In article .com,
help bot wrote:
To sum up: the chess algorithms used for Cray Blitz
were specifically written and optimized to exploit the
killer advantage of its speed and power. They were
not vastly-superior-chess-algorithms which just
happened to have been written for a very fast machine,
whose vast superiority was merely an added bonus.

There *are* no "vastly-superior-chess-algorithms" out
there. Apart from Sanny, we all use alpha-beta with assorted
heuristics and other tweaks. Further, until commercial interest
raised its head, most of us happily swapped our algorithms, so
that any bright ideas [such as "null move"] soon spread to all
active implementations. But you seem to be missing the point
that the algorithms for CB were not *just* exploiting the extra
horsepower of the computer, but were also exploiting potential
for parallel programming. This was simply not possible on the
normal minis and micros of the period; so it would have been
very stupid to attempt to run CB on a Vax. But it would have
been just as stupid to run code written for the Vax on a Cray.

[...] In fact,
at tournament speeds (which is how these
programs got rated) the addition of even two
plys back then should have given the Cray a
big edge in head-to-head competition.

Such as perhaps being able to win the World Championship?

[...]. So you think that if we took the one guy
who had access to a Cray supercomputer and
"switch him" with the absolute best chess programmer
on the PC, there might well be no improvement
whatsoever?

I don't think anyone has suggested that. But there would
have been a very steep learning curve for both programmers, and
no great expectation that a PC expert would be equally expert on
a Cray or vv.

That's amazing. It's akin to suggesting
that if I were to jump out of my car and let A.J.Foyt
get in and drive, he would not get to McDonald's any
sooner than I would driving his race car, because we
think differently?

I know nothing about A. J. Foyt, but I'm not at all sure
that I would trust Michael Schumacher to drive my car. AJF/MS
have at least the advantage that they [presumably] also drive
normal cars, whereas I [and presumably you] have no experience
at all of driving a F1 racing car. You should perhaps also
note that RMH may not be the world's absolute best programmer
on a PC, but Crafty, despite being "open source" and thus
available as a base platform for everyone else to use and to
learn from, has maintained a position among leading programs
over more than a decade, so he is certainly no slouch.

--
Andy Walker, School of MathSci., Univ. of Nott'm, UK.

"Martin Brown" wrote in message
ups.com...
On May 7, 12:09 pm, "Chess One" wrote:
"Chess One" wrote in message

news:P9m%h.1960$s7.772@trndny05...

Martin, there is one other point here to which I can add a note which is
a
specific about playing chess rather than generalities about play or the
components aspects of the program which may or may not be useful

- I set up a Russian Chess journalist [a GM] to annotate his own game [in
our series, this mean he had to annotate a loss] and provide below his
own
comments - there is a certain level of play where the engine doesn't see
enough ahead and in fact at this level of play are not taken too
seriously
/because/ as you will see in this game tactical considerations count for
very little, and a deep positional move seems to win the game, and in
fact
calls into question the entire line - here the program is not used as
much
for its computational ability, but for its database of possibilities -
this
GM criticises himself for over-reliance on using the db, even though he
plays out various lines against his computer - he is playing inferior
lines.

Which engine is he using though? Fritz8 lacks discrimination in this
sort of position. I don't have Fritz9 or X.

He didn't mention it specifically, but Fritz is a good guess - though Rybka
is Russian, it wasn't much about at the time.

All the engines I have tried on this position prefer the natural Be6
above all else (as does powerbooks2006). Nc6 is also playable.

Yes - that is a sort of book line,

Rybka
briefly preferred the novelty Be7 over all other options by ply 20 it
is second to Be6.

What this example might illustrate for OUR conversation is if any chess
engine can find what Bezgogov's could not. Here is the key text:

"13. Nh4 . I underestimated this plan in my preparation. White's knight
arrives at the very strong g6-square and its almost impossible to drive
it
away from this station. The knight distinctly complicates Black's
defense.

Shredder10 sees this at ply 17. It might have seen it at ply 16 but I
was on the phone.
It clearly fears the knight move as it has also put Be7 in the running
(AFAIK a novelty).

Ply (scores in cp)
17 18 19 Line

22 18 24 10. ... Be6 11. Ng5 Nd7 12. Bf4 Nb6 13. a4
76 61 72 10. ... Nc6 11. Bxc4 Be6 12. Bxe6 fxe6 13. Rb1
80 85 71 10. ... Be7 11. Bxc4 Be6 12. Bxe6 fxe6 13. Ng5
104 92 96 10. ... h6 11. 11. Bxc4 Be6 12. Bxe6 fxe6 13. Nh4
96 92 103 10. ... Ke8 11. Bxc4 Be7 12. Ng5 Bxg5 13. Bxg5

hese would have been viable alternatives, though did it find even 10. h6?!
at all. If you insert h6, what sort of evaluation does it give compared to
other 10th moves above?

