

There you have it. The math to explain why current pondering is the way to
go based on tree sizes, and some empirical data to support the conclusion as
well. Note that the math I gave ignores null-move, but that factoring it in
only makes the first branch even *larger* with respect to the remaining branches
as the empirical data shows...

Hope that helps...

Bob

[Komputer Korner]

mclane wrote:
snipped

>
> Dear KK: If I would count all the examples where you have made a fool
> out of yourself, I would have to count a long long time.
>
> Thats typical for you. Please do believe in his words, but let me
> alone with my ideas. I have not said what I think about your postings
> because I thought: ough - he has too learn much. 89 % of what he is
> posting is wrong. So I watched how the others corrected you.

I sent you a private email so that I wouldn't embarrass you in the
r.g.c.c. because often you have a lot to contribute here. However
often you put your own foot in your mouth and you certainly are
doing it on this thread. It would be very interesting to see who has
made more mistakes in postings, me or you, but then you never admit
when you are wrong and that is a shame.

[Komputer Korner]

chrisw wrote:

> Stand by to be flamed Thorsten.
>
> You don't TELL me to do anything, right ?
>
> I'm TELLING you to get your arse over here to Burford and do some work,
> instead of lounging around on german unemployment pay ........... :)
>
> Chris Whittington
>
> >: to implement it in
> > : the next CSTal version, just to show you THAT IT MAKES SENSE to use a
> > : more senseful HUMAN way than the 50:50 chances we use today.
> >
> > : Thanks.
> >
> >
> > Suit yourself of course. But Chris can do the math and determine that
> any
> > other method is inferior, unless the prediction rate is down in the
> 10-20%
> > range, something I've not seen from any program...
> >
> >
> >

Sorry Chris for including some more CSTAL beta comments here, but
emailing
you directly is a risky business and most of these comments have been
made
before.
1) When you exit the program , it doesn't save the player settings
human- human. A lot of time I like to experimnt on monitor mode thus
it is a pain to have to switch back each time.
2) I created an opening book in CSTAL from a FISCHER database of moves.
I pruned it off at 3 ply. I tried to add lines beyond that but no lines
ever got added. When the Learn line box came up I clicked on line good
for both. Sure the evaluation score could be changed for those
first 3 ply but I couldn't add any moves beyond that. I didn't
optimize the book but except for speed that shouldn't matter. I couldn't
optimize it anyway, because the optimize program wasn't there. I
searched for it under both the spellings. I then
created another book with 4 ply. Same story. The correct
user book was specified in book control. You say that the latest beta
has not changed regarding the opening book. What gives here?

[Komputer Korner]

Robert Hyatt wrote:
snipped
snipped

>
> In this position, assume you want to produce the PV for two different moves,
> not just Qd8, but Qd8 + another move. That is not just bad, it is *very*
> bad. Qd8 takes 2 minutes to produce the PV, the second move takes 2 minutes
> to produce a PV, and the remaining W-2 moves will take a few seconds. You
> just *doubled* the search time. If you want the PV for 3 moves, you triple
> the search time. And there's not any way to get around this when using
> alpha/beta..
>
> There you have it. The math to explain why current pondering is the way to
> go based on tree sizes, and some empirical data to support the conclusion as
> well. Note that the math I gave ignores null-move, but that factoring it in
> only makes the first branch even *larger* with respect to the remaining branches
> as the empirical data shows...
>
> Hope that helps...
>
> Bob

Thanks for the very detailed explanation, but then why did Ingo
Althofer say that he had given Frans Morsch some ideas that
enabled a significant speedup in k variation mode compared to your
conclusions? He did make the statement that you and I read when we were
discussing the 3 HIRN experiments. I did some testing on the
Hiarcs 6 engine in Extreme and on my Pentium 133 in 11 variation mode
it took 55 seconds to complete 8 ply calculating from the initial
position in infinite mode. On 10 variation mode it took 53 seconds.
Perhaps I counted wrongly here.
On 9 variation mode it took 55 seconds !!!!! On 8 variation mode it
took 49 seconds. On 7 variation mode it took 42 seconds.
On 6 variation mode it took 33 seconds. On 5 variation mode it took
28 seconds. On 4 variation mode it took 24 seconds. On 3 variation mode
it took 21 seconds. On 2 variation mode it took 17 seconds. On 1
variation it took 12 seconds to finish the 8 ply. These numbers don't
jive with your conclusions.

[Robert Hyatt]

Komputer Korner (kor...@netcom.ca) wrote:
: Robert Hyatt wrote:
: snipped

: snipped
: >
: > In this position, assume you want to produce the PV for two different moves,

: > not just Qd8, but Qd8 + another move. That is not just bad, it is *very*
: > bad. Qd8 takes 2 minutes to produce the PV, the second move takes 2 minutes
: > to produce a PV, and the remaining W-2 moves will take a few seconds. You
: > just *doubled* the search time. If you want the PV for 3 moves, you triple
: > the search time. And there's not any way to get around this when using
: > alpha/beta..
: >
: > There you have it. The math to explain why current pondering is the way to
: > go based on tree sizes, and some empirical data to support the conclusion as
: > well. Note that the math I gave ignores null-move, but that factoring it in
: > only makes the first branch even *larger* with respect to the remaining branches
: > as the empirical data shows...
: >
: > Hope that helps...
: >
: > Bob

: Thanks for the very detailed explanation, but then why did Ingo

: Althofer say that he had given Frans Morsch some ideas that
: enabled a significant speedup in k variation mode compared to your
: conclusions? He did make the statement that you and I read when we were
: discussing the 3 HIRN experiments. I did some testing on the
: Hiarcs 6 engine in Extreme and on my Pentium 133 in 11 variation mode
: it took 55 seconds to complete 8 ply calculating from the initial
: position in infinite mode. On 10 variation mode it took 53 seconds.
: Perhaps I counted wrongly here.
: On 9 variation mode it took 55 seconds !!!!! On 8 variation mode it
: took 49 seconds. On 7 variation mode it took 42 seconds.
: On 6 variation mode it took 33 seconds. On 5 variation mode it took
: 28 seconds. On 4 variation mode it took 24 seconds. On 3 variation mode
: it took 21 seconds. On 2 variation mode it took 17 seconds. On 1
: variation it took 12 seconds to finish the 8 ply. These numbers don't
: jive with your conclusions.
: --
: Komputer Korner

: The inkompetent komputer.

They do, but a lot has to do with the engine being used. If a program
is simply full-width, then it *must* match my numbers, because alpha/beta
is a simple mathematical law that can not be relaxed for any reason. If
a program is selective, it changes. In your case, you also picked 8 plies
rather than an odd number. Run your 8 ply initial search (D=8) thru the
numbers I posted. Even ply searches have less disparity in node counts
because of the odd/even formula (floor()... and ceil()...) So that for
even ply searches, the difference between the time to search the first
move and the remaining moves. So the overhead is somewhat less. However,
the point is still there... compare your 12 seconds to 60 seconds. That is
still 5:1 and is a horrible handicap, because that only costs you about 1.5
plies of search depth. A *huge* cost.

On second thought, maybe I should do the calculations? There are a couple of
sqrt's in there, and with your history of sqrt() problems... :) (just kidding)

[Robert Hyatt]

Komputer Korner (kor...@netcom.ca) wrote:
: Robert Hyatt wrote:
: snipped

: snipped
: >
: > In this position, assume you want to produce the PV for two different moves,

: > not just Qd8, but Qd8 + another move. That is not just bad, it is *very*
: > bad. Qd8 takes 2 minutes to produce the PV, the second move takes 2 minutes
: > to produce a PV, and the remaining W-2 moves will take a few seconds. You
: > just *doubled* the search time. If you want the PV for 3 moves, you triple
: > the search time. And there's not any way to get around this when using
: > alpha/beta..
: >
: > There you have it. The math to explain why current pondering is the way to
: > go based on tree sizes, and some empirical data to support the conclusion as
: > well. Note that the math I gave ignores null-move, but that factoring it in
: > only makes the first branch even *larger* with respect to the remaining branches
: > as the empirical data shows...
: >
: > Hope that helps...
: >
: > Bob

: Thanks for the very detailed explanation, but then why did Ingo

: Althofer say that he had given Frans Morsch some ideas that
: enabled a significant speedup in k variation mode compared to your
: conclusions? He did make the statement that you and I read when we were
: discussing the 3 HIRN experiments. I did some testing on the
: Hiarcs 6 engine in Extreme and on my Pentium 133 in 11 variation mode
: it took 55 seconds to complete 8 ply calculating from the initial
: position in infinite mode. On 10 variation mode it took 53 seconds.
: Perhaps I counted wrongly here.
: On 9 variation mode it took 55 seconds !!!!! On 8 variation mode it
: took 49 seconds. On 7 variation mode it took 42 seconds.
: On 6 variation mode it took 33 seconds. On 5 variation mode it took
: 28 seconds. On 4 variation mode it took 24 seconds. On 3 variation mode
: it took 21 seconds. On 2 variation mode it took 17 seconds. On 1
: variation it took 12 seconds to finish the 8 ply. These numbers don't
: jive with your conclusions.
: --
: Komputer Korner

: The inkompetent komputer.

also the anomoly above is *bad* news. Here is what Fritz is producing,
and you don't realize it: When it searches the first move, it stores
stuff in the hash table. It then uses this in the second search, which
will influence the final score somewhat. In the multi-variation mode in
Crafty, I specifically clear the hash table so that the multiple PV's you
see are "untainted by their cousins"... Otherwise the last PV has a lot
of extra information that the first PV didn't, and if you were to go back
and research the first move with the same hash table, the PV there would
likely change. It's a "fudge" that is not bad, but not completely
accurate as it is being presented, unless you clear things between *every*
PV search... Change the move order at the root, and you change some of
the PV's, which is ugly...

[Tom C. Kerrigan]

Komputer Korner (kor...@netcom.ca) wrote:

> Tom, the argument here has 2 threads.
> a) Is pondering or Permanent Brain less efficient overall than normal
> thinking time? The answer is NO, as Bob Hyatt has proved in another
> post.

In other words, I'm right.

> flawed unless Deep Blue'e guess rate is 100%. Flawed because by

In other words, I'm right again, considering the original post. I can see
how you get tired of refuting my arguments. You just never win. That can't
be fun.

As I recall, the original post suggested pondering multiple moves, which
could easily bring the guess rate to 100%. No flaws.

Cheers,
Tom

[Komputer Korner]

Robert Hyatt wrote:
>
> Komputer Korner (kor...@netcom.ca) wrote:
> : Robert Hyatt wrote:
> : snipped

>

> also the anomoly above is *bad* news. Here is what Fritz is producing,
> and you don't realize it: When it searches the first move, it stores
> stuff in the hash table. It then uses this in the second search, which
> will influence the final score somewhat. In the multi-variation mode in
> Crafty, I specifically clear the hash table so that the multiple PV's you
> see are "untainted by their cousins"... Otherwise the last PV has a lot
> of extra information that the first PV didn't, and if you were to go back
> and research the first move with the same hash table, the PV there would
> likely change. It's a "fudge" that is not bad, but not completely
> accurate as it is being presented, unless you clear things between *every*
> PV search... Change the move order at the root, and you change some of
> the PV's, which is ugly...

Well, hash tables or not, it was my Hiarcs engine that showed this.
I imagine Fritz would have the same numbers prorated for the
difference in speed. I could test Fritz in Chessbase but the dll is
slower
than an engine. So the end result that with hash tables and k
variation feature, the idea of only 1 principal variation on pondering
may
not be the way to go. On 4 variations we get an average of 6 sec per
variation which is better than 12 seconds for one variation. Are you
saying that the difference is all because of hash hits?

[Komputer Korner]

Tom C. Kerrigan wrote:
>
> Komputer Korner (kor...@netcom.ca) wrote:
>

How would moving instantly for several moves in a row and only
pondering during this time increase one's chances unless one was
able to achieve a 100% GUESS rate?

[Tom C. Kerrigan]

Komputer Korner (kor...@netcom.ca) wrote:

> > As I recall, the original post suggested pondering multiple moves, which
> > could easily bring the guess rate to 100%. No flaws.

> How would moving instantly for several moves in a row and only
> pondering during this time increase one's chances unless one was
> able to achieve a 100% GUESS rate?

Interesting... after half a dozen flames in every direction, you finally
grasp the intent of the original post...

To refresh your memory... Kasparov is also thinking on DB's time. If DB
doesn't take any time, Kasparov doesn't think as long and plays a weeker
game. To not take any time, DB has to ponder several moves simultaneously,
which hits its speed, so it plays a weaker game too. The original poster
was wondering if this was a net gain or a net loss.

My view is that it could be a net gain for DB if it pondered the
best-looking handful of moves. That way, it "only" gets hit a ply, whereas
Kasparov gets half his time sucked away.

Now that you understand what's going on, I hope, perhaps you have some
views of your own to share?

BTW, any programmer worth spit could achieve a 100% "guess" rate by
pondering all the moves.

Cheers,
Tom

[Robert Hyatt]

Komputer Korner (kor...@netcom.ca) wrote:
: Robert Hyatt wrote:
: >

: > Komputer Korner (kor...@netcom.ca) wrote:
: > : Robert Hyatt wrote:
: > : snipped

: >
: > also the anomoly above is *bad* news. Here is what Fritz is producing,

The he has the same problem. I received lots of reports, when I
first did the N-variation code that went like this:

Bob, the search pv/score values look funny.. I get:

.321 xxxxxxxxxxxxxxxxx
.300 yyyyyyyyyyyyyyyyy
.336 zzzzzzzzzzzzzzzzz

why is the last score higher than the supposedly best score? The
answer is that the last score had information from the first two
PV's that could influence it. When I cleared the hash table after
each PV, the problem went away as expected. If you don't clear the
hash, you get faster searches, but non-uniform values. If you do
clear the hash table, the performance loss is serious (it's bad
enough when you don't clear it...)

[Robert Hyatt]

Komputer Korner (kor...@netcom.ca) wrote:

: Tom C. Kerrigan wrote:
: >
: > Komputer Korner (kor...@netcom.ca) wrote:
: >

: > > Tom, the argument here has 2 threads.

: > > a) Is pondering or Permanent Brain less efficient overall than normal
: > > thinking time? The answer is NO, as Bob Hyatt has proved in another
: > > post.
: >
: > In other words, I'm right.
: >
: > > flawed unless Deep Blue'e guess rate is 100%. Flawed because by
: >
: > In other words, I'm right again, considering the original post. I can see
: > how you get tired of refuting my arguments. You just never win. That can't
: > be fun.

: >
: > As I recall, the original post suggested pondering multiple moves, which

: > could easily bring the guess rate to 100%. No flaws.

: >
: > Cheers,
: > Tom

: How would moving instantly for several moves in a row and only

: pondering during this time increase one's chances unless one was
: able to achieve a 100% GUESS rate?

: --
: Komputer Korner

: The inkompetent komputer.

only way to achieve 100% is to ponder *every* possible opponent move, then
you can't move instantly... because you've only searched a reply to his move
for < 10% of the total time you want to search before making a move... Why
is this topic still going on???

[Eric Wang]

Komputer Korner (kor...@netcom.ca) wrote:
>> This strategy is flawed for the following 2 reasons.
>> 1) Kasparov will then have longer thinking times than opponent.

kerr...@merlin.pn.org (Tom C. Kerrigan) writes:
> Maybe in your universe. In my universe, X units of time is generally
> just about as long as X units of time.

You mis-read it, Tom. The strategy KK mentioned was that Deep Blue
should intentionally play at 10 seconds per move to deny Kasparov
additional thinking time. Obviously, if Deep Blue thinks for ten
seconds per move and Kasparov thinks for three minutes per move,
whatever advantage Kasparov already has over Deep Blue would only be
magnified. Most observers (at the match and on this newsgroup)
think Deep Blue at three minutes per move still isn't quite at
Kasparov's level -- maybe in a few years or decades, but not yet.
Forcing Deep Blue to play at blitz speed would hurt Deep Blue far
more than it hurts Kasparov. Thus, this "strategy" is flawed.
(Forcing both of them to play blitz might be interesting -- Frederic
Friedel reports that Kasparov loses to Fritz at blitz speeds at a
small but non-zero rate, IIRC around 5-6 losses per 100 games.
Against DB, it would probably be somewhat higher.)

> Dear Lord, please think before posting...

This applies more to you than to KK.

>> 2) Deep Blue's P.B. can not match Kasparov's on the move normal
>> thinking process. Thinking on the opponents time is not the computer
>> programs's strong area.

>This is silly. If DB doesn't think on the opponent's time the same way
>every other program does it, I will buy a hat so I can eat it.

No hat necessary; however, KK is not wrong, either, you simply
mis-read him again. __Of course__ DB thinks on its opponent's time;
that's beyond question. KK simply noted that DB is __not as
effective__ thinking on its opponent's time (when it has to divide
its time among several possible responses) as it is when it thinks
on its own time (i.e. after its opponent has moved), and this is
generally true of all chess programs (and probably true of all human
players, too). As pointed out by the other responses in this thread
(and the announcers at the match), strong chess programs generally
anticipate their opponents' move about 50% of the time, so their
resulting analysis is directly relevant and they get all the
advantages of thinking on their opponent's time. The other 50% of
the time, their opponent plays something else (which, by the nature
of alpha-beta search, is either an "inferior" move, or is deeper
than the search's horizon), and the program has to start
re-calculating from scratch. The stated strategy of having DB move
at blitz speeds fails because DB would then rely almost entirely on
GK's time for its own thinks, which isn't as cost-effective as
simply using its own time.

Eric Wang
wan...@uiuc.edu

[mclane]

hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

>: I don't have to write a chess program to find out that somebody else
>: is an idiot.

>I don't believe you are smart enough to find out someone is an idiot.
>You have to be smarter than that person. I can't think of even one you
>could make that claim about...

Rubbish !

You don't have to be raped to find out that raping is not ok, you
don't have to be smart to find out that somebody is an idiot.

Even:
You don't have to be a programmer to understand about programs.
You don't have to be a writer to criticize about books.
You don't have to be a song-writer to sing better than Julio Iglesias
You don't have to build deep blue to find out the machine is garbage
You don't have to a chess player to find out that this match is
arranged.

Your "society" want to stop people saying something that is against
your ideas just because this people are not programmers.

Thats brilliant. Then of course we are not allowed to say anything
about the turkish-state killing the kurdish-inhabitants, just because
we are no turks.

Thats not my society. My society is open for anybody.

