Why don't people use computers when annotating games? Flawed analysis

Date: 18 Oct 2008 06:05:41
From: raylopez99
Subject: Why don't people use computers when annotating games? Flawed analysis

I'm looking at an analysis of the recent third game of the Kramnik -
Anand game on Chessbase (http://www.chessbase.com/newsdetail.asp?
newsid=4963) and am shocked that the annotator apparently did not use
chess software when analyzing the game.

Why? In this day and age there's no excuse for not using chess
playing software. Computers have shown that they are superior to
humans playing chess.

For example, a simple and fast analysis (five seconds per move) shows
the critical moves in the third game as follows (and nobody seems to
have caught these):

19. Nxd4?! (loses). Instead, 19. Rxd4! = . This bad move probably
lost the game for Kramnik. Yet apologists assume that Kramnik's
losing move of 19. Nxd4?! was a 'planned' sacrifice ("19.Nxd4. What a
shot, Kramnik chooses to sacrifice a piece himself and attacks the
rook on g4."). Yeah right. Like his one move blunder a while ago
against a computer that lost a won game was also a 'planned
sacrifice'? Not. Simply put, Kramnik overlooked 19...h5.

32. f3? (loses worse). Instead, 32. Rd3, while still losing, offers
greater resistance for White.

Ironically, the annotator at Chessbase misses these moves (or does not
understand them, see above). Instead the annotator claims that other
moves were at fault, when they clearly were not. For example, it is
claimed "27.a4? With time ebbing away Kramnik errs. The silicon
consensus is 27.Rc1"... in fact, 27. a4 is the preferred move for
Chessbase (Fritz engine) over 27. Rc1. Likewise, the annotator frets
over "33. Bd3?" but in fact at that point the game is over anyway.

Lesson: please lern to use a computer. It's really the only way to
play chess. Even commentators such as Wlod on this board apparently
don't use PCs. I've heard some grandmasters don't use PCs in their
analysis, feeling that it somehow diminishes from the game, but these
eccentric types should be far and few between, in light of the
advances made in chess playing software over the last generation.

RL

Date: 29 Oct 2008 08:34:33
From: raylopez99
Subject: Re: Why don't people use computers when annotating games? Flawed

On Oct 27, 2:00=A0am, Martin Brown <

Date:
From: Martin Brown
Subject: Re: Why don't people use computers when annotating games? Flawed

Date: 24 Oct 2008 07:11:03
From: raylopez99
Subject: Re: Why don't people use computers when annotating games? Flawed

On Oct 24, 2:37=A0am, Martin Brown <

Date: 26 Oct 2008 11:11:33
From: David Richerby
Subject: Re: Why don't people use computers when annotating games? Flawed

raylopez99 <[email protected] > wrote:
> Correct? I think so. Check and mate. Imagine that. I, a non-
> specialist, de novo figured out something that the best minds in the
> business took decades to solve.

No. You appear, again, to be assuming that the whole tree is stored
in memory, which it isn't. The tree never even exists, except in the
recursive execution pattern of the search function.

Imagine you have a pencil, two pads of paper and a chess board. The
first pad of paper is going to be your hash table; the second pad is
your temporary working space. This is how an exhaustive N-ply search
works.

1. You see position on the chess board.
2. If N=0, then
2.1. Evaluate the position you see in front of you.
2.2. Write in the first pad, `Position X evaluates to Y at 0 ply.'
2.3. Go back to the previous level of the search and look at the
next move.
3. Look at the first pad. If it already says that the evaluation of
the current position to N ply is X then
3.1. Go back to the previous level of the search and look at the
next move.
4. In the second pad, write down the list of all legal moves in the
position.
5. Try out each move in turn as follows, keeping track of what the
best move was.
5.1. Make the move on the board
5.2. Turn to a new page of the second pad
5.3. Go to step 1 and do a search to depth N-1
5.4. Erase everything you wrote on the new page of the second pad
5.5. Turn back one page in the second pad
5.6. Cross out the move you just tried
6. Write in the first pad, `Position X evaluates to Y at N ply', where
Y is the score of the best move you found during step 5. Go
back to the previous level of the search and look at the next
move.