I must admit one more serious mistake - I played this endgame against one
of
the strongest chess programs. My 'inanimate' opponent didn't find
White's
13th move and, against any other play Black stands well and I would
defend
Black's position easily."

(( entire game and commentary is
athttp://www.chessville.com/LessonsLearned/2004Jan.htm))

some other notes from the game:-

"GM Alexei Bezgodov: My loss to an excellent grandmaster from Moscow,
Evgeny Najer, one can say consists of two parts. First I prepared badly
for
the game using an incomplete database and already from the opening got a
very difficult (maybe even lost) endgame."

A critical moment comes here
"10...h6?!

He is saying, to summarise the above, that his 10 ... h6 becomes refuted
by
Whites 13. Nh4. And that, Martin, is where I think a current test might
be
useful. Here are the preceeding moves:-

(1) Najer,E - Bezgodov,A [D24]
56th Russian Championship (3.13), 05.09.2003

1.d4 d5 2.c4 dxc4

3.Nf3 Nf6 4.Nc3! c5 5. d5

5...e6 6.e4 exd5 7.e5! 7...Ne4 8.Qxd5 8...Nxc3

9.Qxd8+ Kxd8 10.bxc3

10...h6?! ((Lets call this Alpha Point)

11.Bxc4 11....Be6 12.Bxe6 fxe6 13.Nh4! ((Lets call this Beta Point))

---

A) My assumptions are that (a) a program will play the Alpha point move
at
9, or score it as a strong consideration if it (b) does not see the Beta
point move at 13.

Most engines prefer to avoid losing the pawn on c4 so h6 is well down
the list.

yes - but this is not quite to the point since without Whites Kt manoever at
13, h6 is very comfortable for black

Shredder would probably see this at tournament rates of play.
It also thinks that 13. ... c4 or Nc6 would be better than Be7.

B) Can your chess engine (a) find the Beta point move on its own? and (b)
how does it score that move in you hand-enter it, compared with scores
for
other White options at 13?

Nh4 is about 20-25cp ahead of all other options in Shredder10. Then
Rb1 or Be3.

If you do not force the move on it, does it value it less than other moves?
Does it even see the move?

The position is effective at detecting engines that can be blinded by
swapoff sequences - like Fritz8. Both Shredder10 and Rybka2.3.1 have
no real difficulty in seeing 13. Rh4 as a threat and offer other
alternatives to the line played.

Its too abstract. Specifically here we have a specific positional option
which the computer can see at either Alpha or Beta point, and does it do
either? This is the 'learning problem' or in other words, insufficient
evaluation capability of the engine, not described in general chess
principals, or programming principles, but in specifics of actual play - and
that is the test!

If it can't do either, what does this mean?

Phil Innes

Regards,
Martin Brown

PS I wonder if Google will swallow this without trace....

09.05.2007 01:58, Chess One:
though Rybka is Russian

Although the main distributer is originally from Russia (now officially
located in the UK), the author is not Russian. Actually it's a Czech,
born in the US, currently living in Hungary.

Greetings,
Ralf

On May 8, 1:23 pm, (Dr A. N. Walker) wrote:

You have some problem with the notion that development
of Cray Blitz stopped when the Cray became unavailable, and that
RMH then wrote Crafty for PCs?

No. Where I have a problem is in twin, conflicting
claims by BH and others regarding the connection
between these two programs. In short, I want such
claims to be both: a) true and b) consistent with
one another. (If one cannot *always* tell the truth,
then at least have the courtesy to lie in a consistent
manner!)


Crafty is not CB. People move on. So do programs.


People keep claiming that Cray Blitz is and is not
related to Crafty, and they seem to want to have it
both ways, depending on whim.



What do you propose? To uninvent the other ways in
which programs can learn?

No. But we might consider eliminating everyone
who knows about them, and destroy the evidence.


These days, programs that don't
learn just get smashed on the servers. But neural networks
do not seem to be particularly successful at the sorts of
learning that chess needs. It's not for want of trying.
Temporal differencing, eg, seems to do much better.

Why am I reminded of the ancient debate about
brute-force vs. selective search?


Mr.
Hyatt, the creator of Crafty, simply exploited the raw
speed of a mainframe he had (virtually unique) access to.
Actually, it was Larry Nelson who exploited the raw speed,

[Yes, "Harry", sorry!]