[mclane]

"chrisw" <chr...@cpsoft.demon.co.uk> wrote:

>> : If Bob is against this method, I will tell Chris

>Stand by to be flamed Thorsten.

>You don't TELL me to do anything, right ?

AH brilliant. Thats what friend are for. Helping themselves in the bad
times, sharing emotions and life.

>I'm TELLING you to get your arse over here to Burford and do some work,
>instead of lounging around on german unemployment pay ........... :)

Why should I go to britain and work WITH you (I would never work FOR
you!!) when you know anything better and nobody, including me, would
be allowed to tell you anything ????

Yeah - you and Bob are very similar in this behaviour !
You don't let anybody tell you anything....

>Chris Whittington

[mclane]

hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

> And there's not any way to get around this when using
>alpha/beta..

>There you have it. The math to explain why current pondering is the way to
>go based on tree sizes, and some empirical data to support the conclusion as
>well. Note that the math I gave ignores null-move, but that factoring it in
>only makes the first branch even *larger* with respect to the remaining branches
>as the empirical data shows...

>Hope that helps...

>Bob

Brilliant Bob - you declare it, and we have to believe it. Your main
point is, and that makes the whole thing worthless:
WHEN USING ALPHA BETA.

Again I want to focus your head on the following game, I had to
operate in Den Hague...

[Event "AEGON Chess Tournament"]
[Site "The Hague NED"]
[Date "1997.04.16"]
[Round "1"]
[White "Yona Kosashvili"]
[Black "CHESS SYSTEM TAL"]
[Result "1-0"]
[WhiteTitle "GM"]
[WhiteElo "2560"]
[WhiteCountry "ISR"]

1. e4 c5 2. Nf3 e6 3. c4 Nc6 4. Nc3 Nf6 5. Be2 d5 6. exd5 exd5 7. d4
Be6 8.
cxd5 Nxd5 9. O-O Nxc3 10. bxc3 cxd4 11. Nxd4 Nxd4 12. cxd4 Bd6 13.
Bb5+ Kf8
14. Re1 Qh4 15. h3 Bxh3 16. gxh3 Qxh3 17. Re5 Bxe5 18. dxe5 Qf5 19.
Qd4 Kg8
20. Be3 Qg6+ 21. Kf1 a6 22. Rd1 h6 23. Bd7 b5 24. Qd5 Rf8 25. Bc5 Rd8
26. e6
fxe6 27. Bxe6+ Kh7 28. Bf5 Rxd5 29. Bxg6+ Kxg6 30. Rxd5 Rc8 31. Rd6+
Kf5 32.
Bb4 Ra8 33. Bc3 g6 34. Kg2 Ra7 35. Rc6 h5 36. Rf6+ Kg5 37. Kh3 Re7 38.
f4+
Kh6 39. Be5 a5 40. Ra6 Rf7 41. Ra8 1-0

Maybe you could input this little game into your Crafty, and others
into their programs.

Maybe this makes the whole discussion somehow more after my taste,
because the way you handle a discussion is more the way moses was
presented the 10 commandments , by god.

I don't like this pedagogical way of yours.

Let crafty (and anybody if he has another program can try it) compute
on move 15...bxh3.

The programs (some, e.g. Hiarcs6 plays a6, but e.g. Fritz4 and other
programs) play Bxh3.
They think this could be a nice attack.

In the 17th move they all consider normally with f4 as the
permanent-brain move and different repetition-main-lines , sometimes
with draw score.
Why? After they throw the bishop into h3, they later find that white
is weaker. Because white is weaker they think : it should repeat the
position !
Of course when white plays f4 and would really repeat the position,
than CSTal and especially me would have been very happy.

Because this would have meant that CSTal would have made a draw
against the tournament-winner of Den Hague, Yona Kosashvili, who
played 6 out of 6, pardon, if our stupid friends would be right, 5.5
out of 6.

I would have been pleased with such a result. Nice - but very easy
bishop sac, and check check check and draw!

Now lets see: all programs think that 17.f4 has to come, which allows
the black queen some checks. And because the programs think white has
a weaker position they are happy with a draw, if they see the chance.

Brillian Alpha-Beta logic. If this is the result of 40 years working
mankind on chess, and you tell me that anything like this is OK, then
of course research further 40 years and I will not longer focus on
computerchess. I would then say: let these old guys research with
their crazy alpha-beta and I would go sleep.

Helpfully Yona Kosashvilil did not what the stupid-programs expected.
He did a very normal and easy move, many chess-players would have
done: gave back the Rook with 17.Re5.

This is "brilliant" because we have a good example why machines are
stupid. And also it stops possible treatment by blacks troops, as
chris would express himself.

The machines wasted permanent-brain computing on f4 main-lines, mainly
they found out that the position will end in a draw.

Suddenly Yona plays a WEAKER move than the BEST move f4, and and ?
The evaluation drops after this move has been made, and they would
maybe all lose the position because black has ONLY A QUEEN against an
army of white pieces threatening the black king.
But a queen alone, cannot mate without the help of another piece ! But
which piece ?

SO: this is my chess-example to the discussion about strategies and
permanent brain. And I am pleased to see your results.

Why did the programs lost although the opponent made the weaker move.
Because after your Rebel-Nimzo match you are an EXPERT for explaining
why your craty lost although Rebel made WEAKER moves, I am pleased to
hear from you further.

Istead of telling the people that one can better discuss with a tree ,
you should discuss (what would be ON TOPIC!) this above CHESS-TREE
stuff. I hope you can satisfy me , and the others.

I hope - because this was my point to the discussion - that we can
learn something about the mistakes of others (here Chess System Tal's
mistake of throwing Bxh3).

When I am right, Yona played h3 by purpose, to let us run into this
FALSE sac. He could have played g3, but in the moment he played h3, I
thought: he did it by purpose. Maybe Yona has seen Re5 while playing
h3.

[mclane]

hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

>: You have not answered to my question WHY all your programs play e.g.
>: Bxh3 and expect later a draw by repetition and plan f4 for Kosashvili
>: because THEY think f4 is the best move, and than Yona plays better Re5
>: instead, and suddenly the evaluation of all programs drop, although
>: Re5 should be a weaker move.

hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
If you look at the title of this thread, and you read back through
"most"
of the postings, this has been about "what is the correct way to think
on the opponents time?" and why... It doesn't have a thing to do
with
any specific game of chess that was played, that was some off-topic
rambling you introduced. If you want to start a thread on this game,
please post the pgn, specific moves you are interested in, and I'll be
happy to join in with whatever I can offer. However the definition of
"thread" is "postings that are about a common topic."..

My answer:
You are trying to spray fog arround. Also this is very rude. You don't
have to explain me in a discussion what I am supposed to do, because
this is off-topic.
The game I speak about HAS TO DO WITH THE TOPIC we/you and others talk
about. The problem is: you don't see this.

I don't want to start a thread about this game, because the game
itself is irrelevant. It is just an example.
I don't like your theoretical way of talking about topics the way many
old people talk when they meat younger people or less experienced
people: you talk the way:

I have known about these facts when you were not born.

Although this is very nice, and the old guys seem to need some
youngsters looking brave and with open mouth to the stories the old
guys talk wisely, this is a communication-setting I would only allow
to people who are wise. Some people get white hair without any
wiseness. They are just old.

>: You always behave - here you show the same behaviour again -
>: as if OTHERS have not understand some main law you know for 40 years,
>: and therefore the others (in the PB-discussion me) cannot follow you.
>: How do you know that I don't know about AB algorithm ?
>: How do you know better about Ed Schroeders Source Codes?

Hyatt answered:

(1) I know that you don't know beans about alpha/beta by reading your
posts here. It is "intuitively obvious to the casual observer" based
on nothing more than your comments and obvious misconceptions. But
one
example: saying that some program or another would spend 60% of it's
time
thinking about moveA, 10% about move B, etc. Sorry, but that "ain't
possible
using alpha/beta" and I am aware of *no* program that doesn't use
alpha/beta.
Therefore, the only other explanation is that the claim is
incorrect...

My comments on that:
Again I must say, reading your above stuff: you are crazy ! Absolutely
crazy and full of emotinal prejudice.

My program is in a position. I let it compute about move a for a
while. I save the main-line and the evaluation.
Then I take back.
I make next best or cut the move out of the move list and let it
compute about the next move. Lets say 10 % of the time of move A.
Then I write down the main-lines, the evaluation.
Then I take back the move B.
Now I start another search, by cutting move A/B from the move list or
easier , by doing another search on the position.
This way doing it, the programs would have registered Re5 is good, in
the permanent brain time of move 17!!
They would find out that Re5 is better than the supposed f4, and now
the clue comes:
Using this method they would resort the branches and instead of
wasting more time, they could try answers on Re5 meanwhile.
I know that after Re5 THERE ARE NO ANSWERS ANYMORE:
but there could be positions in theory where the right change of the
PB move could help.

Hyatt said:
(2) You might as well drop the horsecrap about Ed's source codes. I
don't
have any of his source. However, there are some things that are
certainly
being done by him, and alpha/beta is one of them. There are others.
And they
cen be discovered *without* source code, by anyone with enough sense
to take
the time. This is true for any program...

My answer:
Right. Thats what I am doing. I have to find out about programs
without seeing the sources. And I don't know why you know better than
I do, in this case. This was the discussion we have. And your way of
saying: IT IS NOT POSSIBLE. is not helping much.

>: If your computers would have used the PB time, and instead of assuming
>: the shit f4 would come, computed on whatever strange moves, would have
>: computed anyhow different , maybe not only looking for the best moves,
>: they would MAYBE have seen Re5. If they would have seen it, they could
>: have better used PB-time instead of been shocked about the move, and
>: then dropped dropped and lost the game.

Hyatts words:

nonsense... utter nonsense... and again you overlook alpha/beta. If
I
were to spread the search time over multiple moves, I would lose
probably
two plies of search depth in the same time limit. That's insane...

My comments:
The nonsense is coming from your side.
If you relying on ONE move, this makes only sense, if this move is
often a hit.
That would be the case if chess would be a game, where many moves
would be forced. But in chess we have very often situations where not
only ONE move is the best, but a few moves are maybe equal.
I guess ANY chess player can subscibe to this. Also you must have
recognized this doing your job.

So I assume that we have normally a chance of 1:3 that the move YOU do
your pb about, is the right move.
This is not much. Your two plies do not help here anyway, because YOUR
program would NOT get the 2 plies. It would have to compute by sero
because it has chosen the wrong PB-move.

>: When your alpha-beta is so stupid that it cannot handle easy cases of
>: chess, than throw it away and create an algorithm that handles these
>: cases.

>big words. Exactly what do you propose for all of us to use instead of
>alpha/beta?

I would propose that an algorithm that oversees main and important
moves of the opponent, just because the opponent is not always playing
the BEST move, is a shit algorithm.
Especially when I know that in chess there are no best moves , this
algorithm fails. Of course in your theoretical world, where your chess
program is always RIGHT, there your alpha - beta always works, but not
in Den Hague...

I would propose to use your head and think about the problem.
Maybe this would be the first step. No chess player relys on only one
PB-move, and you should consider that this makes sense.

>: The best way to spare your breath would be to die. This would spare
>: much breath to me either. Please... do it or stop this :

>Here's what I'll do. Since you only post drivel much of the time (you
>do post interesting stuff at times of course) I'm simply going to stop
>reading *anything* you write, then I won't respond to *anything* from you
>and you won't have to deal with my comments. That will at least solve my
>problem... caio...

Brilliant. Thats exactly the method I prejudice about you:
Seeing a problem. Not able to handle it. Closing the eyes and putting
the head into the ground.

I wish you many years doing this method. Because this way one time
Ananse will have overtaken you...

Take my regards for this day in forward !

>: I can only hope that there comes the day somebody writes an
>: intelligent parallel program, and this - with the help of
>: zillion-cpu's kicks all of you stupid-fast-.fanatics in the ass and
>: send you to the back-side of the moon. There you can play with you
>: deep-blue and whatever stupid/hardware-monsters until the last day of
>: the universe.

>That's what I'd expect from you, "I hope someone else..." Most of us are
>trying to do it ourselves... but of course we at least have ideas to try
>and understand what has been done in the past...

The whole discussion shows me, that the amount of ideas in your mind
seems to be limitated on things you have practised well over the
years.
As my old boss always have said: let us do it the usual way, this way
worked all the years before.

>: Maybe the time comes earlier than you would expect it. From my
>: observations it does not take much time until CILK-Chess will show
>: your MONSTER-machines that chess has not only to do with hardware ,
>: and that arrogance will not help you to hide (=HYATT, maybe the best
>: synonym for hiding Artificial Stupidy behind the foggy words) that
>: these software running on fast hardware could only survive because
>: there was no program near, that was more intelligent.

>You really are ignorant. you need to ask Don Dailey about the engine.
>I won't say any more. Ask Don. And then report back. And maybe, just
>maybe, by then you'll have figured out just how stupid some of your
>comments really are. *and* just how *wrong*...

I have talked with him in Den Hague. He is not that ignorant you are
!! :-)

[Robert Hyatt]

mclane (mcl...@prima.ruhr.de) wrote:
: hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

: > And there's not any way to get around this when using
: >alpha/beta..

: >There you have it. The math to explain why current pondering is the way to
: >go based on tree sizes, and some empirical data to support the conclusion as
: >well. Note that the math I gave ignores null-move, but that factoring it in
: >only makes the first branch even *larger* with respect to the remaining branches
: >as the empirical data shows...

: >Hope that helps...

: >Bob

: Brilliant Bob - you declare it, and we have to believe it. Your main

: point is, and that makes the whole thing worthless:
: WHEN USING ALPHA BETA.

Then you trot out a program that doesn't use alpha/beta and we'll talk.
Hint: talk to Chris. He *does* use it... So I suppose I have somehow
(again) overlooked some subtle point to your post, but danged if I can
figure out what it is. Why post a CSTal game when it is subject to the
exact limits I posted. Do you claim it ponders (or permanent brain's)
"differently"???

: Again I want to focus your head on the following game, I had to
: operate in Den Hague...

To quote the warden in "Cool Hand Luke"... "What we have here is a
failure to communicate..." There was *no* pedagogy in my post, only
pure, irreftuable mathematics. If you can't follow, that's o.k... but
don't make stupid statements... in this case math is absolute...

: Let crafty (and anybody if he has another program can try it) compute
: on move 15...bxh3.

what does this have to do with the issue of thinking on the opponent's
time and whether it is better to think about one move or about two?
I suppose I'm overlooking something here...

: The programs (some, e.g. Hiarcs6 plays a6, but e.g. Fritz4 and other

: programs) play Bxh3.
: They think this could be a nice attack.

: In the 17th move they all consider normally with f4 as the
: permanent-brain move and different repetition-main-lines , sometimes
: with draw score.

: Why? After they throw the bishop into h3, they later find that white

I explained thinking on the opponent's time in absolute terms. There is
*no* room for wiggling out of it, so long as the programs continue to use
alpha/beta. Why do you suppose Rebel, Genius, CSTal, Crafty, Ferret, DB,
Hiarcs, Nimzo, Zarkov, Wchess, gnuchess, Rajah, and every other program I've
seen *all* do this the same way? Everyone's wrong I assume and your way
is *the* way? That's exactly what I thought...

: I hope - because this was my point to the discussion - that we can

[Robert Hyatt]

mclane (mcl...@prima.ruhr.de) wrote:
: hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

: >: You have not answered to my question WHY all your programs play e.g.
: >: Bxh3 and expect later a draw by repetition and plan f4 for Kosashvili
: >: because THEY think f4 is the best move, and than Yona plays better Re5
: >: instead, and suddenly the evaluation of all programs drop, although
: >: Re5 should be a weaker move.

: hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
: If you look at the title of this thread, and you read back through
: "most"
: of the postings, this has been about "what is the correct way to think
: on the opponents time?" and why... It doesn't have a thing to do
: with
: any specific game of chess that was played, that was some off-topic
: rambling you introduced. If you want to start a thread on this game,
: please post the pgn, specific moves you are interested in, and I'll be
: happy to join in with whatever I can offer. However the definition of
: "thread" is "postings that are about a common topic."..

: My answer:
: You are trying to spray fog arround. Also this is very rude. You don't
: have to explain me in a discussion what I am supposed to do, because
: this is off-topic.
: The game I speak about HAS TO DO WITH THE TOPIC we/you and others talk
: about. The problem is: you don't see this.

I certainly don't... thinking on the opponent's time is a mathematical
issue. It has nothing to do with games... It has to do with the mechanics
of using all available time effectively and efficiently... A single position
is just as important/useful...

: I don't want to start a thread about this game, because the game

: itself is irrelevant. It is just an example.
: I don't like your theoretical way of talking about topics the way many
: old people talk when they meat younger people or less experienced
: people: you talk the way:

: I have known about these facts when you were not born.

I didn't say "I have known..." I gave two specific journal references
to provide even more information as well as details about the math I
offered to show why searching only one move is much more effective. It
was wasted effort, it seems, which is typical...

: Although this is very nice, and the old guys seem to need some

Then the ball is in your court. You name *one* program that doesn't use
alpha/beta. Or one program that thinks about a bunch of moves when waiting
for the opponent to move... I gave you a list of those that don't... I
use rebel, genius, fritz, gnuchess and crafty, and know the others pretty
well from watching them play and talking with authors. Somewhere you meant
to turn left, but you went straight, and got badly lost...

: My program is in a position. I let it compute about move a for a

I make the following claim... 2+2=4. I claim it is always true. That's
just as safe as my claim about alpha/beta, but much simpler. Refute that,
then we go on to the bigger project of refuting alpha/beta. However,
alpha/beta is a precise mathematical algorithm. It's not subject to
whims and such... any more than 2+2 = something else...

: >: If your computers would have used the PB time, and instead of assuming

: >: the shit f4 would come, computed on whatever strange moves, would have
: >: computed anyhow different , maybe not only looking for the best moves,
: >: they would MAYBE have seen Re5. If they would have seen it, they could
: >: have better used PB-time instead of been shocked about the move, and
: >: then dropped dropped and lost the game.

Not a chance... because it would take so much longer... In fact, it would
take so much longer that the last search tree (before the move we just made
so we could start pondering) would have just as much information. I showed
why. Most understood...

: Hyatts words:

: nonsense... utter nonsense... and again you overlook alpha/beta. If
: I
: were to spread the search time over multiple moves, I would lose
: probably
: two plies of search depth in the same time limit. That's insane...