After a while, the first pad will get full. When this happens, tear
out the first page (the one that was written on the longest time ago)
and throw it away. Add a new page to the back of the pad.

Notice that you never actually draw a tree in either of your pads of
paper. It's the same with chess search algorithms: they never
explicitly construct the tree so you can never say `Look through the
tree of evaluated positions for a promising leaf.'

Now, there are two types of pruning. The sort that alpha-beta does
happens by inserting the following line:

5.7. If the move you just found is already good enough to refute
the position on the board at stage 1, go on to step 6
without trying any of the other moves.

(And modifying step 6 to write `Position X evaluates to Y (or maybe
even worse) at N ply.')

Selective search is done by modifying step 4 so that, instead of
writing down all legal moves, you just write down the ones that look
interesting.

Move ordering is the art of arranging the list of moves generated in
step 4 so that you search the most promising ones first so that step
5.7 kicks in earlier.

Dave.

--
David Richerby Happy Tree (TM): it's like a tree that
www.chiark.greenend.org.uk/~davidr/ makes your troubles melt away!

Date:
From: Martin Brown
Subject: Re: Why don't people use computers when annotating games? Flawed

Date: 23 Oct 2008 12:33:20
From: raylopez99
Subject: Re: Why don't people use computers when annotating games? Flawed

On Oct 23, 9:02=A0am, David Richerby <[email protected] >
wrote:
> raylopez99 <[email protected]> wrote:
> > From memory, here is how I recall (leaving aside details like "null
> > move", cache, transpositions and hashing, opening book, contempt
> > factor and other such esoteric fluffy stuff) chess engines work,
> > past and present:
>
> > 1/ =A0search to N ply using alpha beta, depth first to N ply to get a
> > 'candidate' that all other nodes/branches of the chess tree are
> > compared with [...]
>
> > 1.2/ order your best moves at the top, using not just mechanical
> > ordering (i.e., 1 rook =3D3D 5 pawns) but some intelligence as well [..=
.]
>
> > 2/ =A0once you exhaustively search the chess tree using 1/, you can pic=
k
> > 'promising lines' to evaluate to M ply using alpha beta. Repeat 1 and
> > 1.2 on these 'promising lines' to M ply.
>
> > Now, based on the above, how close am I to the way: =A0(1) modern chess
> > engines work, or, (2) ancient chess engines work, or, has this above
> > method (in its rough outline) ever been used?
>
> Not very, not very and I don't know, respectively.
>

OK, thanks, at least I'm getting somewhere.

> Modern engines use iterative deepening: search (by alpha-beta with as
> many knobs on as you can make work) to depth 1, then 2, then 3,
> then... until you decide that you've used enough time; then play the
> best move.
>
> Alpha-beta is a depth-first search algorithm so, to minimize the
> breadth of the tree that gets searched, you want to somehow search the
> best move first. =A0This is what move-ordering is. =A0Of course, if you
> had some way of always determining the best move in each position, you
> wouldn't actually need to search; therefore, you end up using various
> heuristics. =A0A good heuristic to start with is just to look at
> material balance or soem other simplified evaluation function.
>

Well that make sense, since by definition if somehow you find the best
move using move ordering on the first pass, like you imply the alpha-
beta algorithm will quickly prune the bad moves more quickly than
otherwise.

> The point of iterative deepening is this. =A0Suppose you're three ply
> into a ten-ply search. =A0What you want to do is evaluate the current
> position to a depth of seven more ply and the evaluation to six ply
> will be a pretty good approximation to that (assuming you actually
> evaluated a quiescent position). =A0There's a good chance that the
> evaluation to six ply is sitting in the hash table because you worked
> it out at the last level of iterative deepening. =A0So you can use that
> as your heuristic for move ordering, where it's available.

OK, now I think you're getting into the problem of "transposition" of
positions/ move order, which is more detail than I want, but it's
interesting.

>
> Your suggested approach seems to be, essentially, a coarse-grained
> version of iterative deepening. =A0Instead of incrementally searching
> one ply, two ply, ..., you first search N ply and then M. =A0But how do
> you work out N and M? =A0

N, M are arbitrarily chosen, depending on your hardware. If you have
Deeper Blue than N, M can be large.