Ah, Larry Nelson was, I believe, one of the two
astronaut characters in "I Dream of Jeannie". You
probably forgot, since *the girl* is all that remains of
that TV show in your dreams, too. :D



and Bob Hyatt and Bert Gower who wrote the chess and the algorithms.
Wrote the chess? What does that mean?

It means that a complete chess program includes a lot of
different aspects -- the user interface, chess knowledge, tree-
searching methods, use of heuristics, .... And exploitation of
particular hardware. This was especially true back then, when
there was a factor of 20+ between a generic C/Fortran/whatever
program and code tuned to a particular machine, even more so when
there were vector instructions available. HN, according to RMH,
knew nothing at all about chess,

Hmm... Could HM be a code name for Sanny,
I wonder?


not even the moves, nor about
the search algorithms and heuristics being used; but he was an
expert on the Cray.


Well, IMO this means that the other guys should
take the fall when CB drops the ball (unless there
is a hardware failure).


[...] The thing is, how can you
(rationally) take credit for awesome programming
skill, summarily dismiss the brute power of the
Cray, and yet watch the competition close the
gap in your rear-view mirror when they are
driving ordinary cars and you're in a Porsche?

Who has "dismissed" the "brute power" of the Cray?


Well, in the old newsgroup discussions, Mr. Hyatt
did this when he summarily dismissed criticisms
based on the Cray having a massive speed advantage
(as opposed to, say, the superior chess knowledge).


And not many of the competition were in "ordinary cars".
Belle may have been a PDP 11/23, but it was coupled to the
best special-purpose hardware that Ken Thompson could lay
his hands on.

Good as this may have been, the fact remains that
Cray Blitz had an easy draw for the taking via Bxh6.
In short, Belle left its King wide open, ala Deep Blue
vs. GM Kasparov, only far, far simpler.


Hitech, and so on, were just as "brute" as
Cray Blitz. You really have to wait until long after the
time of CB before "ordinary" competition got close.

Apparently, all we had to "wait for" was one of those
pesky times when, during a USCF-rated tournament,
there was a problem obtaining the usual, massive
resource allocation BH had come to expect. That's
when the whining started, and I am guessing it had
to do with a competitive failure like the game vs.
Belle (and the much more recent tourney in 2005)
where, again, the excuses sprang up. In short,
whenever anything goes wrong it is blamed on the
lack of what I will call speed, for brevity. Yet when
things go well, it's not "merely" the speed. See an
inconsistency here?


Even
now, the top programs are not winning the world championship
or beating Kramnik on the cheapest PCs from "PC World" --
there's a hint in names like "*Deep* Junior/Fritz/etc".


That's primarily because they have the money to
buy really fast computers, but also because,
without them, the results would be something to
avoid. Why arrange a match if you're just going
to lose?


The real issue was the fact that despite vastly
superior hardware, the Crays and Deep Blues
were unable to maintain a substantive lead for any
length of time over programmers who chose to
work their skills on commonplace hardware.

*No-one* has managed to maintain a substantive lead
for any length of time.

So? My point was that without this hardware/
speed advantage, these programmers would almost
certainly have done worse; yet even WITH the killer
hardware, they were unable to maintain the pole
position for very long.

Imagine, if you will, a situation where one man, and
no one else in the world, can defeat every checkers
program in the world. Here is a case where the guy
has bragging rights. But with chess, the leader is
constantly being replaced, and when you factor in
a big hardware advantage, it seems a bit ludicrous
for anyone with such an advantage to try and take
the credit for writing a first-rate program, when in
fact it may be second-rate, for all we know. It's not
like the other contenders can test their programs
against his on a level playing field. All we know is
that: a) when it won, it had a huge speed advantage,
and b) despite continuing to have superior, if not still
overwhelming speed, it lost its position at #1 to a
program or series of programs which ran on far
inferior hardware. This seems to imply that the guys
who overtook the Cray were better in the chess part
of the equation.

It's been so long now that I don't recall what sort of
speed advantage CB had anymore, but as far as I could
tell, this was the key.


A handful of names [eg, Shredder] have stayed
at or near the top for a long time, suggesting that their
programmers are at least pretty competent.

This directly contradicts a statement you made
just a little while ago. Is it any wonder then, that
you feel somehow obliged to "defend" Mr. Hyatt?

You want it both ways, depending on your whims.
When convenient, you say that nobody stays at
the top for long. Then you turn right around and
proclaim that Shredder did it for "a long time".
Enough of your lunacy.