: My comments:
: The nonsense is coming from your side.
: If you relying on ONE move, this makes only sense, if this move is
: often a hit.
: That would be the case if chess would be a game, where many moves
: would be forced. But in chess we have very often situations where not
: only ONE move is the best, but a few moves are maybe equal.
: I guess ANY chess player can subscibe to this. Also you must have
: recognized this doing your job.

I don't disagree, but to give up two plies of search, means I won't
see any more when "pondering" than I saw *during* the search I just
finished when I made the last move. If I couldn't see it then, I won't
see it now, and it's all wasted effort...

:
: So I assume that we have normally a chance of 1:3 that the move YOU do

I've discovered that it is useless to talk to someone that won't listen,
or write for someone that won't read, nor illustrate for someone that won't
look. Get the idea???

: I wish you many years doing this method. Because this way one time

: Ananse will have overtaken you...

: Take my regards for this day in forward !

: >: I can only hope that there comes the day somebody writes an
: >: intelligent parallel program, and this - with the help of
: >: zillion-cpu's kicks all of you stupid-fast-.fanatics in the ass and
: >: send you to the back-side of the moon. There you can play with you
: >: deep-blue and whatever stupid/hardware-monsters until the last day of
: >: the universe.

: >That's what I'd expect from you, "I hope someone else..." Most of us are
: >trying to do it ourselves... but of course we at least have ideas to try
: >and understand what has been done in the past...

: The whole discussion shows me, that the amount of ideas in your mind
: seems to be limitated on things you have practised well over the
: years.
: As my old boss always have said: let us do it the usual way, this way
: worked all the years before.

You should listen to him. New ideas may work. Old ones generally
do...

:
: >: Maybe the time comes earlier than you would expect it. From my

: >: observations it does not take much time until CILK-Chess will show
: >: your MONSTER-machines that chess has not only to do with hardware ,
: >: and that arrogance will not help you to hide (=HYATT, maybe the best
: >: synonym for hiding Artificial Stupidy behind the foggy words) that
: >: these software running on fast hardware could only survive because
: >: there was no program near, that was more intelligent.

: >You really are ignorant. you need to ask Don Dailey about the engine.
: >I won't say any more. Ask Don. And then report back. And maybe, just
: >maybe, by then you'll have figured out just how stupid some of your
: >comments really are. *and* just how *wrong*...

: I have talked with him in Den Hague. He is not that ignorant you are
: !! :-)

The last conversation I had with him had them working "on searching as
fast as possible..." We'll see...

: >: I don't think that they would ever beat the champion. Maybe Kasparov.

: >: (For me Kasparov is no world-champion because he wanted to make his
: >: own deal in PCA. I don't want him, and I am happy that he left our
: >: organisation.) You can buy him. But you can't buy others.

: >: Maybe Cilk-chess could beat the champ. But Deep Blue ? Ha - they shall
: >: play g5 - because it is forced and you prefer it.

Of course, the small fact that most GM's say it is "the only move" doesn't
matter, because to you it is still a "shit move"... good think your opinion
doesn't carry much weight, or all games would be thrown out...

[Robert Hyatt]

mclane (mcl...@prima.ruhr.de) wrote:
: "chrisw" <chr...@cpsoft.demon.co.uk> wrote:

Warning!! bullshit warning.... change the last line to:

: You don't let everybody tell you anything....
^^^^^^^^^

and it makes more sense. Many here have perfectly good ideas. Only
a selected few don't... Probably *everybody* has good ideas at times.
But most of us figure out when we support something that is untenable
and dry it up...

[Robert Hyatt]

mclane (mcl...@prima.ruhr.de) wrote:
: hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

: >: I don't have to write a chess program to find out that somebody else
: >: is an idiot.

: >I don't believe you are smart enough to find out someone is an idiot.
: >You have to be smarter than that person. I can't think of even one you
: >could make that claim about...

: Rubbish !

: You don't have to be raped to find out that raping is not ok, you

: don't have to be smart to find out that somebody is an idiot.

: Even:
: You don't have to be a programmer to understand about programs.
: You don't have to be a writer to criticize about books.
: You don't have to be a song-writer to sing better than Julio Iglesias
: You don't have to build deep blue to find out the machine is garbage
: You don't have to a chess player to find out that this match is
: arranged.

Add...

you don't have to be a complete idiot to post complete nonsense....

: Your "society" want to stop people saying something that is against

[Robert Hyatt]

mclane (mcl...@prima.ruhr.de) wrote:
: hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

: > And there's not any way to get around this when using
: >alpha/beta..

: >There you have it. The math to explain why current pondering is the way to
: >go based on tree sizes, and some empirical data to support the conclusion as
: >well. Note that the math I gave ignores null-move, but that factoring it in
: >only makes the first branch even *larger* with respect to the remaining branches
: >as the empirical data shows...

: >Hope that helps...

: >Bob

: Brilliant Bob - you declare it, and we have to believe it. Your main

: point is, and that makes the whole thing worthless:
: WHEN USING ALPHA BETA.

Then you trot out a program that doesn't use alpha/beta and we'll talk.
Hint: talk to Chris. He *does* use it... So I suppose I have somehow
(again) overlooked some subtle point to your post, but danged if I can
figure out what it is. Why post a CSTal game when it is subject to the
exact limits I posted. Do you claim it ponders (or permanent brain's)
"differently"???

You keep saying that "when using alpha/beta" is a qualification that
makes the discussions I've given "moot". I claim that your point of view
eliminates *every* chess program currently playing the game, because they
*all* use alpha/beta. Now how do you intend to rationalize that away. I
don't have Ed's source. I don't need it. If you don't think he uses
alpha/beta, ask 'im... This is very much like arguing with a 12-year-old
at times.

I have tried two ways to explain to you why your idea won't work. (Note:
I have not said your idea is a bullshit idea, because as a human I also
happen to consider more than one move for my opponent. I only claim that
your idea won't work in the framework of alpha/beta, which we are all using
at present to develop chess programs...)
I have tried a simple "lecture format" of discussion, which didn't work. I
then produced a precise mathematical formulation of alpha/beta, and showed
how that proves that pondering more than one move is a bad idea. That didn't
work. So my final attempt at this is to show by actually implementing your
idea in Crafty, and then producing the output. This also shows why your idea
won't work. And I'm probably wasting my time for the third time, but I'll
at least try this last method of explaining. Note that this has taken about
three hours of time to put this together. I don't invest that much time
lightly, I'm doing my best to educate you on why your statements are wrong
in the case of "pondering." If you'd spend the next three hours looking over
my previous posts, plus the data in this one, perhaps you'll "see the light".
Of course, to do so, you *must* open your eyes...

Here's the game score if you don't have it handy:

[Event "Deep Blue vs Kasparov match #2"]
[Round "2"]
[Date "1997.05.04"]
[White "Deep Blue"]
[Black "Kasparov"]
[Result "1-0"]
[ECO "C93"]
[NIC "RL.22"]
[LongECO "Ruy Lopez: closed, Smyslov defense"]

1. e4 e5 2. Nf3 Nc6 3. Bb5 a6 4. Ba4 Nf6 5. O-O Be7
6. Re1 b5 7. Bb3 d6 8. c3 O-O 9. h3 h6 10. d4 Re8
11. Nbd2 Bf8 12. Nf1 Bd7 13. Ng3 Na5 14. Bc2 c5 15. b3 Nc6
16. d5 Ne7 17. Be3 Ng6 18. Qd2 Nh7 19. a4 Nh4 20. Nxh4 Qxh4
21. Qe2 Qd8 22. b4 Qc7 23. Rec1 c4 24. Ra3 Rec8 25. Rca1 Qd8
26. f4 Nf6 27. fxe5 dxe5 28. Qf1 Ne8 29. Qf2 Nd6 30. Bb6 Qe8
31. R3a2 Be7 32. Bc5 Bf8 33. Nf5 Bxf5 34. exf5 f6 35. Bxd6 Bxd6
36. axb5 axb5 37. Be4 Rxa2 38. Qxa2 Qd7 39. Qa7 Rc7 40. Qb6 Rb7
41. Ra8+ Kf7 42. Qa6 Qc7 43. Qc6 Qb6+ 44. Kf1 Rb8 45. Ra6
{Black resigns} 1-0

Here's the test. DB vs Kasparov game 2. I randomly picked a position, with
Crafty to move, in my case, playing DB's side at move 28. I allowed it to search
for the normal time limit and make a move. It would expect black to play a5, and
the normal crafty would ponder that move. Let's assume that Kasparov is going to
take exactly three minutes to make his move. Here is the ponder search from Crafty
looking at *only* the expected 28. ... a5 response by Kasparov:

depth time score variation (1)
6 0.38 -2.210 a5 axb5 Rcb8 Rxa5 Rxa5 Rxa5 Bxb5
6-> 0.39 -2.210 a5 axb5 Rcb8 Rxa5 Rxa5 Rxa5 Bxb5
7 0.82 ++ a5!!
7 1.23 -0.758 a5 axb5 axb4 Rxa8 Rxa8 Rxa8 Qxa8 Qxc4
Qa1+ Kf2 bxc3
7-> 1.33 -0.758 a5 axb5 axb4 Rxa8 Rxa8 Rxa8 Qxa8 Qxc4
Qa1+ Kf2 bxc3
8 3.13 -0.758 a5 axb5 axb4 Rxa8 Rxa8 Rxa8 Qxa8 Qxc4
Qa1+ Kf2 bxc3
8-> 3.23 -0.758 a5 axb5 axb4 Rxa8 Rxa8 Rxa8 Qxa8 Qxc4
Qa1+ Kf2 bxc3
9 6.70 -0.898 a5 axb5 axb4 Rxa8 Rxa8 Rxa8 Qxa8 Qxc4
Qa1+ Kf2 bxc3 b6
9-> 6.79 -0.898 a5 axb5 axb4 Rxa8 Rxa8 Rxa8 Qxa8 Qxc4
Qa1+ Kf2 bxc3 b6
10 14.20 -1.090 a5 axb5 axb4 Rxa8 Rxa8 Rxa8 Qxa8 Qxc4
Qa1+ Kf2 Qxc3 Qxc3 bxc3 b6
10-> 14.34 -1.090 a5 axb5 axb4 Rxa8 Rxa8 Rxa8 Qxa8 Qxc4
Qa1+ Kf2 Qxc3 Qxc3 bxc3 b6
11 27.96 ++ a5!!
11 42.25 -0.736 a5 axb5 axb4 Rxa8 Rxa8 Rxa8 Qxa8 Qxc4
Qa1+ Kf2 Qxc3 Qxc3 bxc3 b6 Bd6
11-> 42.35 -0.736 a5 axb5 axb4 Rxa8 Rxa8 Rxa8 Qxa8 Qxc4
Qa1+ Kf2 Qxc3 Qxc3 bxc3 b6 Bd6
12 1:37 -1.021 a5 axb5 axb4 Rxa8 Rxa8 Rxa8 Qxa8 Qxc4
Qa1+ Kf2 Qxc3 Qxc3 bxc3 b6 Bd6 b7 Kf8
12-> 1:37 -1.021 a5 axb5 axb4 Rxa8 Rxa8 Rxa8 Qxa8 Qxc4
Qa1+ Kf2 Qxc3 Qxc3 bxc3 b6 Bd6 b7 Kf8
time: 3:00 cpu:98% mat:0 n:13502931 nps:75974

This is what a ponder search would look like. Note that crafty *only* searched
a5 for black, and then goes about finding the best move in response, which it
believes (in this case) is Kf2. Note that also, the way this output is given,
that the real ponder search is one ply less than the dept shown, because the
first move is *always* a5 here. But I am going to use the depth value from
above because it is consistent with the depth for the next test. So, we hit
depth=12 in 3 minutes (tiny hash table, nothing tricky here, so this should
be reproducible by any crafty user by simply saying "search a5" and "go"...

Now, let's try your idea of searching more than one move. The first problem is
to figure out which is the best set of moves to search. Alpha/beta won't give us
that without a lot of effort. So, by magic, using crafty's N-variation annotate
command I had it figure out which were the best 10 moves, and I'm using those 10
moves for the next test. So, here we go. Crafty is going to search each of 10
different moves for Kasparov, rather than only searching the move that it thought
was best. Here's the 10 separate searches:

Note... since we have exactly 180 seconds to ponder, before Kasparov
moves (in this hypothetical scenario) and since we are going to ponder the best
"10" moves which have been handed to us by magic (it would take even longer to
figure out which are the best 10 moves...) I'm going to let crafty analyze
(ponder) for each of these 10 moves, but am going to cut things off at 18
seconds, since 10 moves x 18 seconds each = 180 seconds. You can go back
and take the best 5 moves and we could add more analysis if you like, 10 is
arbitrary. It's much larger than 1, but much less than the total legal moves,
many of which simply lose material instantly and are hopeless.

In the data below, first let's assume Kasparov actually plays a5, which he
didn't. If we had pondered a5 alone, for 180 seconds, we would have a 12 ply
search done, and we could easily move instantly. but in the data below, after
kasparov plays a5, we have only considered that move for 18 seconds. I would
not play instantly, but would say "hmmm... I want to search for 180 seconds,
and I've already spent 18 seconds on the right move, so I just keep going for
an additional 162 seconds, and I save 18 seconds on my clock, rather than 180
seconds. Not so good. But let's suppose that Kasparov plays Ne8 instead, Crafty's
second-best choice (which he actually played, BTW). After Kasparov moves, we have
done a deep 8 ply search using 18 seconds. Do we search deeper or move instantly?
We search deeper, because 18 seconds is not enough time to find all the tactical
things that are going on. So we again have to search for 162 seconds. In fact,
if we target 180 seconds for *every* move in the game, we will actually have to
search for 162 seconds for every move (assuming Kasparov also takes exactly 180
seconds of course). So we save 18 seconds per move, *every* move, or for 40
moves we save 40*18 = 720 seconds, or a grand total of 12 minutes out of the
first two hours. In this game, for the first 40 moves, Crafty correctly predicted
28 of kasparovs 40 moves. if we cancel the first 8 predictions because they were
in and out of book, then we get about 20 out of 30 moves predicted correctly.

Searching/pondering the way *all of us* are doing, for 20 of Kasparov's moves
we would have spent the full 180 seconds on the right move, and could move
instantly if we wanted, saving exactly 60 minutes on the clock (compared to 12
in the best-10 search). Which is more reasonable? I believe the answer is
obvious. Here's the data, just so you can see what the numbers look like. If
this doesn't explain the "why" of pondering only one move, then perhaps it is
hopeless to continue trying...

depth time score variation (1)
6 0.47 -2.210 a5 axb5 Rcb8 Rxa5 Rxa5 Rxa5 Bxb5
6-> 0.52 -2.210 a5 axb5 Rcb8 Rxa5 Rxa5 Rxa5 Bxb5
7 0.86 ++ a5!!
7 1.27 -0.758 a5 axb5 axb4 Rxa8 Rxa8 Rxa8 Qxa8 Qxc4
Qa1+ Kf2 bxc3
7-> 1.37 -0.758 a5 axb5 axb4 Rxa8 Rxa8 Rxa8 Qxa8 Qxc4
Qa1+ Kf2 bxc3
8 3.16 -0.758 a5 axb5 axb4 Rxa8 Rxa8 Rxa8 Qxa8 Qxc4
Qa1+ Kf2 bxc3
8-> 3.26 -0.758 a5 axb5 axb4 Rxa8 Rxa8 Rxa8 Qxa8 Qxc4
Qa1+ Kf2 bxc3
9 6.73 -0.898 a5 axb5 axb4 Rxa8 Rxa8 Rxa8 Qxa8 Qxc4
Qa1+ Kf2 bxc3 b6
9-> 6.83 -0.898 a5 axb5 axb4 Rxa8 Rxa8 Rxa8 Qxa8 Qxc4
Qa1+ Kf2 bxc3 b6
10 14.22 -1.090 a5 axb5 axb4 Rxa8 Rxa8 Rxa8 Qxa8 Qxc4
Qa1+ Kf2 Qxc3 Qxc3 bxc3 b6
10-> 14.32 -1.090 a5 axb5 axb4 Rxa8 Rxa8 Rxa8 Qxa8 Qxc4
Qa1+ Kf2 Qxc3 Qxc3 bxc3 b6

depth time score variation (1)
5 0.31 -1.430 Ne8 axb5 Bxb5 Qf5 Bd6 R3a2
5-> 0.37 -1.430 Ne8 axb5 Bxb5 Qf5 Bd6 R3a2
6 1.05 -1.294 Ne8 axb5 Bxb5 Nf5 Nd6 Nxd6 Bxd6 Qf5
6-> 1.10 -1.294 Ne8 axb5 Bxb5 Nf5 Nd6 Nxd6 Bxd6 Qf5
7 2.28 -1.292 Ne8 axb5 Bxb5 Bd1 Nf6 Qf5 Bd6 Bf3
7-> 2.32 -1.292 Ne8 axb5 Bxb5 Bd1 Nf6 Qf5 Bd6 Bf3
8 12.28 -1.130 Ne8 axb5 axb5 Ra6 Rxa6 Rxa6 Bd6 Nf5
Bc7 Ng3
8-> 12.46 -1.130 Ne8 axb5 axb5 Ra6 Rxa6 Rxa6 Bd6 Nf5
Bc7 Ng3

depth time score variation (1)
6 1.10 -1.180 Rcb8 a5 Bd6 Nf5 Bc7 Ng3
6-> 1.14 -1.180 Rcb8 a5 Bd6 Nf5 Bc7 Ng3
7 2.69 -1.304 Rcb8 Qf3 Bd6 axb5 Bxb5 Nf5 Rc8 Ra5
<HT>
7-> 2.78 -1.304 Rcb8 Qf3 Bd6 axb5 Bxb5 Nf5 Rc8 Ra5
<HT>
8 9.90 -1.180 Rcb8 a5 Bd6 Qf3 Ne8 Nf5 f6 <HT>
8-> 9.98 -1.180 Rcb8 a5 Bd6 Qf3 Ne8 Nf5 f6 <HT>

depth time score variation (1)
6 0.42 -1.543 Nh7 axb5 Bxb5 Nf5 Bd6 Nxd6 Qxd6 Qf5
6-> 0.47 -1.543 Nh7 axb5 Bxb5 Nf5 Bd6 Nxd6 Qxd6 Qf5
7 3.23 -1.333 Nh7 Nf5 f6 Qf2 Rab8 axb5 axb5 Ba7
7-> 3.28 -1.333 Nh7 Nf5 f6 Qf2 Rab8 axb5 axb5 Ba7
8 14.48 -1.257 Nh7 axb5 axb5 Ra6 Rxa6 Rxa6 f6 Bb6
Qe7 Ra7
8-> 14.54 -1.257 Nh7 axb5 axb5 Ra6 Rxa6 Rxa6 f6 Bb6
Qe7 Ra7