> One advantage of iterative deepening is that
> you can just stop at any time and know that the best move you've found
> is the best up to the depth you've searched fully and still pretty
> good at the depth you've maybe only searched partially. =A0

OK, now a disconnect. Since this forum is not an online learning
forum, feel free to ignore me, but it seems when you say "searched
partially" you are implying step 2/ above in my post. This is my
point. If you have not exhaustively searched the entire tree, then
how exactly do you know if your "iterative deepening" is working? I
posit (and you say it's not close to being the way things work) that
you pick, using rules, the most "promising" initial positions at the
end of your event horizon (which I define as those nodes/branches of
the chess tree at the very bottom of the chess tree portion that you
have just exhaustively searched). That is, at the bottom of the tree
(from the starting node, which is the top), you have a bunch of leaves/
branches that let's suppose are all quiet positions. Which do you
pick as the starting point for a partial search of the tree for
another M plys? You pick the ones that look "promising" according to
your move ordering algorithm. But you imply that you can just pick
any arbitrary node, because the way chess engines work, you are going
to search the whole tree anyway for the next M plys.

But, as I type this, I see a synthesis of our "competing" views that
offers a way out. Let's just say that if your 'move ordering'
algorithm is done right, it _always_ picks the 'most promising' move,
at any node. So thus both our viewpoints are accommodated. I would
call the 'partial search' a 'selective search' while you call it
'iterative deepening'. It raises the question that a 'partial search'
or 'iterative deepening' can, because the chess tree portion is not
exhaustively searched, give a 'false positive' that will eliminate a
best move, but since now we're moving away from the original thread
I'll leave it at that.

In any event, if the preceding paragraph is correct, then essentially
we've been arguing over nothing because our lingo was incorrect. My
apologies.

> But what
> happens if you decide you need to move now, half way through your
> higher-depth search? =A0You're essentially only confident in your move
> up to depth N.

That's right, see "false positive" above. I can see where iterative
deepening can actually (by mistake) eliminate your best move.

RL

Date: 26 Oct 2008 10:36:19
From: David Richerby
Subject: Re: Why don't people use computers when annotating games? Flawed

raylopez99 <[email protected] > wrote:
> David Richerby <[email protected]> wrote:
>> The point of iterative deepening is this. Suppose you're three ply
>> into a ten-ply search. What you want to do is evaluate the current
>> position to a depth of seven more ply and the evaluation to six ply
>> will be a pretty good approximation to that (assuming you actually
>> evaluated a quiescent position). There's a good chance that the
>> evaluation to six ply is sitting in the hash table because you
>> worked it out at the last level of iterative deepening. So you can
>> use that as your heuristic for move ordering, where it's available.
>
> OK, now I think you're getting into the problem of "transposition"
> of positions/ move order, which is more detail than I want, but it's
> interesting.

I'm not talking about move order but about move ordering. Move order
is a matter of classical chess analysis: things like black playing
1.d4 Nf6 2.c4 e6 in the hope that white will play 3.Nf3 to avoid the
Nimzo-Indian, after which black can respond 3... d5, entering a QGD
where white has already committed to Nf3.

Move ordering is a matter for alpha-beta-like chess algorithms.
Alpha-beta and its friends don't search the entire tree but they
produce the same result as searching the entire tree. The reasoning
is as follows. Suppose that, in a particular level position, you can
play 22.Nd4, leading to another level position. As an alternative,
you can play 22.Nb4, to which black has a bunch of possible responses.
22... f5 leads to a level position but 22... Bxb4 wins the knight
outright. At this point, you don't need to do any more analysis on
22.Nb4. It might be that black can also play 22... Rxd2, winning your
queen but you don't care -- you already know that playing 22.Nb4 loses
at least a knight, while 22.Nd4 maintains equality.

But you wasted time considering 22... f5 as a possible refutation of
22.Nb4. Indeed, it would have been better to consider 22... Rxd2
first, as `you lose your queen' is a more powerful refutation than
`you lose your knight'. Move ordering is a series of heuristics
designed to get the right move considered first, to avoid wasting time
on searching branches of the tree that don't contribute to the final
evaluation.