-- help bot

On May 8, 2:20 pm, (Dr A. N. Walker) wrote:

I know nothing about A. J. Foyt, but I'm not at all sure
that I would trust Michael Schumacher to drive my car. AJF/MS
have at least the advantage that they [presumably] also drive
normal cars, whereas I [and presumably you] have no experience
at all of driving a F1 racing car.

How tough can it be -- you just stay on the track
and keep turning left (again and again and again).
I would stay near the grass, so the other drivers
could pass (again and again and again).


You should perhaps also
note that RMH may not be the world's absolute best programmer
on a PC, but Crafty, despite being "open source" and thus
available as a base platform for everyone else to use and to
learn from, has maintained a position among leading programs
over more than a decade, so he is certainly no slouch.

My comments regarding the programming skills
of Bob Hyatt were mainly regarding the program
Cray Blitz, which as I have decided (on a whim),
is *not* Crafty.

-- help bot

On May 9, 12:58 am, "Chess One" wrote:
"Martin Brown" wrote in message
ups.com...

On May 7, 12:09 pm, "Chess One" wrote:
"Chess One" wrote in message
news:P9m%h.1960$s7.772@trndny05...

Which engine is he using though? Fritz8 lacks discrimination in this
sort of position. I don't have Fritz9 or X.

He didn't mention it specifically, but Fritz is a good guess - though Rybka
is Russian, it wasn't much about at the time.

All the engines I have tried on this position prefer the natural Be6
above all else (as does powerbooks2006). Nc6 is also playable.

Yes - that is a sort of book line,

Indeed.

Rybka
briefly preferred the novelty Be7 over all other options by ply 20 it
is second to Be6.

What this example might illustrate for OUR conversation is if any chess
engine can find what Bezgogov's could not. Here is the key text:

"13. Nh4 . I underestimated this plan in my preparation. White's knight
arrives at the very strong g6-square and its almost impossible to drive
it
away from this station. The knight distinctly complicates Black's
defense.

Shredder10 sees this at ply 17. It might have seen it at ply 16 but I
was on the phone.
It clearly fears the knight move as it has also put Be7 in the running
(AFAIK a novelty).

Ply (scores in cp)
17 18 19 Line

22 18 24 10. ... Be6 11. Ng5 Nd7 12. Bf4 Nb6 13. a4
76 61 72 10. ... Nc6 11. Bxc4 Be6 12. Bxe6 fxe6 13. Rb1
80 85 71 10. ... Be7 11. Bxc4 Be6 12. Bxe6 fxe6 13. Ng5
104 92 96 10. ... h6 11. 11. Bxc4 Be6 12. Bxe6 fxe6 13. Nh4
96 92 103 10. ... Ke8 11. Bxc4 Be7 12. Ng5 Bxg5 13. Bxg5

These would have been viable alternatives, though did it find even 10. h6?!

It seems you are not paying attention. Your belief that engines cannot
do these things has blinded you to the fact that it had the entire
best line following 10. ... h6?! as its fourth least bad option in the
table above from ply16 onwards. It also found a slightly better
response for black after 13. Nh4 within about 5 minutes elapsed time.

at all. If you insert h6, what sort of evaluation does it give compared to
other 10th moves above?

See the table above.

"GM Alexei Bezgodov: My loss to an excellent grandmaster from Moscow,
Evgeny Najer, one can say consists of two parts. First I prepared badly
for
the game using an incomplete database and already from the opening got a
very difficult (maybe even lost) endgame."

A critical moment comes here
"10...h6?!

He is saying, to summarise the above, that his 10 ... h6 becomes refuted
by
Whites 13. Nh4. And that, Martin, is where I think a current test

Most engines prefer to avoid losing the pawn on c4 so h6 is well down
the list.

yes - but this is not quite to the point since without Whites Kt manoever at
13, h6 is very comfortable for black

Shredder would probably see this at tournament rates of play.
It also thinks that 13. ... c4 or Nc6 would be better than Be7.

B) Can your chess engine (a) find the Beta point move on its own? and (b)
how does it score that move in you hand-enter it, compared with scores
for
other White options at 13?

Nh4 is about 20-25cp ahead of all other options in Shredder10. Then
Rb1 or Be3.

If you do not force the move on it, does it value it less than other moves?
Does it even see the move?

It sees the entire line from before 10. ... h6 once the search depth
exceeds 16 ply ~4min for Shredder or 13 ply in 15s for Rybka. Both of
these top engines have no trouble at all with this position. Fritz is
blinded as are lesser engines and will probably never see the tricky
line no matter how much time they are given - preferring material gain
or equal swaps.