depth time score variation (1)
6 0.75 -1.302 Qe8 a5 Bd6 Nf5 Bxf5 Qxf5 Qe7
6-> 0.77 -1.302 Qe8 a5 Bd6 Nf5 Bxf5 Qxf5 Qe7
7 3.83 -1.399 Qe8 Qf3 Qd8 Nf5 Bxf5 <HT>
7-> 3.86 -1.399 Qe8 Qf3 Qd8 Nf5 Bxf5 <HT>
8 12.12 -1.271 Qe8 a5 Qd8 Bb6 Qe8 Nf5 Bxf5 Qxf5 Bd6
8-> 12.18 -1.271 Qe8 a5 Qd8 Bb6 Qe8 Nf5 Bxf5 Qxf5 Bd6

depth time score variation (1)
6 0.45 -1.100 Bd6 a5 Ne8 Rd1 f6 Qf3
6-> 0.47 -1.100 Bd6 a5 Ne8 Rd1 f6 Qf3
7 2.55 -1.232 Bd6 Nf5 Ne8 axb5 axb5 Nxd6 Nxd6 Rxa8
Rxa8 Rxa8 Qxa8
7-> 2.60 -1.232 Bd6 Nf5 Ne8 axb5 axb5 Nxd6 Nxd6 Rxa8
Rxa8 Rxa8 Qxa8
8 4.16 -1.258 Bd6 Nf5 Ne8 axb5 Bxb5 Qf3 f6 Qg4 Qc7
8-> 4.18 -1.258 Bd6 Nf5 Ne8 axb5 Bxb5 Qf3 f6 Qg4 Qc7
9 12.07 -1.232 Bd6 Nf5 Ne8 axb5 axb5 Nxd6 Nxd6 Ra6
Qf8 Rxa8 Rxa8 Rxa8 Qxa8
9-> 12.15 -1.232 Bd6 Nf5 Ne8 axb5 axb5 Nxd6 Nxd6 Ra6
Qf8 Rxa8 Rxa8 Rxa8 Qxa8

depth time score variation (1)
6 0.58 -1.160 Be7 a5 Bd6 Nf5 Bc7 Ng3
6-> 0.60 -1.160 Be7 a5 Bd6 Nf5 Bc7 Ng3
7 2.09 -1.284 Be7 Qf3 Bd6 axb5 Bxb5 Nf5 Rab8 Ra5
7-> 2.11 -1.284 Be7 Qf3 Bd6 axb5 Bxb5 Nf5 Rab8 Ra5
8 6.60 -1.180 Be7 a5 Ne8 Bb6 Nc7 Nf5 f6 Nxe7+ Qxe7
Rd1
8-> 6.64 -1.180 Be7 a5 Ne8 Bb6 Nc7 Nf5 f6 Nxe7+ Qxe7
Rd1

depth time score variation (1)
6 0.52 -1.211 Rab8 axb5 axb5 Ra7 Bd6 Qf3 Ra8
6-> 0.54 -1.211 Rab8 axb5 axb5 Ra7 Bd6 Qf3 Ra8
7 1.71 -1.294 Rab8 axb5 axb5 Nf5 Rc7 Qf3 Bxf5 Qxf5
Bd6
7-> 1.75 -1.294 Rab8 axb5 axb5 Nf5 Rc7 Qf3 Bxf5 Qxf5
Bd6
8 5.83 -1.243 Rab8 a5 Bd6 Nf5 Ne8 Ba7 Ra8 Bb6
8-> 5.86 -1.243 Rab8 a5 Bd6 Nf5 Ne8 Ba7 Ra8 Bb6

depth time score variation (1)
6 0.58 -1.270 Kh7 a5 Bd6 Nf5 Bc7 Ng3
6-> 0.60 -1.270 Kh7 a5 Bd6 Nf5 Bc7 Ng3
7 1.64 -1.304 Kh7 axb5 Bxb5 Qf3 Bd6 Nf5 Kg8 Ra5
7-> 1.66 -1.304 Kh7 axb5 Bxb5 Qf3 Bd6 Nf5 Kg8 Ra5
8 8.73 -1.271 Kh7 a5 Kg8 Bb6 Qe8 Nf5 Bxf5 Qxf5 Bd6
8-> 8.76 -1.271 Kh7 a5 Kg8 Bb6 Qe8 Nf5 Bxf5 Qxf5 Bd6

depth time score variation (1)
6 0.39 -1.373 Be8 axb5 Bxb5 Nf5 Bd6 Nxd6 Qxd6 Qf5
6-> 0.41 -1.373 Be8 axb5 Bxb5 Nf5 Bd6 Nxd6 Qxd6 Qf5
7 1.29 -1.380 Be8 axb5 Bxb5 Nf5 Bd6 Nxd6 Qxd6 Bc5
Qc7
7-> 1.33 -1.380 Be8 axb5 Bxb5 Nf5 Bd6 Nxd6 Qxd6 Bc5
Qc7
8 3.75 -1.423 Be8 axb5 Bxb5 Ba4 Rab8 Bxb5 axb5 Qf5
Bd6 Ra7
8-> 3.79 -1.423 Be8 axb5 Bxb5 Ba4 Rab8 Bxb5 axb5 Qf5
Bd6 Ra7
9 15.48 -1.325 Be8 axb5 Bxb5 Qf2 Ne8 Bb6 Nc7 Qf5
Bd6 Be3
9-> 15.57 -1.325 Be8 axb5 Bxb5 Qf2 Ne8 Bb6 Nc7 Qf5
Bd6 Be3

The problem with your rhetoric is this: The old guys, the young guys,
the middle-aged guys, the smart guys, the not-so-smart guys, and all the
rest are *still* using alpha/beta my friend... You rant as though there
is an alternative that is being used by someone in the "commercial" ranks.
Produce it...

Here's crafty's analysis:

depth time score variation (1)
4-> 0.15 -0.046 Rd8 Be3 Qe4 Qd3
5 0.24 -0.001 Rd8 Be3 Bd5 Qg4 Qxg4 hxg4
5 0.43 0.078 a6 Bf1 Bd5 g3 Qf6
5-> 0.58 0.078 a6 Bf1 Bd5 g3 Qf6
6 0.99 -0.086 a6 Bf1 Bd5 g3 Qf6 Be3
6 1.99 -0.065 Rd8 Qd3 Bd5 Be3 Kg8 Qf5 Qe4 <HT>
6 2.68 -0.023 Bb4 Re3 Rd8 Bb2 Bc5 Rd3
6-> 2.89 -0.023 Bb4 Re3 Rd8 Bb2 Bc5 Rd3
7 4.04 -0.022 Bb4 Re2 Rd8 Bd2 Bxd2 Rxd2 Rd5 Bc4
7 6.03 0.056 a6 Bf1 Bd5 g3 Qf6 Bg2 Bxg2 Kxg2
7-> 7.54 0.056 a6 Bf1 Bd5 g3 Qf6 Bg2 Bxg2 Kxg2
8 10.99 -0.014 a6 Bf1 Rd8 Be3 Bc5 g3 Qf6 Qb1 Bxd4
Bxd4 Qxd4 Qxb7
8 22.56 0.000 Bxh3 gxh3 Qxh3 f4 Qg3+ Kh1 Qh3+ Kg1
8-> 23.30 0.000 Bxh3 gxh3 Qxh3 f4 Qg3+ Kh1 Qh3+ Kg1
9 32.69 0.000 Bxh3 gxh3 Qxh3 f4 Qg3+ Kh1 Qh3+ Kg1
9-> 1:39 0.000 Bxh3 gxh3 Qxh3 f4 Qg3+ Kh1 Qh3+ Kg1
10 2:12 0.000 Bxh3 gxh3 Qxh3 f4 Qg3+ Kh1 Qh3+ Kg1
10-> 3:20 0.000 Bxh3 gxh3 Qxh3 f4 Qg3+ Kh1 Qh3+ Kg1
11 4:49 0.000 Bxh3 gxh3 Qxh3 f4 Qg3+ Kh1 Qh3+ Kg1

so it agrees... repetition...

: Helpfully Yona Kosashvilil did not what the stupid-programs expected.
: He did a very normal and easy move, many chess-players would have
: done: gave back the Rook with 17.Re5.

depth time score variation (1)
9 9.11 0.340 Bxe5 dxe5 Qe6 Qd4 Kg8 Bg5 Rc8 Bd3
h6
9-> 17.51 0.340 Bxe5 dxe5 Qe6 Qd4 Kg8 Bg5 Rc8 Bd3
h6
10 36.08 0.325 Bxe5 dxe5 Qe6 Qd4 a6 Bd3 Re8 Ba3+
Kg8 Bd6 h6
10-> 55.65 0.325 Bxe5 dxe5 Qe6 Qd4 a6 Bd3 Re8 Ba3+
Kg8 Bd6 h6
11 1:47 0.269 Bxe5 dxe5 Qe6 Qd4 Kg8 Be3 Rc8 Rd1
Qg6+ Kf1 Qc2 f4
11-> 4:45 0.269 Bxe5 dxe5 Qe6 Qd4 Kg8 Be3 Rc8 Rd1
Qg6+ Kf1 Qc2 f4
12 9:04 0.230 Bxe5 dxe5 Qe6 Bf4 Kg8 Qf3 Qg6+ Kh2
Qf5 Bd3 Qd7 Be4 Rb8

So it still agrees, white's move was a mistake. I assume you have a point
here? That for some reason, the program should be searching inferior moves
for the opponent?

: This is "brilliant" because we have a good example why machines are
: stupid. And also it stops possible treatment by blacks troops, as
: chris would express himself.

The machines are stupid if they misevaluate the game. But it has *nothing*
to do with which move they happen to search. In this game. pondering on Re5
rather than f4 wouldn't have affected the game *at all* had crafty been playing
black. Because it would not have searched any deeper than 12 plies, and it
wouldn't have changed it's mind. If it had been expecting f4, and white played
Re5, it would have started over, and only reached a depth of 11 plies in 4:45 or
less, which is exactly the depth it would have reached had it predicted Re5 in
that same time limit. So no effect on the game's outcome at all. If CSTal played
stupidly in your mind, that's a different issue. But it *wasn't* a result of not
thinking on the opponent's time in your so-called "more intelligent" approach.

: The machines wasted permanent-brain computing on f4 main-lines, mainly
: they found out that the position will end in a draw.

And what's your point. The programs will continue to think that f4 is
better, and that if white plays this move, the program is even better
off than before...

: Suddenly Yona plays a WEAKER move than the BEST move f4, and and ?
: The evaluation drops after this move has been made, and they would
: maybe all lose the position because black has ONLY A QUEEN against an
: army of white pieces threatening the black king.
: But a queen alone, cannot mate without the help of another piece ! But
: which piece ?

: SO: this is my chess-example to the discussion about strategies and
: permanent brain. And I am pleased to see your results.

: Why did the programs lost although the opponent made the weaker move.
: Because after your Rebel-Nimzo match you are an EXPERT for explaining
: why your craty lost although Rebel made WEAKER moves, I am pleased to
: hear from you further.

It lost because *it* played inferior moves too, or else *it* misevaluated
the position and thought it went to a better position when, in fact, it
went to a worse one. This sounds like ranting from someone that is really
mentally unbalanced.

: Istead of telling the people that one can better discuss with a tree ,
: you should discuss (what would be ON TOPIC!) this above CHESS-TREE
: stuff. I hope you can satisfy me , and the others.

I don't hear *anyone* else ranting on and on about this. I've seen no
one take issue with the math I posted, nor is it likely with the empirical
data I've posted above. You do notice, I hope, that I generally try to
answer such discussions with real data when possible? You ought to try
that sometime, sometimes you might answer your own question rather than
keeping a discussion alive that should have died days ago...

: I hope - because this was my point to the discussion - that we can
: learn something about the mistakes of others (here Chess System Tal's
: mistake of throwing Bxh3).

might not be a mistake. But that's not the point. That had nothing to
do with his pondering only one move for the opponent. It would have played
that if you had pondering turned off. As did crafty in the analysis I posted
above.

: When I am right, Yona played h3 by purpose, to let us run into this
: FALSE sac. He could have played g3, but in the moment he played h3, I
: thought: he did it by purpose. Maybe Yona has seen Re5 while playing
: h3.

I really strongly doubt it. You should have asked, rather than resorting
to your usual speculation. Facts are so much nicer to work with. Speculation
is like swatting at a swarm of bees... it only makes things worse...

[Komputer Korner]

Robert Hyatt wrote:

>
> only way to achieve 100% is to ponder *every* possible opponent move, then
> you can't move instantly... because you've only searched a reply to his move
> for < 10% of the total time you want to search before making a move... Why
> is this topic still going on???

It is still going on because Tom Kerrigan still doesn't understand
that 1 K variation pondering is the optimum as long as the guess rate
is >= to 50%.

[mclane]

hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

> Why post a CSTal game when it is subject to the
>exact limits I posted. Do you claim it ponders (or permanent brain's)
>"differently"???

I hope you do not think I want to advertise CSTal. The point is, why I
post this game in thi context: I know about this game because I
operated the machine, and I knew it PERMANENT BRAINED on f4, but f4
did not come. Also I told you, there was/is a dedicated chess computer
that computes on more than ONE move in PB. It is from saitekl and it
is years ago...

>what does this have to do with the issue of thinking on the opponent's
>time and whether it is better to think about one move or about two?
>I suppose I'm overlooking something here...

I know that you are overlooking something. Thats your main way of
building an opinion, by NOT looking on facts, or opinions other people
have. Your brain-algorithm works very selective, I wonder if you use
alpha-beta yourself.

>I explained thinking on the opponent's time in absolute terms. There is
>*no* room for wiggling out of it, so long as the programs continue to use
>alpha/beta. Why do you suppose Rebel, Genius, CSTal, Crafty, Ferret, DB,
>Hiarcs, Nimzo, Zarkov, Wchess, gnuchess, Rajah, and every other program I've
>seen *all* do this the same way? Everyone's wrong I assume and your way
>is *the* way? That's exactly what I thought...

Of course you commanded them to do it this way, did you forgot? Moses
and his ten commandments? :-)
If you would concentrate on the game, instead of again IGNORING it,
you would see that Rebel/Genius/CST/... make a mistake in ONLY PB f4.
But is is not your day, you are not capable of relying to chess-facts.
Instead you post wishi-washi answers a 10 year old boy could post when
he is in bad mood.
Concentrate on the game. Or let it-.

[mclane]

hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

>Add...

>you don't have to be a complete idiot to post complete nonsense....

No - sometimes it is enough to be Bob Hyatt, of course. Enough to not
asnwer on chess-data instead of answer on these off-topic
Korinthenkacker discussion here.

Take the game. Replay it. If you want you can delete the CSTAl in the
header, or if your idle view needs it, replace it with crafty...

[mclane]

hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

>: The game I speak about HAS TO DO WITH THE TOPIC we/you and others talk
>: about. The problem is: you don't see this.

>I certainly don't... thinking on the opponent's time is a mathematical
>issue. It has nothing to do with games... It has to do with the mechanics
>of using all available time effectively and efficiently... A single position
>is just as important/useful...

Than your mathematics sucks. Any 10 year old chess player could tell
you, that thinking on 1 move is inefficient.
If you would behave like your chess-program in a decision process in
life, you would make as much mistakes as your program does.
Ah - maybe that is the reason you behave this faulty in discussion.
YOU ARE USING ONLY ONE PB-move !!

Now I see it.... sorry.

>I didn't say "I have known..." I gave two specific journal references
>to provide even more information as well as details about the math I
>offered to show why searching only one move is much more effective. It
>was wasted effort, it seems, which is typical...

If you search on one move, and it is not coming, what happens often,
it must happen often, because it is systemimmanent of chess, you wated
computing time. Chess has no OFTEN used best moves... chess has often
many good moves, you cannot exactly say which one is better.
I guess many chess-programs don't try this, because it would be to
complicate to implement.

>: >: If your computers would have used the PB time, and instead of assuming
>: >: the shit f4 would come, computed on whatever strange moves, would have
>: >: computed anyhow different , maybe not only looking for the best moves,
>: >: they would MAYBE have seen Re5. If they would have seen it, they could
>: >: have better used PB-time instead of been shocked about the move, and
>: >: then dropped dropped and lost the game.

>Not a chance... because it would take so much longer... In fact, it would
>take so much longer that the last search tree (before the move we just made
>so we could start pondering) would have just as much information. I showed
>why. Most understood...

Use the game, show me why the programs expect f4 and I show you that a
next-best search about the position takes 2 seconds for the programs
to find out: Re5 is better than f4.
So if they waste 40 seconds on thinking: f4 is the move to come, what
can I play against f4, they would have seen in 2 seconds next best,
that re5 is the move to be prepared and could pb on Re5 instead.

But - as you said before.... it would take so much longer.
And when GOD says, than it is true. Sorry. I always forgot that I am
not allowed to even think different than you allow or extrapolate it.

Try it out in the position. I have tried it out before the whole
discussion was involved, and my machines say: I am right.
Whatever you blow in the sky here, it is just brabbling of words
instead of using the game I posted and give concrete data on it.

Thanks your majesty.

[Robert Hyatt]

mclane (mcl...@prima.ruhr.de) wrote:
: hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

: > Why post a CSTal game when it is subject to the

: >exact limits I posted. Do you claim it ponders (or permanent brain's)
: >"differently"???

: I hope you do not think I want to advertise CSTal. The point is, why I

: post this game in thi context: I know about this game because I
: operated the machine, and I knew it PERMANENT BRAINED on f4, but f4
: did not come. Also I told you, there was/is a dedicated chess computer
: that computes on more than ONE move in PB. It is from saitekl and it
: is years ago...

I'd bet you are incorrect, because it is an illogical way of thinking on
the opponent's time. Dredge up the model, I'll find out who wrote the
program and we can ask them directly.

: >what does this have to do with the issue of thinking on the opponent's

: >time and whether it is better to think about one move or about two?
: >I suppose I'm overlooking something here...

: I know that you are overlooking something. Thats your main way of

: building an opinion, by NOT looking on facts, or opinions other people
: have. Your brain-algorithm works very selective, I wonder if you use
: alpha-beta yourself.