>> Your suggested approach seems to be, essentially, a coarse-grained
>> version of iterative deepening. Instead of incrementally searching
>> one ply, two ply, ..., you first search N ply and then M. But how
>> do you work out N and M?
>
> N, M are arbitrarily chosen, depending on your hardware. If you
> have Deeper Blue than N, M can be large.

That doesn't answer the question because it doesn't consider the
clock. If I have five minutes left, how do I chose N and M to
guarantee that the search finishes quickly enough that I don't lose on
time now or in the next few moves?

>> One advantage of iterative deepening is that you can just stop at
>> any time and know that the best move you've found is the best up to
>> the depth you've searched fully and still pretty good at the depth
>> you've maybe only searched partially.
>
> OK, now a disconnect. Since this forum is not an online learning
> forum, feel free to ignore me, but it seems when you say "searched
> partially" you are implying step 2/ above in my post.

No. When I say `searched partially', I mean that you started to do a
search to some depth and then you realised that you didn't have enough
time to finish it.

> This is my point. If you have not exhaustively searched the entire
> tree, then how exactly do you know if your "iterative deepening" is
> working? I posit (and you say it's not close to being the way
> things work) that you pick, using rules, the most "promising"
> initial positions at the end of your event horizon (which I define
> as those nodes/branches of the chess tree at the very bottom of the
> chess tree portion that you have just exhaustively searched). [...]

I think this points to a fundamental misunderstanding. You appear to
be assuming that the whole tree (or, perhaps, some kind of pruned
tree) exists in memory. It doesn't. It's too big.

> But, as I type this, I see a synthesis of our "competing" views that
> offers a way out. Let's just say that if your 'move ordering'
> algorithm is done right, it _always_ picks the 'most promising' move,
> at any node. So thus both our viewpoints are accommodated. I would
> call the 'partial search' a 'selective search' while you call it
> 'iterative deepening'.

No. Iterative deepening involves using something like alpha-beta at
higher and higher search depths. Alpha-beta is guaranteed to return
the same result as an exhaustive search of the tree. Selective search
involves only looking at parts of the tree in a way that means you
might miss the best move (because it didn't look promising, initially)
and, therefore, might misevaluate a position.

> I can see where iterative deepening can actually (by mistake)
> eliminate your best move.

Iterative deepening of alpha-beta can never do this. There are
exactly two reasons for iterative deepening coming up with a bad move:

1. you just didn't search deep enough;
2. your evaluation function misled you somehow at a leaf.

Dave.

--
David Richerby Disposable Evil Hi-Fi (TM): it's like
www.chiark.greenend.org.uk/~davidr/ a music system but it's genuinely evil
and you never have to clean it!

Date: 23 Oct 2008 07:20:04
From: raylopez99
Subject: Re: Why don't people use computers when annotating games? Flawed

On Oct 23, 3:04=A0am, Martin Brown <

Date: 23 Oct 2008 17:02:15
From: David Richerby
Subject: Re: Why don't people use computers when annotating games? Flawed

raylopez99 <[email protected] > wrote:
> From memory, here is how I recall (leaving aside details like "null
> move", cache, transpositions and hashing, opening book, contempt
> factor and other such esoteric fluffy stuff) chess engines work,
> past and present:
>
> 1/ search to N ply using alpha beta, depth first to N ply to get a
> 'candidate' that all other nodes/branches of the chess tree are
> compared with [...]
>
> 1.2/ order your best moves at the top, using not just mechanical
> ordering (i.e., 1 rook =3D 5 pawns) but some intelligence as well [...]
>
> 2/ once you exhaustively search the chess tree using 1/, you can pick
> 'promising lines' to evaluate to M ply using alpha beta. Repeat 1 and
> 1.2 on these 'promising lines' to M ply.
>
> Now, based on the above, how close am I to the way: (1) modern chess
> engines work, or, (2) ancient chess engines work, or, has this above
> method (in its rough outline) ever been used?

Not very, not very and I don't know, respectively.