The position is effective at detecting engines that can be blinded by
swapoff sequences - like Fritz8. Both Shredder10 and Rybka2.3.1 have
no real difficulty in seeing 13. Rh4 as a threat and offer other
alternatives to the line played.

Its too abstract. Specifically here we have a specific positional option
which the computer can see at either Alpha or Beta point, and does it do
either?

Can't you read? Shredder and Rybka both have no difficulty in seeing
the entire line at around 4 mins per move. Rybka definitely has the
edge though seeing it faster and at a 3ply shallower search depth.

This is the 'learning problem' or in other words, insufficient
evaluation capability of the engine, not described in general chess
principals, or programming principles, but in specifics of actual play - and
that is the test!

If it can't do either, what does this mean?

That Fritz and certain other engines are still a bit too
materiallistic. Hint: not all chess engines are Fritz 5.32

Incidentally the original Najer, Bezgodov 2003 game was a lot more
nearly equal than his commentary implies.

Most of the engines think black actually stands better at -0.60 after
30. ... Kc4 or 30. ... Kb3 but it is beyond my analytical abilities to
verify this.

It was the inaccuracy of 31. ... Kb2 that really started his terminal
decline 31. ... Kb3 or 31. ... Kb4 looks drawish. And then 33. ...
Bxe5 effectively seals his fate. Although I do suspect something funny
is going on in blundercheck. I didn't like Rybkas annotation value for
35. ... d2?+- 3.28/14 so I ran it with a clean slate and got instead.

move35 ply 14 15 16 17 18 19 20
.... Re7 18 24 28 50 245 245 320
.... c3 93 97 97 108 121 245 244
.... d2 93 97 103 108 124 245 257

d2 never evaluates at anything like 3.28 and the only values left by
ply 19 are 2.45 and 3.71 for Rybka. Although it is possible that this
reflects the true situation it could also be a quirk of unwanted hash
collisions in the dirty transposition tables. There are some pretty
wild changes in evaluation at ply 19 here. Shredder10 still sees a
range of distinct values. Both prefer ... Re7 out to ply 17.

I guess that and the final disaster of 38. .... Kb4 were made in time
trouble - it looks to me like 38. ... Re1 could hold out for a good
while longer and maybe even draw against slightly inaccurate play by
white.

The Kasparov Anand at Riga 1995 game that helpbot pointed out has far
higher discrimination against weaker engines. So far only Rybka has
got the right answer. Other engines can see the right lines but score
them with wildy incorrect evaluations that lead to them never being
played. Worst case error on this one is 90cp!

NB it isn't always the top line that gets the crazy score in engine
evaluations. It is worth leaving a spectrum of the top 5-10 lines
visible during infinite analysis when looking at awkward test
positions to gain a better insight. Sometimes it is obvious to a human
that a lower scoring line is much better for reasons that the engine
cannot see.

Regards,
Martin Brown

PS Google is dropping my posts on the floor again. Why can't they fix
it?

"Martin Brown" wrote in message
ups.com...

Shredder10 sees this at ply 17. It might have seen it at ply 16 but I
was on the phone.
It clearly fears the knight move as it has also put Be7 in the running
(AFAIK a novelty).

Ply (scores in cp)
17 18 19 Line

22 18 24 10. ... Be6 11. Ng5 Nd7 12. Bf4 Nb6 13. a4
76 61 72 10. ... Nc6 11. Bxc4 Be6 12. Bxe6 fxe6 13. Rb1
80 85 71 10. ... Be7 11. Bxc4 Be6 12. Bxe6 fxe6 13. Ng5
104 92 96 10. ... h6 11. 11. Bxc4 Be6 12. Bxe6 fxe6 13. Nh4
96 92 103 10. ... Ke8 11. Bxc4 Be7 12. Ng5 Bxg5 13. Bxg5

These would have been viable alternatives, though did it find even 10.
h6?!

It seems you are not paying attention. Your belief that engines cannot
do these things has blinded you to the fact that it had the entire
best line following 10. ... h6?! as its fourth least bad option in the
table above from ply16 onwards. It also found a slightly better
response for black after 13. Nh4 within about 5 minutes elapsed time.

I just wanted to be clear if the moves are input, or the prog finds them
itself, then the evaluation value for each line analysed - especially
comparing the Alpha point [move 10] with the Beta point [13.]

11.Ng5 Ke8 12.e6!?÷ Bareev E.- Azmaiparashvili Z., Plovdiv 2003, and the
game was drawn on the 55th move; 11...Be6 12.Bxe6 fxe6 13.Ng5 Kd7 ½-½ Dreev
A. - Azmaiparashvili Z., Dos Hermanas 2001 (38).

at all. If you insert h6, what sort of evaluation does it give compared
to
other 10th moves above?