: >I explained thinking on the opponent's time in absolute terms. There is

: >*no* room for wiggling out of it, so long as the programs continue to use
: >alpha/beta. Why do you suppose Rebel, Genius, CSTal, Crafty, Ferret, DB,
: >Hiarcs, Nimzo, Zarkov, Wchess, gnuchess, Rajah, and every other program I've
: >seen *all* do this the same way? Everyone's wrong I assume and your way
: >is *the* way? That's exactly what I thought...

: Of course you commanded them to do it this way, did you forgot? Moses

: and his ten commandments? :-)
: If you would concentrate on the game, instead of again IGNORING it,
: you would see that Rebel/Genius/CST/... make a mistake in ONLY PB f4.
: But is is not your day, you are not capable of relying to chess-facts.
: Instead you post wishi-washi answers a 10 year old boy could post when
: he is in bad mood.

They make a mistake *on that move* only. What about on the moves where they
do predict correctly and save the time? I posted analysis with real data from
Crafty to explain why doing it your way won't work very well. Nothing wrong
with your *idea* as I mentioned. But within the framework of current computer
chess using alpha/beta it is *totally* unworkable... *lots* of ideas are
unworkable... I've pointed out the "discernability" problem of a program
choosing to lose obviously rather than choosing to lose subtly, because in that
case obvious was a few "millipawns" better. Idea is good. But alpha/beta doesn't
offer up a solution...

: Concentrate on the game. Or let it-.

[Robert Hyatt]

mclane (mcl...@prima.ruhr.de) wrote:
: hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

: >and it makes more sense. Many here have perfectly good ideas. Only

: >a selected few don't... Probably *everybody* has good ideas at times.
: >But most of us figure out when we support something that is untenable
: >and dry it up...

: You hjave not answered to my example game. Instead you trying to
: convince people different paths. This is off-topic.
: In the example game the permanent-brain move was NOT played, and the
: reason is: why did the programs not see that f4 is weaker than Re5.
: I guess Alpha-beta problem.
: Therefore I would say: if this algorithm is all you have invented over
: the years, than you will never win against grandmasters, because you
: oversee important moves.

I have posted analysis for your example game. Get off your butt, find it
and study it. Then come back with something that makes sense. Anybody else
reading here that did not see my computer analysis of his game? Or is he
the only one???

[mclane]

[Peter Osterlund]

On Thu, 8 May 1997, mclane wrote:

: If you would concentrate on the game, instead of again IGNORING it,
: you would see that Rebel/Genius/CST/... make a mistake in ONLY PB f4.

: Concentrate on the game. Or let it-.

Your claim is that in this game at the position you describe, it would
have been better to ponder both Re5 and f4. This is correct because the
move actually played was Re5, so pondering that move would have saved some
time on the computers clock.

Hyatt claims that in order to save as much time as possible (on the
average), you should only ponder the one move your opponent is most likely
to play. This is also correct.

You then try to conclude that Hyatt's claim is wrong because your own
claim is correct, and this is where your logic fails. Do you suggest that
we should use a pondering strategy that on average is much worse, just
because in certain cases it would be slightly better?

--
Peter Österlund Email: peter.o...@mailbox.swipnet.se
Sköndalsvägen 35 f90...@nada.kth.se
S-128 66 Sköndal Homepage: http://home1.swipnet.se/~w-15919
Sweden Phone: +46 8 942647

[Torstein Hall]

mclane <mcl...@prima.ruhr.de> skrev i artikkelen
<5krr67$js0$4...@steve.prima.ruhr.de>...

> hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
>
>
> >: I don't have to write a chess program to find out that somebody else
> >: is an idiot.
>
> >I don't believe you are smart enough to find out someone is an idiot.
> >You have to be smarter than that person. I can't think of even one you
> >could make that claim about...
>
> Rubbish !
>

> You don't have to be raped to find out that raping is not ok, you

> don't have to be smart to find out that somebody is an idiot.
>
> Even:
> You don't have to be a programmer to understand about programs.
> You don't have to be a writer to criticize about books.
> You don't have to be a song-writer to sing better than Julio Iglesias
> You don't have to build deep blue to find out the machine is garbage
> You don't have to a chess player to find out that this match is
> arranged.

Perhaps you are leaving a line out here?

You don't have to be a psycologist to find out that mclane is
.............

--
Torstein Hall
tors...@eunet.no
http://login.eunet.no/torshall/

[Komputer Korner]

Robert Hyatt wrote:
snipped

Actually the math is quite simple as Bob has correctly pointed out.
It depends on the guess rate. Here is the chart:

Number of principal variations in P.B.
------------------------------------------------------
1 2 3 4 5
___________ ___________ __________ __________ __________

% Guess
rate 50% 33% 25% 20%

Note that for the chart, I purposely put the 50 % rate between the 1 and
the 2.
Whenever the guess rate is less than the number posted above you will
gain
by choosing that number of variations to choose from but only when you
are
100% certain that the correct move lies within those variations. Since
that
would only realistically happen around the 5 move limit, then only when
the guess rate is less than 20% would you gain from having 5 move
pondering . Therefore this whole argument revolves around the guess
rate.
Since all empirical evidence shows that pondering is in the 50% range
on average, then 1 move pondering is the optimum strategy. Now Fritz and
Hiarcs have implemented the k variation mode to take advantage of hash
hits so that the first 4 variations take only twice as long as the first
one. Even with this, as long as the guess rate is >=50%, 1 move
pondering
is still optimal.

[Robert Hyatt]

Komputer Korner (kor...@netcom.ca) wrote:
: Robert Hyatt wrote:

: >
: > only way to achieve 100% is to ponder *every* possible opponent move, then
: > you can't move instantly... because you've only searched a reply to his move
: > for < 10% of the total time you want to search before making a move... Why
: > is this topic still going on???

: It is still going on because Tom Kerrigan still doesn't understand

: that 1 K variation pondering is the optimum as long as the guess rate
: is >= to 50%.
: --
: Komputer Korner

: The inkompetent komputer.

I think I'm in the twilight zone here. I thought his point had to do with time
handicapping and pondering? the original thing that started this thread off?
He ponders just like the rest of us and understands why I think... but when
you handicap through time, pondering suddenly takes on a different character...

[Robert Hyatt]

mclane (mcl...@prima.ruhr.de) wrote:
: hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

: >: The game I speak about HAS TO DO WITH THE TOPIC we/you and others talk

: >: about. The problem is: you don't see this.

: >I certainly don't... thinking on the opponent's time is a mathematical

: >issue. It has nothing to do with games... It has to do with the mechanics
: >of using all available time effectively and efficiently... A single position
: >is just as important/useful...

: Than your mathematics sucks. Any 10 year old chess player could tell
: you, that thinking on 1 move is inefficient.
: If you would behave like your chess-program in a decision process in
: life, you would make as much mistakes as your program does.
: Ah - maybe that is the reason you behave this faulty in discussion.
: YOU ARE USING ONLY ONE PB-move !!

: Now I see it.... sorry.

You don't see a blasted thing, sorry. You are continually telling me how
a human does this. And I have agreed with you. But then I explain to you
*why* a computer can't do it like that, and you respond with how a human does
this and any other way is silly. I give you a mathematical reason why a computer
can't do it efficiently, and you respond with how a human does it. I give you
actual program output doing it "your way" which shows how inefficient it is, and
you respond with how a human does it. *one* of us is stuck in a rut...

: >I didn't say "I have known..." I gave two specific journal references

: >to provide even more information as well as details about the math I
: >offered to show why searching only one move is much more effective. It
: >was wasted effort, it seems, which is typical...

: If you search on one move, and it is not coming, what happens often,
: it must happen often, because it is systemimmanent of chess, you wated
: computing time. Chess has no OFTEN used best moves... chess has often
: many good moves, you cannot exactly say which one is better.
: I guess many chess-programs don't try this, because it would be to
: complicate to implement.

not too complicated. I did it this morning in an hour. Too inefficient to
use however, because it reduces the overall strength of the program. Analysis
was provided to explain why.

: >: >: If your computers would have used the PB time, and instead of assuming

: >: >: the shit f4 would come, computed on whatever strange moves, would have
: >: >: computed anyhow different , maybe not only looking for the best moves,
: >: >: they would MAYBE have seen Re5. If they would have seen it, they could
: >: >: have better used PB-time instead of been shocked about the move, and
: >: >: then dropped dropped and lost the game.

: >Not a chance... because it would take so much longer... In fact, it would
: >take so much longer that the last search tree (before the move we just made
: >so we could start pondering) would have just as much information. I showed
: >why. Most understood...

: Use the game, show me why the programs expect f4 and I show you that a
: next-best search about the position takes 2 seconds for the programs
: to find out: Re5 is better than f4.
: So if they waste 40 seconds on thinking: f4 is the move to come, what
: can I play against f4, they would have seen in 2 seconds next best,
: that re5 is the move to be prepared and could pb on Re5 instead.

Not true. I posted a search after Re5 and the score is worse for white, after a
deep search. Therefore avoiding the search of f4 would not let the machine learn
something about Re5, it already had found Re5 was worse. At the point Re5 was
played, it still believes f4 is best. I'd bet CSTal did also...

: But - as you said before.... it would take so much longer.

: And when GOD says, than it is true. Sorry. I always forgot that I am
: not allowed to even think different than you allow or extrapolate it.

you really have a serious mental problem with this religious fixation of
yours, and you probably need some sort of help. Was Sir Isaac Newton GOD
when he proved that E=M*V^2? He exposed a basic physical law of physics,
he didn't *create* that law. I didn't *create* the alpha/beta algorithm,
but I do know how it works. *completely*. And I tried to expose the flaw
in your reasoning so you could understand it... and all you can toss up is
comments about alpha/beta is stupid. That's not the issue. Of course it is
different than how humans search. No one disputes that. But it is the best
search strategy we've found so far. And everyone uses it. And apparently
everyone is stupid for doing so...

: Try it out in the position. I have tried it out before the whole

: discussion was involved, and my machines say: I am right.
: Whatever you blow in the sky here, it is just brabbling of words
: instead of using the game I posted and give concrete data on it.

Jesus, I believe I posted *concrete* data myself, on your game too, in fact.
Did you even bother reading it? Of course you didn't... but the search output
for Bxh3 was there, and the search output after Re5 was also there... as well
as a lot of other output to show why thinking about more than one move on the
opponent's time doesn't work efficiently. If anyone is "blowing", it is not
me. At least I give solid math, or real data to support what I said...

: Thanks your majesty.

You are welcome, I think...

[Komputer Korner]

Well, here is a direct quote from one of Kerrigan's postings:

"My view is that it could be a net gain for DB if it pondered the
best-looking handful of moves. That way, it "only" gets hit a ply,
whereas
Kasparov gets half his time sucked away."

Looks to me that Kerrigan doesn't understand Alpha Beta either along
with
understanding the math involved with the difference in ply lost between
instantaneous replies by Deep Blue and 3 minute thinks. If Deep Blue
adopts a policy of always replying instantly and if it's guess rate is
50%, then for any 4 ply interval, we have 0, 180 seconds, 0 , and 0
for a total of 180 seconds of thinking time. Under normal thinking
times,
Deep Blue would have 180,180,180, and 0 for a total of 540 seconds of
thinking time. Contrast this with Kasparov's thinking time under
Kerrigan's proposed strategy and normal strategy. Kerrigan's strategy
-Kasparov would have 0+180+0+180 =360 seconds. Normal strategy-
Kasparov
would have 720 seconds minus whatever wastage he has considering moves
that are irrelevant according to Deep Blue. So who is better off with
Kerrigan's strategy? Well, Deep Blue under Kerrigan's strategy gets
only an average of 45 seconds per move whereas Kasparov gets 90 seconds
per move. Of course, eventually Deep Blue gets an extra 2 hours at the
end but of course, Kasparov will also think on Deep Blue's extra 2
hours.
In En Passant # 105, I wrote an original treatise titled
Tactical Speed and Thinking Time. In it I related how time handicaps
were only really effective if one side couldn't ponder and I showed
the mathematical relationship between highest effective time
handicapping
and thinking on the opponents time. It is the inverse of the guess rate.
If the guess rate is 50%, then the highest effective time handicap
assuming pondering or P.B. is 2 to 1. Even if the time handicap is
a 1000 to 1 the ratio never goes above 2 which is the inverse of
the guess rate. So even at the end, Kasparov will benefit an hour from
Deep Blue's time, assuming Kasparov's time wastage is also 50% on the
oppoents time. Obviously Kerrigan's policy, ( I am calling it that
because
I can't remember the original poster and Kerrigan is now supporting it)
would make Kasparov very happy indeed. The only time that it would work
is
when Kasparov is in severe time pressure.

[Komputer Korner]

Peter Osterlund wrote:
snipped

> Hyatt claims that in order to save as much time as possible (on the
> average), you should only ponder the one move your opponent is most likely
> to play. This is also correct.
>
> You then try to conclude that Hyatt's claim is wrong because your own
> claim is correct, and this is where your logic fails. Do you suggest that
> we should use a pondering strategy that on average is much worse, just
> because in certain cases it would be slightly better?
>
> --
> Peter Österlund Email: peter.o...@mailbox.swipnet.se
> Sköndalsvägen 35 f90...@nada.kth.se
> S-128 66 Sköndal Homepage: http://home1.swipnet.se/~w-15919
> Sweden Phone: +46 8 942647

What Thorsten doesn't seem to understand is that the human's selective
search strategy has a great deal to do with why the human's pondering
is better, but it does bring up an important point. The pondering
strategy could be adjusted in programs that are highly selective search.
However almost all programs are basically full width with partial
selective search, so even this doesn't really work with today's
programs.

[Komputer Korner]

Robert Hyatt wrote:

>
> .321 xxxxxxxxxxxxxxxxx
> .300 yyyyyyyyyyyyyyyyy
> .336 zzzzzzzzzzzzzzzzz
>
> why is the last score higher than the supposedly best score? The
> answer is that the last score had information from the first two
> PV's that could influence it. When I cleared the hash table after
> each PV, the problem went away as expected. If you don't clear the
> hash, you get faster searches, but non-uniform values. If you do
> clear the hash table, the performance loss is serious (it's bad
> enough when you don't clear it...)

I don't see how clearing the hash table helps anything unless you
don't have enough RAM. However, we can end this thread in agreement
that as long as chess programs do a form of full width search,
alpha beta is optimal both for P.B. and normal thinking mode.

[Serge Desmarais]

Peter Osterlund wrote:
> =

> On Thu, 8 May 1997, mclane wrote:

> =

> : If you would concentrate on the game, instead of again IGNORING it,
> : you would see that Rebel/Genius/CST/... make a mistake in ONLY PB f4.

> =

> : Concentrate on the game. Or let it-.

> =

> Your claim is that in this game at the position you describe, it would
> have been better to ponder both Re5 and f4. This is correct because the

> move actually played was Re5, so pondering that move would have saved som=

e
> time on the computers clock.

> =

> =

> --
> Peter =D6sterlund Email: peter.o...@mailbox.swipnet.se
> Sk=F6ndalsv=E4gen 35 f90...@nada.kth.se
> S-128 66 Sk=F6ndal Homepage: http://home1.swipnet.se/~w-15919

> Sweden Phone: +46 8 942647

Maybe Maclane's point was also about the preceeding move which only anticip=
ated ...f4 =

as a drawing move? Was it a bishop sacrifice on h3? Also, when in a loosing=
position, =

and the opponent repeat the position to gain time on the clock before achie=
ving the =

wounded beast, maybe the computer should do something more useful than repe=
ating the =

losing position?

Serge

[Serge Desmarais]

Robert Hyatt wrote:
>
> mclane (mcl...@prima.ruhr.de) wrote:
> : hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
>

> : Sorry Bob, but IMO this totally wrong, the only argument you have is
> : that almost all pc-programs are doing it the way crafty is doing it
> : (50:50 chance).
>
> : Humans do it - of course - different.
> : If I have to extrapolate behaviour of an unknown opponent, I will OF
> : COURSE NOT ONLY THIN ABOUT
>
> : O N E
>
> : possibility. I would be a fool only to consider about ONE possible
> : way. It is also not very intelligent, therefore no humans are doing
> : it.
>
> : A human (i guess) is thinking about a few answers his opponent could
> : make, and he has an aswer for each branch, or the human is not clever.
> : If he would not have something in petto, he would oversee many threads
> : and moves against him.
>
> But you aren't a computer using alpha/beta either. You can compare two
> moves quite easily, and not figure out just which is better, but by how
> much it is better. And that has nothing to do with this discussion. I
> gave the simple math... take it apart, or refute it, but please don't
> change the subject with wild hand-waving and gesticulation... You don't
> like 50%? Pick a number... I'll revise my math using that number. But
> for my case, 50% is low... I have data to support this...
>
> : >If you analyze alpha/beta, searching the position below *one* move is about
> : >twice as efficient as searching the positions below *two* moves, because if
> : >you ponder one move, you only have to produce *one* PV, which is very
> : >efficient. If you ponder two different moves, you have to produce two
> : >different PV's, which is about 1/2 as efficient as only producing one.
>
> : >Therefore, searching (pondering) two moves by the opponent is a total loss
> : >in terms of time saved, when you compare everything...
>
> : I do neither understand your strange sentences above about searching 2
> : moves should be less efficient, nor do I think that you EXPLAIN
> : anything. The only thing you do is to spread bubbles of not filled
> : claims arround , instead of telling anything.
>
> : Your remarks about this topic remember me on the Flintstones. Of
> : course for them it would be much easier only to save 1 pv, because it
> : is so difficult to get the pv into stone.
>
> We aren't living in the same world, because you don't have a clue about
> alpha/beta. The above is perfectly clear to *anyone* that knows beans
> about alpha/beta minimax. If you don't understand AB, then you sure aren't
> going to follow the above discussion... Therefore I elect to save my
> breath...
>
> : We are not living in this time.
>
> : >If that's not clear enough, feel free to ask for more details...
>
> : How nice of you. As if you would know anyhing better. Oh - sorry, I
> : forgot you have Ed's and Chris' source-codes, what a pity you don't
> : have my sources...
>
> I have lots of blank pages... In the case of real computer chess internals,
> that seems to be your source code...

Just an idea : the more PV's are checked by the computer, the more time
it takes, but we don't know how long the opponent will thnink. So if you
use, say, 4 PV's while the opponent is thinking, when he will play,
assuming it is one of the 4 moves computed, the computer could then
get rid of the 3 others and continue thinking where it was in the tree...
That would mean that it would not play instantly very often... But
couldn't it reach more depth than starting from scratch cause the
opponent played a non computed move? I mean, if, when the opponent plays,
we were at a depth of 7 plies, then, we start with a 7 ply full width
tree for the move played? Of course, I may be wrong...