Modern engines use iterative deepening: search (by alpha-beta with as
many knobs on as you can make work) to depth 1, then 2, then 3,
then... until you decide that you've used enough time; then play the
best move.

Alpha-beta is a depth-first search algorithm so, to minimize the
breadth of the tree that gets searched, you want to somehow search the
best move first. This is what move-ordering is. Of course, if you
had some way of always determining the best move in each position, you
wouldn't actually need to search; therefore, you end up using various
heuristics. A good heuristic to start with is just to look at
material balance or soem other simplified evaluation function.

The point of iterative deepening is this. Suppose you're three ply
into a ten-ply search. What you want to do is evaluate the current
position to a depth of seven more ply and the evaluation to six ply
will be a pretty good approximation to that (assuming you actually
evaluated a quiescent position). There's a good chance that the
evaluation to six ply is sitting in the hash table because you worked
it out at the last level of iterative deepening. So you can use that
as your heuristic for move ordering, where it's available.

Your suggested approach seems to be, essentially, a coarse-grained
version of iterative deepening. Instead of incrementally searching
one ply, two ply, ..., you first search N ply and then M. But how do
you work out N and M? One advantage of iterative deepening is that
you can just stop at any time and know that the best move you've found
is the best up to the depth you've searched fully and still pretty
good at the depth you've maybe only searched partially. But what
happens if you decide you need to move now, half way through your
higher-depth search? You're essentially only confident in your move
up to depth N.

Dave.

--
David Richerby Expensive Impossible Cat (TM): it's
www.chiark.greenend.org.uk/~davidr/ like a cat but it can't exist and
it'll break the bank!

Date: 23 Oct 2008 07:06:18
From: raylopez99
Subject: Re: Why don't people use computers when annotating games? Flawed

On Oct 22, 5:42=A0pm, Simon Krahnke <[email protected] > wrote:

Seems we're talking past each other, because you're accustomed to
different lingo, but insofar as I can tell, I took from this thread
that what you (perhaps) are calling "forward pruning" is what I call
"not searching the entire chess tree exhaustively past a certain
ply". Both of these terms assume that the alpha-beta search algorithm
will be used (as it always is, since otherwise you have to search too
many branches of the chess tree). And both these terms rely on a
function that is going to pick the 'most promising' quiescent move
lines/positions.

In any event, this is my last reply to you since it's not productive
to communicate further. I've noticed this with certain engineers I
work with in my professional capacity--if you don't use the exact
lingo they are accustomed to, they look at you like you're from Mars.
Nevermind that on closer examination it turns out they don't really
understand the lingo itself, but that's another matter.

RL

Date: 26 Oct 2008 11:30:01
From: David Richerby
Subject: Re: Why don't people use computers when annotating games? Flawed

raylopez99 <[email protected] > wrote:
> In any event, this is my last reply to you since it's not productive
> to communicate further. I've noticed this with certain engineers I
> work with in my professional capacity--if you don't use the exact
> lingo they are accustomed to, they look at you like you're from
> Mars. Nevermind that on closer examination it turns out they don't
> really understand the lingo itself, but that's another matter.

It's not a matter of being accustomed to some particular language, as
if there are equally valid alternatives. In this scenario, there is
one language, used by a wide community of people, and another
language, used only by you. You need to learn the community language
in order to communicate effectively with the community. We'll try to
help you with that but there comes a time when it simply isn't worth
the effort to keep remembering that `A piston is a part of a car
engine but Ray calls those ``explosion chamber oscillators'' and uses
the term ``piston'' to mean the brake pedal.'

If you wish to discuss precise, technical things, it is necessary to
use precise, technical language. That way, people know that
misunderstandings are because of misunderstanding of the fundamental
concepts, not just of the words being used to describe them.

Dave.

--
David Richerby Moistened Hi-Fi (TM): it's like a
www.chiark.greenend.org.uk/~davidr/ music system but it's moist!