See the table above.
-------------

If you do not force the move on it, does it value it less than other
moves?
Does it even see the move?

It sees the entire line from before 10. ... h6 once the search depth
exceeds 16 ply ~4min for Shredder or 13 ply in 15s for Rybka. Both of
these top engines have no trouble at all with this position.

Okay, I understand. At 16 or 19 ply, how much /time/ does it take to review
the options you give above?

Fritz is
blinded as are lesser engines and will probably never see the tricky
line no matter how much time they are given - preferring material gain
or equal swaps.

So Fritz is going to have trouble evaluating positively your inovation 10.
.... Ke8 11. Bxc4 Be7 12. Ng5 Bxg5 13. Bxg5

The position is effective at detecting engines that can be blinded by
swapoff sequences - like Fritz8. Both Shredder10 and Rybka2.3.1 have
no real difficulty in seeing 13. Rh4 as a threat and offer other
alternatives to the line played.

Its too abstract. Specifically here we have a specific positional option
which the computer can see at either Alpha or Beta point, and does it do
either?

Can't you read? Shredder and Rybka both have no difficulty in seeing
the entire line at around 4 mins per move. Rybka definitely has the
edge though seeing it faster and at a 3ply shallower search depth.

Good - I just want to be sure we are discussing the same thing - which is
indeed the positional evaluation problem raised in the book - and Fritz is
embarassed by this /specific/ example. I was just trying to avoid
generalisms, since while this seems typical of a blindness in Fritz, is it
really?

From a purely chessic p.o.v. the issue of bishopXknight exchanges seems
indicated in this instance 10. ... Ke8 11. Bxc4 Be7 12. Ng5 Bxg5 13. Bxg5
as if Fritz doesn't want to exchange 13... Bxg5, since it may be giving up
its piece differential of maybe 0.25 points, while Rybka & Shredder can
consider the exchange since it loses less than the positional disadvantage
of not exchanging BxN.

---

That Fritz and certain other engines are still a bit too
materiallistic. Hint: not all chess engines are Fritz 5.32

Incidentally the original Najer, Bezgodov 2003 game was a lot more
nearly equal than his commentary implies.

Most of the engines think black actually stands better at -0.60 after
30. ... Kc4 or 30. ... Kb3 but it is beyond my analytical abilities to
verify this.

yes - this is very interesting analysis - you are good to keep plugging away
on the specific line, and my motive in questioning it rests somewhat in
differing engines piece evaluation and positional evaluation - ie, above I
credited Rybka with more postional sense, but is that true, or does it
actually vary the worth of the B&N?

Did your engine see this:-

29.f6 Kc4 The king hurries to d5 in order to limit White's offensive on the
kingside. But a spectacular bishop sacrifice could face Black with
insoluble problems: 30.Bxh6! gxh6 31.Ke4! (Not impossible is 31.Kf4??
Kd5!-+) 31...Kb3 32.Kf5 Kc2 33.Rh1 d3 34.e6 d2 35.e7 Bxe7 36.f7 Rf8 37.Ke6
Bf6 38.Kxf6+-

It was the inaccuracy of 31. ... Kb2 that really started his terminal
decline 31. ... Kb3 or 31. ... Kb4 looks drawish. And then 33. ...
Bxe5 effectively seals his fate. Although I do suspect something funny
is going on in blundercheck. I didn't like Rybkas annotation value for
35. ... d2?+- 3.28/14 so I ran it with a clean slate and got instead.

move35 ply 14 15 16 17 18 19 20
... Re7 18 24 28 50 245 245 320
... c3 93 97 97 108 121 245 244
... d2 93 97 103 108 124 245 257

d2 never evaluates at anything like 3.28 and the only values left by
ply 19 are 2.45 and 3.71 for Rybka. Although it is possible that this
reflects the true situation it could also be a quirk of unwanted hash
collisions in the dirty transposition tables. There are some pretty
wild changes in evaluation at ply 19 here. Shredder10 still sees a
range of distinct values. Both prefer ... Re7 out to ply 17.

Fascinating!