Serge

[mclane]

hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

> Why post a CSTal game when it is subject to the
>exact limits I posted. Do you claim it ponders (or permanent brain's)
>"differently"???

Why post a Kasparov-deepblue game when I don't like Kasparov nor Deep
Blue, but when I remember a game much better, because I have sit the
whole time behind the machine.
Also YOU have to construct that this could have happen or Kasparov
could have done this. I knew that f4 was pb and that Re5 came, so MY
example is more realistic.
The reason why you like to use your example is, that in my case the
permanent brain move was NOT played meanwhile in your example the PB
move WAS played. And this is also the reason why your whole example
and all your work sucks. IF the PB move is chosen by the opponent,
than of course if would have been better to do it the way YOUR program
and all the other programs do it.

So : this is the thing you have to claim a priori, otherwise your
whole construction will sink into its own weight.

I don't think that 10 moves are the general number of playable moves.
I have found out playing and analysing mail-chess-games, that normally
3 or 4 moves, are enough.

Of course the question: which moves shall we use is SILLY.
It is a shame I have to misuse the name of this eastern-german
pop-group in this context, but:
you get the moves AND the order in which you should use them from the
search before.
In my example the programs can handle this.
Ok - you do your example and I do mine:
I take Fritz4.01 and let it run over move 17 in analysis mode. I also
let it compute 4 different main-lines, that you will see how the move
order is, and

We have
1. f4 with 0,00
2. Re5 with -0,13
3.Bd7 with -1,63
4.Ba3 with -1,75.

You see, in my position I need only 2 branches, the rest is to weak...

Ok, Fritz is now in the 10 ply , arround 3 minutes computing time.
Still the main line gives stupid draw evaluation.
It takes some time , then branch 2 Re5 is also evaluated 0,00.
But white could consider 30 minutes or more, still f4 would be chosen.
When I now input Re5 it takes not more than 7 seconds and the score is
+0,03, after 26 seconds +0,09.

I would say that 0,09 is better than 0,00 ! (After 4'47" Fritz finds
out that Re5 is worth +0,38 - I would say Re5 wins the game, but Fritz
is of course a program that has no knowledge, so the values are not
very precise...)

Fritz would have found out with some easy thoughts, that Re5 is
better.
After these searching for WHICH MOVE IS THE BEST PB-MOVE, the program
could e.g. switch to the new candidate and compute on Re5 instead of
f4.
When , you speak about 3 minutes, when my machine has 3 minutes, it
has still arround 2 minutes (1 minute was spent on f4/Re5/Bd7/Ba3
evaluating.). My way, although it is not based on alpha - beta, would
increase the chances to ponder the right move, also it would have
helped in this position.

Your whole theory relys on presumptians that do not happen, because
your way of handling it is only working if the move is chosen.

If the chosen move is different, and this is more often the case than
you like, your method does not help. It is wasted time, and the
program could better calculate some tetris games instead of
calculating a wrong move.

If I would have the choice between a program finding the pondering
move, and a program finding it 1/4 times less, I would choose the
program that finds the pondering move.

>: Let crafty (and anybody if he has another program can try it) compute
>: on move 15...bxh3.

>what does this have to do with the issue of thinking on the opponent's
>time and whether it is better to think about one move or about two?
>I suppose I'm overlooking something here...

Right. You do overlook something. You should ask yourself :
why do I overlook such an easy thing.
I know the answer....

>The problem with your rhetoric is this: The old guys, the young guys,
>the middle-aged guys, the smart guys, the not-so-smart guys, and all the
>rest are *still* using alpha/beta my friend... You rant as though there
>is an alternative that is being used by someone in the "commercial" ranks.
>Produce it...

The problem with your rethoric is:
your argumentation is not more right just because NOBODY has ever
tried it out. This argumentation is typical for fat old men:
they say:
There are no stones falling from heaven, because in heaven there are
no stones.
Or: we don't have to test hiarcs on fast machines, because we KNOW
that hiarcs is not a strong program.
This rethorics is - of course - so smart, that only King Kongs would
follow it...

>Here's crafty's analysis:

> depth time score variation (1)
> 4-> 0.15 -0.046 Rd8 Be3 Qe4 Qd3
> 5 0.24 -0.001 Rd8 Be3 Bd5 Qg4 Qxg4 hxg4
> 5 0.43 0.078 a6 Bf1 Bd5 g3 Qf6
> 5-> 0.58 0.078 a6 Bf1 Bd5 g3 Qf6
> 6 0.99 -0.086 a6 Bf1 Bd5 g3 Qf6 Be3
> 6 1.99 -0.065 Rd8 Qd3 Bd5 Be3 Kg8 Qf5 Qe4 <HT>
> 6 2.68 -0.023 Bb4 Re3 Rd8 Bb2 Bc5 Rd3
> 6-> 2.89 -0.023 Bb4 Re3 Rd8 Bb2 Bc5 Rd3
> 7 4.04 -0.022 Bb4 Re2 Rd8 Bd2 Bxd2 Rxd2 Rd5 Bc4
> 7 6.03 0.056 a6 Bf1 Bd5 g3 Qf6 Bg2 Bxg2 Kxg2
> 7-> 7.54 0.056 a6 Bf1 Bd5 g3 Qf6 Bg2 Bxg2 Kxg2
> 8 10.99 -0.014 a6 Bf1 Rd8 Be3 Bc5 g3 Qf6 Qb1 Bxd4
> Bxd4 Qxd4 Qxb7
> 8 22.56 0.000 Bxh3 gxh3 Qxh3 f4 Qg3+ Kh1 Qh3+ Kg1
> 8-> 23.30 0.000 Bxh3 gxh3 Qxh3 f4 Qg3+ Kh1 Qh3+ Kg1
> 9 32.69 0.000 Bxh3 gxh3 Qxh3 f4 Qg3+ Kh1 Qh3+ Kg1
> 9-> 1:39 0.000 Bxh3 gxh3 Qxh3 f4 Qg3+ Kh1 Qh3+ Kg1
> 10 2:12 0.000 Bxh3 gxh3 Qxh3 f4 Qg3+ Kh1 Qh3+ Kg1
> 10-> 3:20 0.000 Bxh3 gxh3 Qxh3 f4 Qg3+ Kh1 Qh3+ Kg1
> 11 4:49 0.000 Bxh3 gxh3 Qxh3 f4 Qg3+ Kh1 Qh3+ Kg1

>so it agrees... repetition...

Briliant . Your program, using YOUR alpha-beta and YOUR way of PB
would waste time and permanent-brain on f4!

>: Helpfully Yona Kosashvilil did not what the stupid-programs expected.
>: He did a very normal and easy move, many chess-players would have
>: done: gave back the Rook with 17.Re5.

> depth time score variation (1)
> 9 9.11 0.340 Bxe5 dxe5 Qe6 Qd4 Kg8 Bg5 Rc8 Bd3
> h6
> 9-> 17.51 0.340 Bxe5 dxe5 Qe6 Qd4 Kg8 Bg5 Rc8 Bd3
> h6
> 10 36.08 0.325 Bxe5 dxe5 Qe6 Qd4 a6 Bd3 Re8 Ba3+
> Kg8 Bd6 h6
> 10-> 55.65 0.325 Bxe5 dxe5 Qe6 Qd4 a6 Bd3 Re8 Ba3+
> Kg8 Bd6 h6
> 11 1:47 0.269 Bxe5 dxe5 Qe6 Qd4 Kg8 Be3 Rc8 Rd1
> Qg6+ Kf1 Qc2 f4
> 11-> 4:45 0.269 Bxe5 dxe5 Qe6 Qd4 Kg8 Be3 Rc8 Rd1
> Qg6+ Kf1 Qc2 f4
> 12 9:04 0.230 Bxe5 dxe5 Qe6 Bf4 Kg8 Qf3 Qg6+ Kh2
> Qf5 Bd3 Qd7 Be4 Rb8

>So it still agrees, white's move was a mistake. I assume you have a point
>here? That for some reason, the program should be searching inferior moves
>for the opponent?

YOUR program - crafty - is so stupid that is even does not SEE after
the move has been made, that after Bxe5 WHITE is better.
Even Fritz evaluates better than crafty. Sorry. I have not thought
about this, I thought what Fritz sees would be easy for crafty.
I was wrong.
Beleive me, the position is better for white. But programs thinking
Re5 is better for white do also ponder on f4.

The POINT I have here, is the REASON I have chosen THIS example. It
was NOT my intention to post another game of CSTal, what you claimed
between the lines.

>: This is "brilliant" because we have a good example why machines are
>: stupid. And also it stops possible treatment by blacks troops, as
>: chris would express himself.

There is a little cheating in it, because the DRAW lines in the first
computation make any search come deeper very fast, meanwhile the
normal search in the position after Re5 is a normal search with no
repetition case.

>The machines are stupid if they misevaluate the game. But it has *nothing*

>to do with which move they happen to search. ... But it *wasn't* a result of not

>thinking on the opponent's time in your so-called "more intelligent" approach.

My point is here the following:
I do let a program compute 15 minutes.
It finds f4 but does not like Re5.
Than I input Re5 and it sees in a few seconds:
Re5 was better.
From my point of view, something must be wrong with these programs.
Because MANY programs do this error, I guess:
There is somehting in the programs , although they have all different
evaluations, that made them all make THIS major mistake.
If we can find out about this mistake, we can stop them expecting f4
instead of Re5 and maybe they would not have played Bxh3 at all.

My guess is, all programs play f4 because of alpha-beta or some
strange draw-behaviour, or because whatever. Let us find out.

>: The machines wasted permanent-brain computing on f4 main-lines, mainly
>: they found out that the position will end in a draw.

>And what's your point. The programs will continue to think that f4 is
>better, and that if white plays this move, the program is even better
>off than before...

You are right. In this case it does not change anything if the
computer had seen Re5 or not. Maybe if the computer would have seen
Re5 at move 15, but not at move 17.

But I could construct situations where the loss of pb
will cause significant changes to the game.

I - for myself - do still not believe that you like it that your
program ponders on the wrong move.
Computing on a few moves would help. Also help to better find the
right pondering move, as we have seen from this example.

>It lost because *it* played inferior moves too, or else *it* misevaluated
>the position and thought it went to a better position when, in fact, it
>went to a worse one. This sounds like ranting from someone that is really
>mentally unbalanced.

My point is:
Take a program that understands that Re5 is better than f4.
(Not yours - it does not understand like +0,23 shows me)
Again also this program will consider that f4 is the right move and
evaluates the position as a draw.
And than , after Re5 is on board, they suddenly see that f4 draw is
not the best score white could/should get.

My second point is:
If we have situations where one ply more seems to change the
point-of-view of a program in a drastically way, then it should be
senseful to do this in PB.
Lets gve the program the choice of finding out about the moves,
instead of trusting a search one interation before.
Experience show that it makes more sense.

>might not be a mistake. But that's not the point. That had nothing to
>do with his pondering only one move for the opponent. It would have played
>that if you had pondering turned off. As did crafty in the analysis I posted
>above.

It would have played this with or without pondering. No doubt about
this. But what if the move after Re5 would have been a sensible time
using move, and you have wasted your pondering on the wrong move?
With the alenative method you could have found out that Re5 was better
and would have pondered on the right move. This could have increased
the playing strength in a critical situation.

I do not understand
a) that you don't see this
b) don't see that my example of a position where almost any program
ponders on the wrong move, is a part of the pondering discussion.
Especially when my analysis with Fritz and other programs SHOW that
they need a very small amount of time to evaluate Re5 higher than f4
if they search seriell and not one ply before with alpha-beta.
Because with normal search they ALL play f4.

>: When I am right, Yona played h3 by purpose, to let us run into this
>: FALSE sac. He could have played g3, but in the moment he played h3, I
>: thought: he did it by purpose. Maybe Yona has seen Re5 while playing
>: h3.

>I really strongly doubt it. You should have asked, rather than resorting
>to your usual speculation. Facts are so much nicer to work with. Speculation
>is like swatting at a swarm of bees... it only makes things worse...

We can ask him, no problem. We get his email and we ask him.
I have registered that many players in den haag tried to let the
programs run into a trap. Playing h3 instead of g3 is:
making a compter trap. Despite of hiarcs6 , many programs would trap
into it.

[Robert Hyatt]

vince (vj...@doubled.com) wrote:

: fwiw, 2 rhetorical questions:
: 1. if higher plies generate less return, why not search alternate
: variations after plies reach a certain level of diminishing returns?

possible to do.

: 2. suppose there are 3 nearly equal variations? the chances could be 34%,
: 33%, 32% that they would be chosen by the opponent. wouldnt it be best to
: search and hash all three?

This is back to the same old problem however... it is not so easy to do,
because of the problems of searching multiple moves with a non-zero alpha
beta window...

: vince

[mclane]

Komputer Korner <kor...@netcom.ca> wrote:

> Since all empirical evidence shows that pondering is in the 50% range
>on average, then 1 move pondering is the optimum strategy.

I don't believe that chess programs, especially when they play against
humans, have pondering rates of 50 %.

>The inkompetent komputer.

That would also imply that chess is more based on binaric-trees ...

No - 50 % is to high, they have maybe 30% or less.

[mclane]

Peter Osterlund <peter.o...@mailbox.swipnet.se> wrote:

>Your claim is that in this game at the position you describe, it would
>have been better to ponder both Re5 and f4. This is correct because the

>move actually played was Re5, so pondering that move would have saved some

>time on the computers clock.

Although Bob is absolutely right: it would have not changed the game.
But there could be different positions where the right pondering can
help much.

>Hyatt claims that in order to save as much time as possible (on the
>average), you should only ponder the one move your opponent is most likely
>to play. This is also correct.

But WHAT is the one move ?! How can Hyatt know before ???
That is the reason HIS argumentation is RIGHT WHEN the ponder move
comes. But how often is that the case ?

>You then try to conclude that Hyatt's claim is wrong because your own
>claim is correct, and this is where your logic fails. Do you suggest that
>we should use a pondering strategy that on average is much worse, just
>because in certain cases it would be slightly better?

No - my point is that hyatt-ideas, statictics, maths, knowledge about
alpha-beta is all correct, but does not help the main problem that in
a chess-game there are often MORE than ONE good move.
If you ponder on ONE you reduce the chances that this move is chosen.
I know that ALL programs do it the way that it is decribed by Hyatt. I
also know the reasons WHY this is done.
Still this does not satisfy me, because I often see that programs
ponder wrong moves because the progams do use the main-line of the
last move, take the 2nd ply and MOVE it meanwhile computing in the
background and waiting for keyboard.

This will not work when the position is balanced and opponent can move
several moves.

Whatever you think about the methods, IMO you could try to enhance the
way that it is done in the moment.
I have many ideas about this and would like to implement some, but of
course if somebody reacts the way Bob does it, he provokes my
anti-force.

>--
>Peter Österlund Email: peter.o...@mailbox.swipnet.se
>Sköndalsvägen 35 f90...@nada.kth.se

>S-128 66 Sköndal Homepage: http://home1.swipnet.se/~w-15919

[mclane]

hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

>mclane (mcl...@prima.ruhr.de) wrote:
>: hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

>I'd bet you are incorrect, because it is an illogical way of thinking on
>the opponent's time.

Hey - you want to provoke me, he ?
You think I am incorrect although I tested this machine and wrote the
article some years ago although YOU have never seen the machine and
have no knowledge about dedicated-chess-computers anyway.
How can I call this in other words, others than IGNORANCE and IDLE
RANTING.
"I have written an article about this machine long time before you
were born!" I could say now.
Unbelievable. This man behaves like a GOD.

>Dredge up the model, I'll find out who wrote the
>program and we can ask them directly.

I have to look into my COMPUTER-SCHACH & Spiele editions, because it
is years ago. I hope it is not under the editions I lend to Tom
Kerrigan...
BTW: TOM, where are you. Any chance to get my CSS-editions back?
I would come by car and catch them. Of course if you are in US, that
would be a rough travel.

>They make a mistake *on that move* only. What about on the moves where they
>do predict correctly and save the time? I posted analysis with real data from
>Crafty to explain why doing it your way won't work very well. Nothing wrong
>with your *idea* as I mentioned. But within the framework of current computer
>chess using alpha/beta it is *totally* unworkable... *lots* of ideas are
>unworkable... I've pointed out the "discernability" problem of a program
>choosing to lose obviously rather than choosing to lose subtly, because in that
>case obvious was a few "millipawns" better. Idea is good. But alpha/beta doesn't
>offer up a solution...

The idea is: one ply later many programs see the position different.
E.g. Fritz and maybe Hiarcs and others (not crafty and zarkov4.1, ok)
WOULD see that Re5 is better and would exchange the PB move with the
better move. That's still not a guaranty that Yona plays Re5.
But I would have a main-line and hash and evals for candidate 4 moves.
Also in my case, and from my experience, I think there are many
positions where this strange effects happen, that must have to do with
alpha - beta or pruning effects, systemimmenant of the program, that
cause:

Search of 10 plies finds move a in the whole computation time of 3
minutes.

You input another answer than the expected, and suddenly they drop
with evals into seconds.

This is not possible in a full-width program. But we don't have these
brute-forcers anymore.

So I guess in our times your statistics do not work anymore, because
the programs have changed.

In the old days of Fidelity Mach3 or Mephisto2 or Novag Forte A it was
not useful to do permanent brain ideas like multiple pb with
exchanging the ponder move by finding out different result.

But in todays time the programs are much more selective and I colletc
positions where the programs behave the above described way.
I mean they prune away a move they think this move is weak.
You can let them compute a long time. They don't come to any new
ideas. Than you input the KILLER-MOVE to this pos. and they instantly
SEE within seconds that their whole search tree that was build minutes
before was shit.

This happens in last years more and more often. You cannot trust them
anymore because they do more and more forward pruning or neat
search-techniques that have problems with special cases.

Of course some programs (crafty, ferret, stobor, zarkov) do not react
often like the above.

But it is true for:
Genius (especially for genius because of the asymmetric search,
remember the Karpov-Topalov Rxe6 position), Mchess, Hiarcs, Fritz,
CSTal...

I am collecting these positions because I like them to fail with the
search. And sometimes they fail because misevaluation.

BTW: please do me a favour, don't speak in these absolute terms with
me. Believe me, I have programs at home without search tree, only
having static-search (Chat) and others only making 4-10 NPS (Mephisto
III) and others like Gandalf doing one ply with extensions, and also
dedicated chess computers with very strange effects.

I am - over the years - not because I am an expert, but because I do
this hobby such a long time, somehow having a routine into it.