I guess that and the final disaster of 38. .... Kb4 were made in time
trouble -

the player identified an earlier problem 31.Rc1+ Kb2?

but his comment to 38. ...Kb4 was
"A nervous move which one obviously ought to make. Certainly in my
calculations I hoped on the move 38...Re1 but after an obligatory 39.Rh3+
Kb4 (39...Kc2 40.Rh2+-) 40.Rd3 I did not see the refined maneuver 40...Re7!
(40...d1Q 41.Rxd1 Rxd1+ 42.Ke6 Rd8 43.f6+-) 41.Rxd2! (An untimely pawn's
transformation to a Queen deprives White of his advantage. 41.g8Q Rd7+)
41...Rxg7 42.Rd4+ Kc3 43.Rf4+- However here also White's victory is quite
simple in spite of a material equality."

it looks to me like 38. ... Re1 could hold out for a good
while longer and maybe even draw against slightly inaccurate play by
white.

The Kasparov Anand at Riga 1995 game that helpbot pointed out has far
higher discrimination against weaker engines. So far only Rybka has
got the right answer. Other engines can see the right lines but score
them with wildy incorrect evaluations that lead to them never being
played. Worst case error on this one is 90cp!

NB it isn't always the top line that gets the crazy score in engine
evaluations. It is worth leaving a spectrum of the top 5-10 lines
visible during infinite analysis when looking at awkward test
positions to gain a better insight. Sometimes it is obvious to a human
that a lower scoring line is much better for reasons that the engine
cannot see.

Thank you for continuing these explorations. Quite HOW programs are able to
see or not see lines is an area of great interest. Although Rybka gets much
right, we are all a bit puzzled /how/ it does that - or put the other way,
why other programs cannot do that.

I am personally puzzled by the way programs evaluate piece values as the
game progresses, compared with positional evaluations - examining specific
lines seems to aid the discovery process, and is suggestive of benchmark
ideas [such as the Alpha-Beta one above] to 'put the question' to any
engine.

Its typical to read responses which refer to the process of the software -
and the test of whether that is of the slightest consequence is to look at
specific lines /in chess terms/ not in program machinations. It was very
interesting for example to read your real-time reference above that Rybka
found the line at 16 ply in 4 mins.

And Martin, we sell all Convekta software at Chessville, so I am familiar
with Rybka, but I don't want to turn this into an advertisement as much as
look at the more general question; 'how we know what we know'.

Cordially, Phil Innes

Regards,
Martin Brown

PS Google is dropping my posts on the floor again. Why can't they fix
it?

Usenet has its own and very simple format, which google in all its
complexity, has difficulty with.

In article .com,
help bot wrote:
[...] Where I have a problem is in twin, conflicting
claims by BH and others regarding the connection
between these two programs. [...]
People keep claiming that Cray Blitz is and is not
related to Crafty, and they seem to want to have it
both ways, depending on whim.

Possibly that's because you are asking for a yes/no
answer to a more complex problem? Clearly Crafty is *related*
to CB. RMH himself says on his web site that Crafty is a
"direct descendent" of CB, that it's "derivative". But it's
not a clone, and Bob will have had to throw away much of CB
to get it to work on a PC. *Especially* those elements of CB
that made it distinctively *Cray*; ie, those bits that made
it interestingly different from other programs of the period.
If there are many people who make the mistake of saying that
Crafty owes its success to Cray, then it's not surprising that
Bob gets a bit ticked off and points out that Crafty has
nothing at all to do with Cray [the machine, as opposed to CB,
the program derived from Blitz].

[...] But neural networks
do not seem to be particularly successful at the sorts of
learning that chess needs. It's not for want of trying.
Temporal differencing, eg, seems to do much better.
Why am I reminded of the ancient debate about
brute-force vs. selective search?

Difficult to know; what debate was that? Successful
programs, then and now, used brute force *and* selection. TINA.

[...]
Good as this may have been, the fact remains that
Cray Blitz had an easy draw for the taking via Bxh6.

There are not many bug-free chess programs around ....
We all have war stories to tell.

*No-one* has managed to maintain a substantive lead
for any length of time.
So? My point was that without this hardware/
speed advantage, these programmers would almost
certainly have done worse; yet even WITH the killer
hardware, they were unable to maintain the pole
position for very long.

Not sure it's much of a point. Sure, if CB and DB
had been running on a PDP 11/23, they would have done worse.
Not only because the PDP 11/23 was slower, but also because
the programs weren't written with the PDP 11 architecture in
mind. Play Bach on a modern piano or Chopin on a harpsichord,
and it's rubbish -- not because Bach/Chopin were stupid, but
because the instruments are different. We'll never know what
they might have written had they swapped periods; not even
whether they would still have been great composers.

Meanwhile, note that if CB and DB are no longer top
programs, it owes not a little to the fact that Cray and IBM
stopped supporting chess. It's not as though DB is now
languishing at number 40 or whatever; it's dead, which is a
decent enough reason for not maintaining pole position.