When you suddenly tell: It cannot be. Such a machine cannot exists.

I have the same problem with it e.g. like I would come into rgcc and
would post:

Deep Blue plays 2000 ELO chess and is not existing, just an idea of
some capitalist.

You would attack me heavily if I would do so. Please recognize that
your way of saying

>I'd bet you are incorrect, because it is an illogical way of thinking on
>the opponent's time.

is provoking me, because I HAVE SEEN THIS MACHINE AND WE TESTED THE
STUFF WITH THE MANY PONDERING MOVES.

Of course it was years ago, ok, but e.g. Saitek has created some very
crazy machines. In the low-cost area the programmers try some new
ideas. That has happened to Julio Kaplans low-costs Saitek machines as
it has happened to Kittinger's low-costs machines. Also Frans Morsch
tried out strange ideas in the low-costs machines. NOT IN THE TOP
machines. But in the other machines.
Also David Levy's machines were all very crude and had uneblievable
details and behaviour (I just want to mention that Mark 5 simulated,
as if it computed abot the move, although the move was played out of
the book !).

Here one thing: Don't underestimate the creative ideas of people.
I have learned to honor the work of people over the years.
So many things I have seen in the past were outstanding, and due to
the love and passion of chess-programmers who made my life exiting
with their work.

I have an 8bit chess computer that does not play a sac when I input
the position.
But it plays the sac of bishop Bd6xa3 when the machine has moved the
bishop before.
Here the game with the strange 8-Bit machine behaviour....

BTW: don't think that the machines have played BORING chess.
I always had a faible for chess-programs that play interesting chess.
In these times some nasty distributors called me Thorsten Saitek, only
because I liked the programs of Julio Kaplan, who were never at the
top, but played good positional stuff, meanwhile other testers
preferred the big and expensive Mephisto machines.

Here Ed Schroeders Mephisto Milano (6502, 5Mhz, 48KROM) has to fight
against Julio Kaplans Saitek Maestro D++ 10 Mhz.
It was a tournament game 40 in 120.
Julio must have implemented some experimental stuff into this machine
because it plays like the 8-Bit versio of Chess System Tal.
Because the saitek machines can be connected with the PC, here the
output...

Mephisto Milano-Saitek Maestro D++
40 in 120
1. c2-c4 e7-e6
2. d2-d4 d7-d5
3. Nb1-c3 Ng8-f6
4. Bc1-g5 Bf8-e7
5. c4xd5 e6xd5
6. e2-e3 c7-c6
7. Bf1-d3 O-O
8. Ng1-e2 Nb8-d7
9. Qd1-c2 Rf8-e8
10. O-O-O Nd7-f8
11. h2-h3 a7-a5
12. g2-g4 Nf8-g6
13. f2-f3 h7-h6
14. Bd3xg6 h6xg5
15. Bg6-d3 Be7-b4
16. e3-e4 Bc8-e6
17. a2-a3 Bb4-e7
18. Rh1-e1 b7-b5
19. e4-e5 Nf6-d7
20. f3-f4 b5-b4
21. f4-f5 b4xc3
22. Ne2xc3 Nd7xe5
23. d4xe5 Be6-c8
24. e5-e6 f7xe6
25. f5xe6 Be7-d6
26. Bd3-h7+ Kg8-h8
27. Qc2-g6 Bc8xe6
28. Re1xe6 Re8xe6
29. Qg6xe6 Kh8xh7
30. Kc1-c2 Qd8-f8
31. Nc3-e4 Bd6-f4
32. Ne4-c3 Qf8-c5
33. Kc2-b1 Ra8-f8
34. Nc3-a4 Qc5-c4
35. Na4-c3 Qc4-b3
36. Kb1-a1 Rf8-f6
37. Qe6-e2 Bf4-d6
Sendinfo = 00:04 1 -00207 f6-f7
Sendinfo = 00:05 1 -00205 f6-f7 e2-d3 h7-g8
Sendinfo = 00:05 2 -00205 f6-f7 e2-d3 h7-g8
Sendinfo = 00:05 2 -00182 f4-d6 e2-a6 d6-c7
Sendinfo = 00:07 2 -00183 f4-c7 e2-e7 c7-b6
Sendinfo = 00:08 3 -00183 f4-c7 e2-e7 c7-b6
Sendinfo = 00:11 3 -00160 f4-c7 e2-d3 h7-h8 d3-e3 f6-g6 e3-e8 h8-h7
Sendinfo = 00:17 3 -00177 f6-h6 e2-d3 h7-g8 d3-a6 h6-h3 a6-c8 g8-h7
c8-c6
Sendinfo = 00:24 4 -00177 f6-h6 e2-d3 h7-g8 d3-a6 h6-h3 a6-c8 g8-h7
c8-c6
Sendinfo = 00:40 4 -00173 f6-h6 d1-d3 b3-c4 e2-f3
Sendinfo = 01:02 5 -00173 f6-h6 d1-d3 b3-c4 e2-f3
Sendinfo = 01:51 5 -00159 f6-h6 d1-h1 f4-d6 e2-d3 h6-g6 h1-f1
Sendinfo = 04:40 5 -00162 f4-d6 e2-d3 h7-h8 d3-e3 a5-a4 d1-e1 d6-f4
e3-e8
Sendinfo = 07:05 6 -00162 f4-d6 e2-d3 h7-h8 d3-e3 a5-a4 d1-e1 d6-f4
e3-e8
Move Ready 09:52

Without this move, the machine will not play
39...Bxa3. If you input the position AFTER the move 37...Bd6 has
occured and try to replay the game, the machine will NOT play Bxa3.
If you force the 37.move BEFORE the main-line with the Bd6 occurs, and
then try to get 39...Bxa3 it will not work.
So it looks to me that FIRST Bd6 must be found by the program to make
Bxa3 happen. Maybe Kaplan programmed a kind of RULE-BASE and the sac
bishop sac can only be loaded out of the rule-base when the bishop
moved before, or whatever Kaplan does, he has managed to save some
stuff in ram that causes this effect.
And this in an 8bit machine with 16K Ram and NO HASH. Maybe kaplan
saved the main-lines, or whatever, preprocessing?
I don't know.
We found out about the important Bd6 because I published the
SAC-position Bxa3 and some testers having this experimental version
too were unable to reproduce. I setup the sac-position and was unable
to get Bxa3 myself. Than I though: it must happen from the game. So I
setup the position always on move earlier and went backstepping
through the came. I found out how critical Bd6 main-line is for Bxa3 !
38. Qe2-d3+ Kh7-h8
39. Qd3-e3 Bd6xa3

- - - - - - - k
- - - - - - p -
- - p b - r - -
p - - p - - p -
- - - - - - P -
P q N - Q - - P
- P - - - - - -
K - - R - - - -

Sendinfo = 00:00 2 -00184 d6-f4 e3-e8 h8-h7 e8-h5 f6-h6
Sendinfo = 00:06 2 -00119 d6-f4 e3-a7
Sendinfo = 00:07 2 -00151 d6-a3 b2-a3 b3-a3
Sendinfo = 00:09 2 -00172 d6-a3 b2-a3 b3-a3 a1-b1
Sendinfo = 00:14 3 -00172 d6-a3 b2-a3 b3-a3 a1-b1
Sendinfo = 00:17 3 -00176 d6-a3 b2-a3 b3-a3 a1-b1
Sendinfo = 00:46 4 -00176 d6-a3 b2-a3 b3-a3 a1-b1
Sendinfo = 00:50 4 -00173 d6-a3 b2-a3 b3-a3 a1-b1 a3-b3 b1-a1
Sendinfo = 02:01 5 -00173 d6-a3 b2-a3 b3-a3 a1-b1 a3-b3 b1-a1
Sendinfo = 02:21 5 -00173 d6-a3 b2-a3 b3-a3 a1-b1 a3-b3 b1-a1
Move Ready 02:37
Bd6xa3 02:37

40. b2xa3 Qb3xa3+
41. Ka1-b1 Qa3-b3+
42. Kb1-c1 a5-a4
43. Kc1-d2 a4-a3
44. Rd1-b1 Qb3xb1
45. Nc3xb1 a3-a2
46. Nb1-c3 a2-a1/Q
47. Kd2-c2 Qa1-f1
48. Nc3-d1 Qf1-c4+
49. Kc2-d2 Qc4-f4
50. Kd2-d3 Qf4xe3+
51. Kd3xe3 c6-c5
52. Ke3-e2 d5-d4
53. Ke2-d3 Rf6-f3+
54. Kd3-e4

If you would change the parameters the way that you setup the position
after Bd6 or do not wait until Bd6 is found in 4'40", than you get the
boring main-lines in move 39...:
Sendinfo = 00:00 2 -00160 h8-h7 d1-e1 d6-f4 e3-c5 a5-a4
Sendinfo = 00:00 3 -00160 h8-h7 d1-e1 d6-f4 e3-c5 a5-a4
Sendinfo = 00:03 3 -00153 h8-h7 e3-d3 f6-g6 d1-f1
Sendinfo = 00:05 3 -00169 d6-f4 e3-e8 h8-h7 d1-e1
Sendinfo = 00:31 4 -00169 d6-f4 e3-e8 h8-h7 d1-e1
Sendinfo = 00:45 4 -00150 d6-f4 e3-c5 f4-c7 c5-e3 f6-g6 e3-e8 h8-h7

instead of Bxa3 !!!!

BTW: it is a nice game, isn't it ?
Also: If anyone is interested in getting this version for the
Maestro/analyst modules, he can ask me. We can easily burn the
eproms...

[mclane]

"Torstein Hall" <tors...@eunet.no> wrote:

>You don't have to be a psycologist to find out that mclane is
>.............

Please continue this. If you are not wanting to post this in public I
would like you to do it in an email.

You should learn to stand your words and not hide behind
.............

[mclane]

hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

> But then I explain to you
>*why* a computer can't do it like that, and you respond with how a human does
>this and any other way is silly.

Again: A computer CAN do it different than you do it.

> I give you a mathematical reason why a computer
>can't do it efficiently,

wrong again: a computer can do it more efficiently than crafty does
it's pondering.

>and you respond with how a human does it. I give you
>actual program output doing it "your way" which shows how inefficient it is

You have NOT shown me how inefficient my way was, your crafty was
unable to understand the whole position, therefore crarfty would not
have changed the f4 move into Re5, right ?

Also your analysis does not show that it is inefficient.

>not too complicated. I did it this morning in an hour. Too inefficient to
>use however, because it reduces the overall strength of the program. Analysis
>was provided to explain why.

No - analysis was not enough, maybe you have to invest more time .
How have you tried it out?

>Not true. I posted a search after Re5 and the score is worse for white, after a
>deep search. Therefore avoiding the search of f4 would not let the machine learn
>something about Re5, it already had found Re5 was worse. At the point Re5 was
>played, it still believes f4 is best. I'd bet CSTal did also...

RIGHT. For the first time in this discussion you are right. CRAFTY is
so stupid that it is unable to see that Re5 is better than f4.
OF COURSE therefore your RESULTS are useless, because YOUR program
cannot compute any positive and efficient advantages due to a new
permanent brain, especially NOT in this position that crafty seem not
to understand.

>you really have a serious mental problem with this religious fixation of
>yours, and you probably need some sort of help.

Thanks for your care. I can give this back.

>Was Sir Isaac Newton GOD
>when he proved that E=M*V^2?

Isac newton has not proved anything. And Isac Newton, if you would
have ever read anything deeply about Isac Newton, was a very religious
man who beliefed in GOD.

If you have studied physics any better, you would have registered
somehow, that Newton was not the PROTOTYPE of a scientists you
sometimes want to appear here:
Newton has written more about problems of the soul and alchemie and
that what we would call today para-psychology than about his optical
or gravitational topics.
He was very spritistic. He wrote about spirit and light (in his work
about optic):
"Couldn't it be possible, that the materia and the light do exchange
themselves into each other. And wouldn't it not possible that the
materia takes most of their active forces out of the light-particles
that exchanged into it? If so, couldn't the light, as the one main
principle in nature, be the power, that directs the activitates of the
materia ?"
Later in this text he differenciates between 32 kinds of light, the
normal light we see, and the one that is potential, that is somehow
bound within the living organism and is the habit for the spirit.

So Newton thought that the life that comes in organisms comes from the
numenal/potential light that is bound in the materia, as the habit/or
vehicle for the spirit.

Read Jean Zafiropulo and Catherine Monod's : Sensorium Dei, Paris 1976
or P.M.Rattensis views about Newtons work in Science, Medicine and
Society in the Renaissance, London 1972.

Please stop quoting Newton in a misused context with you.
You have not the size of newton.

>Jesus, I believe I posted *concrete* data myself, on your game too, in fact.

again, i have no reason to oversee your posts, because I like to learn
and discuss with you. Your ideas that I would oversee something you
post by purpose is ridiculous. I have never done this. It is not my
way handling anything.

>Did you even bother reading it? Of course you didn't... but the search output
>for Bxh3 was there, and the search output after Re5 was also there... as well
>as a lot of other output to show why thinking about more than one move on the
>opponent's time doesn't work efficiently. If anyone is "blowing", it is not
>me. At least I give solid math, or real data to support what I said...

Wrong. Your data did not show anything imnportant.

[mclane]

Komputer Korner <kor...@netcom.ca> wrote:

>However almost all programs are basically full width with partial
>selective search, so even this doesn't really work with today's
>programs.

I don't understand why you always generalize in such a wrong and very
sensible topic for me.
NO - the programs I know do NOT basically full width with partial
selsctive search.

If so, they would not fail so often in easy combinations although they
reached the search depth.
You are wrong. And I don't know WHY you repeat these wrong statements
over and over.

I don't want to start a discussion about this topic with you, but you
can believe me, the programs are less full-width than you expect them
to be.

[mclane]

hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

>mclane (mcl...@prima.ruhr.de) wrote:
>: hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

>I have posted analysis for your example game. Get off your butt, find it
>and study it. Then come back with something that makes sense. Anybody else
>reading here that did not see my computer analysis of his game? Or is he
>the only one???

Ok - i have studied the analysis YOU gave me, but you not the analysis
I gave you.

YOUR program (like e.g. zarkov 4.1 and maybe others) does not see that
Re5 is better (it evaluates +0,23 for BLACK although it should be
winning for white.) than 0,00 draw evaluation.

Therefore my game does not fit with your crafty in example.
But Fritz and Hiarcs come , as I told you, in a few seconds, with a
positive evaluation >0,00 for white when I force Re5 and let them
play.

I cannot give you sendinfo because these win95 programs do not
generate sendinfo-data.
But I wrote the times and branches and evaluations down manually.

These TIMES show, that it would make sense to compute on Re5 because
the BREAKTHROUGH in changing the opinion came for Fritz4 on my
Pentium120 within 27 seconds!
This would have changed Fritz decision and would not have lost much
time....your statistics about HOW LONG it would take and that it will
not be efficient do not work anymore, because the programs do NOT pure
alpha-beta but prune much away and show many strange behaviours.

Please take into your consideration that the mail you post in US comes
with a special delay onto my usenet server.
I don't know why but I have very often quoted data by european
posters, and later I download the US-original.
I have seen you analysis and will - of course - take a look.
Also I posted the search times and the evals of fritz4.01.
But you did not commented on this. WHY?

[chrisw]

--
http://www.demon.co.uk/oxford-soft

mclane <mcl...@prima.ruhr.de> wrote in article
<5krr5t$js0$3...@steve.prima.ruhr.de>...
> "chrisw" <chr...@cpsoft.demon.co.uk> wrote:
>
>
>
>
> >> : If Bob is against this method, I will tell Chris
>
> >Stand by to be flamed Thorsten.
>
> >You don't TELL me to do anything, right ?
> AH brilliant. Thats what friend are for. Helping themselves in the bad
> times, sharing emotions and life.
>
>
> >I'm TELLING you to get your arse over here to Burford and do some work,
> >instead of lounging around on german unemployment pay ........... :)
>
> Why should I go to britain and work WITH you (I would never work FOR
> you!!) when you know anything better and nobody, including me, would
> be allowed to tell you anything ????
>
> Yeah - you and Bob are very similar in this behaviour !
> You don't let anybody tell you anything....

No, no. I listen to the trees. Bob thinks it good to explain to trees.
That's the difference :)

Chris Whittington

>
> >Chris Whittington
>
>
>
>
>

[Robert Hyatt]

mclane (mcl...@prima.ruhr.de) wrote:
: hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

: > Why post a CSTal game when it is subject to the
: >exact limits I posted. Do you claim it ponders (or permanent brain's)
: >"differently"???

: Why post a Kasparov-deepblue game when I don't like Kasparov nor Deep
: Blue, but when I remember a game much better, because I have sit the
: whole time behind the machine.
: Also YOU have to construct that this could have happen or Kasparov
: could have done this. I knew that f4 was pb and that Re5 came, so MY
: example is more realistic.
: The reason why you like to use your example is, that in my case the
: permanent brain move was NOT played meanwhile in your example the PB
: move WAS played. And this is also the reason why your whole example
: and all your work sucks. IF the PB move is chosen by the opponent,
: than of course if would have been better to do it the way YOUR program
: and all the other programs do it.

And if the PB move is right 50% of the time or greater, what do you
respond then? I now have Crafty keeping up with this statistic. I can
post a few... after the new version has played a few games. If it
predicts 1/3 or better, the current scheme is clearly better. Numbers
I have seen while watching is almost always 50% or greater correct
predictions. In light of that statistic, what would you suggest? Give
up on a >50% correct approach for one that is obviously going to be less
than 50% correct?

: So : this is the thing you have to claim a priori, otherwise your

: whole construction will sink into its own weight.

: I don't think that 10 moves are the general number of playable moves.
: I have found out playing and analysing mail-chess-games, that normally
: 3 or 4 moves, are enough.

To *you* or to *me* yes, but to a computer? That's nonsense. You should
take any program that is selective at the root, and set the width to 4 and
play a game. Easy win. Every time. Because there's no real way to find
out which are the best 4 moves except to do a search, and after you've
done a search to find them, it would be a little redundant to then search
them again.

: Of course the question: which moves shall we use is SILLY.

: It is a shame I have to misuse the name of this eastern-german
: pop-group in this context, but:
: you get the moves AND the order in which you should use them from the
: search before.

No you don't, not *ever*, and we've been over this again and again.
Alpha/Beta does not tell you *anything* other than move N is the best,
with a value of X. You have *no* idea which move is second, or third,
or anything else. You end up with two pools of moves: pool (a) has
only one move, and it is the best; pool (b) has all the others, and
you have no idea about how that pool should be ranked in terms of
which is best and which is worst. No idea at all...

: In my example the programs can handle this.

how???

: Ok - you do your example and I do mine:

: I take Fritz4.01 and let it run over move 17 in analysis mode. I also
: let it compute 4 different main-lines, that you will see how the move
: order is, and

: We have
: 1. f4 with 0,00
: 2. Re5 with -0,13
: 3.Bd7 with -1,63
: 4.Ba3 with -1,75.

: You see, in my position I need only 2 branches, the rest is to weak...

Sure, but what did it cost you to find out that you needed only two
branches? And what about positions where there are several nearly
equal moves? etc...

: Ok, Fritz is now in the 10 ply , arround 3 minutes computing time.

: Still the main line gives stupid draw evaluation.
: It takes some time , then branch 2 Re5 is also evaluated 0,00.
: But white could consider 30 minutes or more, still f4 would be chosen.
: When I now input Re5 it takes not more than 7 seconds and the score is
: +0,03, after 26 seconds +0,09.

: I would say that 0,09 is better than 0,00 ! (After 4'47" Fritz finds

Are you reading the eval correctly? +.09 sounds good for black, since
it is black's move, which is exactly what crafty found. Crafty and Fritz
are similar in search although far different in evaluation. But here there's
no tactics... and the search I gave you for Re5 was 12 plies deep...

: out that Re5 is worth +0,38 - I would say Re5 wins the game, but Fritz

: is of course a program that has no knowledge, so the values are not
: very precise...)

: Fritz would have found out with some easy thoughts, that Re5 is
: better.
: After these searching for WHICH MOVE IS THE BEST PB-MOVE, the program
: could e.g. switch to the new candidate and compute on Re5 instead of
: f4.
: When , you speak about 3 minutes, when my machine has 3 minutes, it
: has still arround 2 minutes (1 minute was spent on f4/Re5/Bd7/Ba3
: evaluating.). My way, although it is not based on alpha - beta, would
: increase the chances to ponder the right move, also it would have
: helped in this position.

: Your whole theory relys on presumptians that do not happen, because
: your way of handling it is only working if the move is chosen.

No, my analysis included pecentages, because every move is not correctly
predicted, but it is predicted correctly over 50% of the time in the games
I have on hand...

: If the chosen move is different, and this is more often the case than

: you like, your method does not help. It is wasted time, and the
: program could better calculate some tetris games instead of
: calculating a wrong move.

I have data to the contrary. *real* data, not imagined..

: If I would have the choice between a program finding the pondering

: >so it agrees... repetition...

As would *every* program. Either we are all wrong, or......

: >: Helpfully Yona Kosashvilil did not what the stupid-programs expected.

May be. I don't agree however...

: The POINT I have here, is the REASON I have chosen THIS example. It

I don't like it at all... but I haven't found a way to stop it either...

: Computing on a few moves would help. Also help to better find the

: right pondering move, as we have seen from this example.

: >It lost because *it* played inferior moves too, or else *it* misevaluated
: >the position and thought it went to a better position when, in fact, it
: >went to a worse one. This sounds like ranting from someone that is really
: >mentally unbalanced.

: My point is:
: Take a program that understands that Re5 is better than f4.
: (Not yours - it does not understand like +0,23 shows me)
: Again also this program will consider that f4 is the right move and
: evaluates the position as a draw.
: And than , after Re5 is on board, they suddenly see that f4 draw is
: not the best score white could/should get.

That's a program problem, *not* an alpha/beta problem. Alpha/Beta
*guarantees* that the move found is the best, based on the evaluation
and material. Anything else is a program issue. Fritz does a lot of
computation at the root of the tree, before it starts the search. This
could easily explain the difference, because I don't like the root-eval
idea for this reason. I'd suspect that causes your Re5 to look slightly
better for white, because after preprocessing the piece/square tables
and whatever else he does, they are different than they were when the
original Bxh3 search was done, and the eval is now different. I don't
do this in Crafty. Many other programs also don't do it, because it
produces wrong answers...

: My second point is:

: If we have situations where one ply more seems to change the
: point-of-view of a program in a drastically way, then it should be
: senseful to do this in PB.
: Lets gve the program the choice of finding out about the moves,
: instead of trusting a search one interation before.
: Experience show that it makes more sense.

: >might not be a mistake. But that's not the point. That had nothing to
: >do with his pondering only one move for the opponent. It would have played
: >that if you had pondering turned off. As did crafty in the analysis I posted
: >above.

: It would have played this with or without pondering. No doubt about
: this. But what if the move after Re5 would have been a sensible time
: using move, and you have wasted your pondering on the wrong move?
: With the alenative method you could have found out that Re5 was better
: and would have pondered on the right move. This could have increased
: the playing strength in a critical situation.

: I do not understand
: a) that you don't see this
: b) don't see that my example of a position where almost any program
: ponders on the wrong move, is a part of the pondering discussion.
: Especially when my analysis with Fritz and other programs SHOW that
: they need a very small amount of time to evaluate Re5 higher than f4
: if they search seriell and not one ply before with alpha-beta.
: Because with normal search they ALL play f4.

you are mixing up lots of issues, because changing its mind from one
move to another is a common characteristic of a program that does part
of its evaluation at the root, and only small pieces at the tips. I
(and most other) programs don't do this for this very reason...

: >: When I am right, Yona played h3 by purpose, to let us run into this

: >: FALSE sac. He could have played g3, but in the moment he played h3, I
: >: thought: he did it by purpose. Maybe Yona has seen Re5 while playing
: >: h3.

: >I really strongly doubt it. You should have asked, rather than resorting
: >to your usual speculation. Facts are so much nicer to work with. Speculation
: >is like swatting at a swarm of bees... it only makes things worse...

: We can ask him, no problem. We get his email and we ask him.
: I have registered that many players in den haag tried to let the
: programs run into a trap. Playing h3 instead of g3 is:
: making a compter trap. Despite of hiarcs6 , many programs would trap
: into it.

*if* it's a trap. He had to return the sac to avoid the draw. CSTal lost
after that. But it's not much of a trap if the position is pretty equal
after the trap is "sprung"...

[chrisw]

--
http://www.demon.co.uk/oxford-soft

Robert Hyatt <hy...@crafty.cis.uab.edu> wrote in article
<5kv97q$r...@juniper.cis.uab.edu>...

This is only partly true.

You can get an better than random ordering of the remaining moves out of
alpha beta.

Its not perfect, and in some cases its well wrong, but its not wholly
random for pool (b).

Chris Whittington

[Robert Hyatt]

Serge Desmarais (psy...@total.net) wrote:
: Robert Hyatt wrote:
: >

: > mclane (mcl...@prima.ruhr.de) wrote:
: > : hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
: >

: > : Sorry Bob, but IMO this totally wrong, the only argument you have is

: > : that almost all pc-programs are doing it the way crafty is doing it
: > : (50:50 chance).
: >
: > : Humans do it - of course - different.
: > : If I have to extrapolate behaviour of an unknown opponent, I will OF
: > : COURSE NOT ONLY THIN ABOUT
: >
: > : O N E
: >
: > : possibility. I would be a fool only to consider about ONE possible
: > : way. It is also not very intelligent, therefore no humans are doing
: > : it.
: >
: > : A human (i guess) is thinking about a few answers his opponent could
: > : make, and he has an aswer for each branch, or the human is not clever.
: > : If he would not have something in petto, he would oversee many threads
: > : and moves against him.
: >
: > But you aren't a computer using alpha/beta either. You can compare two
: > moves quite easily, and not figure out just which is better, but by how
: > much it is better. And that has nothing to do with this discussion. I
: > gave the simple math... take it apart, or refute it, but please don't
: > change the subject with wild hand-waving and gesticulation... You don't
: > like 50%? Pick a number... I'll revise my math using that number. But
: > for my case, 50% is low... I have data to support this...
: >

: > : >If you analyze alpha/beta, searching the position below *one* move is about
: > : >twice as efficient as searching the positions below *two* moves, because if

No... your analysis is right. But if you carry it through, you find that
overall you lose time doing this... Whether that is bad is another issue,
of course...

[Robert Hyatt]

mclane (mcl...@prima.ruhr.de) wrote:
: hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

: > But then I explain to you

: >*why* a computer can't do it like that, and you respond with how a human does
: >this and any other way is silly.

: Again: A computer CAN do it different than you do it.

: > I give you a mathematical reason why a computer

: >can't do it efficiently,

: wrong again: a computer can do it more efficiently than crafty does
: it's pondering.

: >and you respond with how a human does it. I give you
: >actual program output doing it "your way" which shows how inefficient it is

: You have NOT shown me how inefficient my way was, your crafty was
: unable to understand the whole position, therefore crarfty would not
: have changed the f4 move into Re5, right ?

: Also your analysis does not show that it is inefficient.

it does, if you will read it. If you won't read it, it won't show you
anything...

: RIGHT. For the first time in this discussion you are right. CRAFTY is
: so stupid that it is unable to see that Re5 is better than f4.

You haven't done a thing to convince me Re5 is better, other than to
wave your arms wildly... That won't win many games, unfortunately...

: Wrong. Your data did not show anything imnportant.

Then I guess this is hopeless to continue...

[Robert Hyatt]

Komputer Korner (kor...@netcom.ca) wrote:
: Robert Hyatt wrote:

Because the hash entries used to compute the value for .321 will then
affect the tree used to produce the value for .300. and the hash
entries from both of these will then affect the value for the last
search. If you like, I can run crafty with and without clearing and
show you how this happens. Running out of ram would be better, because
that would partially prevent this. However, my machine has 128mb and
this *isn't* a problem...

: The inkompetent komputer.

[Robert Hyatt]

mclane (mcl...@prima.ruhr.de) wrote:
: hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

: >mclane (mcl...@prima.ruhr.de) wrote:
: >: hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

: >I have posted analysis for your example game. Get off your butt, find it

: >and study it. Then come back with something that makes sense. Anybody else
: >reading here that did not see my computer analysis of his game? Or is he
: >the only one???

: Ok - i have studied the analysis YOU gave me, but you not the analysis
: I gave you.

: YOUR program (like e.g. zarkov 4.1 and maybe others) does not see that
: Re5 is better (it evaluates +0,23 for BLACK although it should be

: winning for white.) than 0,00 draw evaluation.

: Therefore my game does not fit with your crafty in example.
: But Fritz and Hiarcs come , as I told you, in a few seconds, with a
: positive evaluation >0,00 for white when I force Re5 and let them
: play.

: I cannot give you sendinfo because these win95 programs do not
: generate sendinfo-data.
: But I wrote the times and branches and evaluations down manually.

: These TIMES show, that it would make sense to compute on Re5 because
: the BREAKTHROUGH in changing the opinion came for Fritz4 on my
: Pentium120 within 27 seconds!
: This would have changed Fritz decision and would not have lost much
: time....your statistics about HOW LONG it would take and that it will
: not be efficient do not work anymore, because the programs do NOT pure
: alpha-beta but prune much away and show many strange behaviours.

: Please take into your consideration that the mail you post in US comes
: with a special delay onto my usenet server.
: I don't know why but I have very often quoted data by european
: posters, and later I download the US-original.
: I have seen you analysis and will - of course - take a look.
: Also I posted the search times and the evals of fritz4.01.
: But you did not commented on this. WHY?

I did comment. I explained that the only way Fritz can change it's
mind on Re5, is because of pre-analysis done at the root altering
the piece/square tables or whatever else he does... Because it did
not search Re5 any deeper than it did f4 according to your analysis.
Therefore, something else happened to make (suddenly) Re5 look better,
that something had to be something altered at the root. I've seen it
happen to very old versions of crafty that did some root analysis. Now
everything is done at the tips, and it doesn't happen any more...

[Robert Hyatt]

chrisw (chr...@cpsoft.demon.co.uk) wrote:

: --
: http://www.demon.co.uk/oxford-soft

: Robert Hyatt <hy...@crafty.cis.uab.edu> wrote in article
: <5kv97q$r...@juniper.cis.uab.edu>...

: > mclane (mcl...@prima.ruhr.de) wrote:
: > : hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:
: >

: > : > Why post a CSTal game when it is subject to the

Maybe. If you use PVS, and if the first move you search is the best one,
you don't get anything at all, because all will be flagges as < alpha,
where alpha is constant for all... If you change your mind a few times,
or if you search with different alpha/beta values, you might discover
something useful...

: Chris Whittington

[Robert Hyatt]

mclane (mcl...@prima.ruhr.de) wrote:
: Komputer Korner <kor...@netcom.ca> wrote:

: > Since all empirical evidence shows that pondering is in the 50% range
: >on average, then 1 move pondering is the optimum strategy.

: I don't believe that chess programs, especially when they play against
: humans, have pondering rates of 50 %.

Here's three games against Mark Diesen vs Crafty:

moves 51, predicted 35.
moves 89, predicted 41.
moves 33, predicted 24.

I added this recently, but don't have many games yet. But those are
pretty good numbers against a pretty good IM player. I have more if
you'd like.

: >The inkompetent komputer.

: That would also imply that chess is more based on binaric-trees ...

: No - 50 % is to high, they have maybe 30% or less.

Again, I have data, not speculation...

[Robert Hyatt]

mclane (mcl...@prima.ruhr.de) wrote:
: Peter Osterlund <peter.o...@mailbox.swipnet.se> wrote:

: >Your claim is that in this game at the position you describe, it would
: >have been better to ponder both Re5 and f4. This is correct because the

: >move actually played was Re5, so pondering that move would have saved some
: >time on the computers clock.

: Although Bob is absolutely right: it would have not changed the game.
: But there could be different positions where the right pondering can
: help much.

: >Hyatt claims that in order to save as much time as possible (on the
: >average), you should only ponder the one move your opponent is most likely
: >to play. This is also correct.

: But WHAT is the one move ?! How can Hyatt know before ???
: That is the reason HIS argumentation is RIGHT WHEN the ponder move
: comes. But how often is that the case ?

for the 4th time, typically 50% of the time, or more. for the 4th time,
supported by real data collected by a real program, playing against real
IM/GM/computer programs, for the 4th time.

: >You then try to conclude that Hyatt's claim is wrong because your own

: >claim is correct, and this is where your logic fails. Do you suggest that
: >we should use a pondering strategy that on average is much worse, just
: >because in certain cases it would be slightly better?

: No - my point is that hyatt-ideas, statictics, maths, knowledge about
: alpha-beta is all correct, but does not help the main problem that in
: a chess-game there are often MORE than ONE good move.

I have *never* disagreed with you here. Not once. Just that using an
alpha/beta search, doing what you suggest is simply inefficient, and would
reduce the overall strength of a program, because it reduces the available
time per move. That is *all* I've said. Your point about pondering the
wrong move is correct... but at present, it's an unsolvable problem.

: If you ponder on ONE you reduce the chances that this move is chosen.

: I know that ALL programs do it the way that it is decribed by Hyatt. I
: also know the reasons WHY this is done.
: Still this does not satisfy me, because I often see that programs
: ponder wrong moves because the progams do use the main-line of the
: last move, take the 2nd ply and MOVE it meanwhile computing in the
: background and waiting for keyboard.

then we agree about everything I believe... except that doing what you
suggest, within the framework of alpha/beta, is really not doable.

: This will not work when the position is balanced and opponent can move

[vince]

fwiw, 2 rhetorical questions:
1. if higher plies generate less return, why not search alternate
variations after plies reach a certain level of diminishing returns?

2. suppose there are 3 nearly equal variations? the chances could be 34%,
33%, 32% that they would be chosen by the opponent. wouldnt it be best to
search and hash all three?

vince

[mclane]

hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

OK - yo win. I accept that you win if the data shows this.
BUT: I would say this works for each program with a different
%, do you confirm to this ??

But still I think my point is clear: I want to make PB better . And I
would not do it your way. I would do it other way.
Also there IS/WAS one chess dedicated chess computer from Saitek, i
will tell you the name and you can prove by asking the programmer.
I still think that you are ignorant to say: it cannot be what I claim
that cannot be. I have seen this machine and tested it.
Your data cannot be tell us anything to this event than SPECULATIONS.
Right ?

[mclane]

"chrisw" <chr...@cpsoft.demon.co.uk> wrote:

>No, no. I listen to the trees. Bob thinks it good to explain to trees.
>That's the difference :)

>Chris Whittington

In the shervood forrest we have many trees. And Guy Guibert and the
Sheriff of Nottingham will never get us. We will rob the rich and give
it to the poor ones. Richard Lionheard is our king, and no one will
stop us from reunion the country and free it from the aristrocrats.

[mclane]

hy...@crafty.cis.uab.edu (Robert Hyatt) wrote:

I have *never* disagreed with you here. Not once. Just that using an
alpha/beta search, doing what you suggest is simply inefficient, and

could

reduce the overall strength of a program, because it reduces the
available
time per move. That is *all* I've said. Your point about pondering
the
wrong move is correct... but at present, it's an unsolvable problem.

Mclane:
Can we come to the point, together, that maybe I was provoked to
insists so strongly because of the way you have done it ?
And can we come together both trying to get data with several programs
on pondering? And come later with more ideas, results and that stuff
and take a break ?! Because Today I got 2 new versions of CSTal and
have to play with them instead of doing usenet.

Hyatt:

then we agree about everything I believe... except that doing what you
suggest, within the framework of alpha/beta, is really not doable.

Mclane:
Ok, agreed then.

[Ingo Althoefer]

Bob Hyatt wrote:
> chrisw (chr...@cpsoft.demon.co.uk) wrote:
> : ...

> : You can get an better than random ordering of the remaining moves out of
> : alpha beta.
> : Its not perfect, and in some cases its well wrong, but its not wholly
> : random for pool (b).

> Maybe. If you use PVS, and if the first move you search is the best one,
> you don't get anything at all, because all will be flagges as < alpha,

That is not true. In this case you may use information given by the node-count:
Let x be the best move (searched first) and y_1, y_2, ..., y_k be the inferior
alternatives. Let n_1, n_2, ..., n_k be the amount of work needed ( = number
of nodes searched ) to refute y_1, y_2, ..., y_k, respectively. Then in the
next iteration these candidates should be searched in decreasing order of
the n_i.

( Philosophy behind it: the more "expensive" it is to refute a move the higher
is the chance that this move will show up superior in the next iteration.
So this candidate should be searched earlier in the next iteration because
positivelywise it will give better bounds for the candidates that are
searched subsequently. )

> where alpha is constant for all... If you change your mind a few times,
> or if you search with different alpha/beta values, you might discover
> something useful...

Ingo Althoefer.