Imagine, if you will, a situation where one man, and
no one else in the world, can defeat every checkers
program in the world. Here is a case where the guy
has bragging rights.

Indeed. And Jon Schaeffer is a pretty bright guy
with lots of innovations to his credit. Nevertheless, you
are talking about a big fish in a small pond. 8x8 checkers
[draughts] is not the most vibrant part of computer games.

But with chess, the leader is
constantly being replaced, and when you factor in
a big hardware advantage, it seems a bit ludicrous
for anyone with such an advantage to try and take
the credit for writing a first-rate program, when in
fact it may be second-rate, for all we know.

Maybe. Get back to us when you have read and
understood the academic papers about CB [and DB/HiTech,
etc]. And while CB may have had a big hardware advantage
over li'l ol' me, it didn't have one over Belle/HiTech
and others. It might also be a little surprising if we
were to find that CB was second-rate but that Bob was
able to up his game when writing Crafty. ...

It's not
like the other contenders can test their programs
against his on a level playing field.

... Whereas we could all test our programs against
Belle, HiTech, DB? It was a different era. I don't think
Bob is so big-headed as to pretend or claim that he is the
world's top programmer, but he did well enough with Blitz
and Crafty to prove that he's not just a hack with access
to fantastic hardware. If you wanted a level playing field,
you could play in the Uniform Platform competition; if you
wanted to avoid the big metal, you could play in the Micro
competitions. All a bit like "Robot Wars".

All we know is
that: a) when it won, it had a huge speed advantage,
and b) despite continuing to have superior, if not still
overwhelming speed, it lost its position at #1 to a
program or series of programs which ran on far
inferior hardware.

Like Belle, HiTech and Deep Blue? Hmm.

This seems to imply that the guys
who overtook the Cray were better in the chess part
of the equation.

In the World Championship, CB won in '83 and '86; the
next winner was Deep Thought in '89. In the ACM events, CB won
in '83 and '84; the next few winners were HiTech, Belle, ChipTest
and Deep Thought. You might well be able to claim that Hans
Berliner was "better in the chess part", but not that any of those
machines were feeble in hardware terms. The big shock was Rebel's
success in the '92 WCCC; but that was 6 years after CB's prime,
and the period when PCs were starting to be useful for serious
computing [and, despite the hype, Rebel/Gideon/ChessMachine wasn't
exactly a mere PC].

A handful of names [eg, Shredder] have stayed
at or near the top for a long time, suggesting that their
programmers are at least pretty competent.
This directly contradicts a statement you made
just a little while ago.

Really? You see a contradiction between "*No-one* has
managed to maintain a substantive lead for any length of time"
and "A handful of names have stayed at or near the top for a
long time"? No distinction between "substantive lead" and "at
or near the top"? Korchnoi has been "at or near the top" for
several decades without ever having a "substantive lead"; and
Botvinnik was "at or near the top" for around 30 years, but had
a "substantive lead" only perhaps for five of those, the rest of
the time being at best "primus inter pares".

[...]
When convenient, you say that nobody stays at
the top for long. Then you turn right around and
proclaim that Shredder did it for "a long time".

If you still think that what I *actually* said was a
contradiction, then perhaps this explains why you are having
problems with what has been said/written about Dr Hyatt, CB
and Crafty.

--
Andy Walker, School of MathSci., Univ. of Nott'm, UK.

In article . com,
help bot wrote:
[...] driving a F1 racing car.
How tough can it be -- you just stay on the track
and keep turning left (again and again and again).
I would stay near the grass, so the other drivers
could pass (again and again and again).

That may or may not be a workable strategy at Indianapolis,
but I strongly advise you not to try it in a F1 racing car. Esp
not at Monaco, where the grass is in short supply [but the shops
and the sea are quite near].

--
Andy Walker, School of MathSci., Univ. of Nott'm, UK.

Dr A. N. Walker wrote:
David Richerby wrote:
But it seems rather unfair to attack their work *without* first
reading it.
True. But they're kind of asking for it by publishing a summary
publicly while only making the full paper available through a
subscription-only journal.

It seems rather unfair to attack *them* because *Chessbase*
publishes a summary

The summary was written by the original authors of the paper.


and even more unfair when that summary includes a link [near the
bottom left] to the full paper, which is available at Chessbase

That was an oversight on my part, for which I apologize.


[free, as you're in the UK] from your nearest decent library.

Excepting that I don't live in the UK.


Dave.

--
David Richerby Gigantic Swiss Pants (TM): it's like
www.chiark.greenend.org.uk/~davidr/ a well-tailored pair of trousers but
it's made in Switzerland and huge!