Date:
From: Martin Brown
Subject: Re: Why don't people use computers when annotating games? Flawed

Date: 22 Oct 2008 12:11:58
From: raylopez99
Subject: Re: Why don't people use computers when annotating games? Flawed

On Oct 22, 10:43=A0am, Simon Krahnke <[email protected] > wrote:
> * raylopez99 <[email protected]> (14:39) schrieb:
>
> > OK, "evaluation function" then. =A0But I believe that "evaluation
> > function" is the key to having a PC search so good at blitz [...]
>
> It is what it makes any good, because the evaluation functions what
> informs the search. That isn't much different in blitz, tournament or
> analysis.

Yes, I agree and understand this, I think.

>
> > --because, like you say, it does a depth first not breath first search,
>
> That's got nothing too do with breadth *first*, alpha-beta always does
> depth first search, it avoids breadth not searching certain subtrees
> that don't promise too be any good (pruning).

Good point, you need to establish a cutoff value by going as deep as
you can then bringing that value back to the root node, I forgot about
that.

>
> > and given that at blitz the PC only has five seconds to evaluate a
> > chess tree that has ~(32)^N positions to search, where N=3D8 or 10 (ie.
> > ply),
>
> In 5 seconds a modern computer spends 12 billion cycles and calculates
> about 2.5 million positions. That's about N=3D4.

Interesting. Back in the early 90s I read a book that implied N=3D5 is
tough, but perhaps that was slow hardware.

>
> > then by necessity you are not going to search even four moves
> > exhaustively in five seconds, but have to rely on the evaluation
> > function to find the best move, I would imagine. At 180 seconds, with
> > a fast PC I'm sure you can do an exhaustive alpha-beta moves up to 4
> > moves (8 ply) deep, but then beyond this, I would imagine again most
> > chess software relies on the evaluation function to 'order' the top 3
> > moves (from which it picks its next move), and to search beyond 4
> > moves if it has time.
>
> The evaluation function doesn't directly order any moves. Alpha-beta
> efficiency depends on early cuts. Thus keeping book of moves that lead
> to cuts helps ordering moves for further early cuts.

Well this is confusing. Something has to order moves, because three
different move sequences (aside from transposition) have to have
different numerical evaluations. For example two bishops are slightly
more powerful than two knights. That's the scoring function. But now
consider this: after 5 moves (10 ply), or maybe 7 moves (14 ply) most
PCs cannot exhaustively evaluate all moves beyond that. So are you
saying they don't even try? Are you saying that modern chess programs
only do "exhaustive" searches of the chess tree, using alpha-beta,
pruning, etc, without going beyond a certain ply? That is news to
me. I assumed they are constantly thinking, constantly digging deeper
into the chess tree. That is, essentially chess programs are always
in "infinite mode" except that if you select five seconds per move,
they will simply give you the top moves (from the "infinite mode"
list) that are present after five seconds. This list of top moves can
change as time progresses (but it usually does not for ~80+% of moves--
that was my original point in this thread). In short, chess programs
don't stop at ply N, but use what I call the 'evaluation function' to
select the 'best' likely move beyond the exhaustive search, and use
alpha-beta and cutoff etc on that line.

Please correct me if I'm wrong.

RL

>
> The evaluation function evaluates positions not moves. It's what tells
> the search how good a position is. It's usually called on the branches
> of the search tree.
>

OK, positions then. A position is a series of moves so to a degree
this is semantics.

RL

Date: 22 Oct 2008 12:02:14
From: raylopez99
Subject: Re: Why don't people use computers when annotating games? Flawed

On Oct 22, 6:07=A0am, Martin Brown <

Date:
From: Martin Brown
Subject: Re: Why don't people use computers when annotating games? Flawed

Date: 22 Oct 2008 05:39:25
From: raylopez99
Subject: Re: Why don't people use computers when annotating games? Flawed

On Oct 21, 6:14=A0pm, Simon Krahnke <[email protected] > wrote:
>
> What you call "move ordering" is the evaluation function. The search
> algorithm is pretty generic without much chess specific things in it. It
> is guided by the results of the evaluation function which evaluates
> positions using chess knowledge.
>
> There's a thing called move ordering in the search algorithm which is
> mostly generic and only chess centric in the manner it treats captures,
> promotions and checks (because these are chess specific concepts).
>
> > This latter component is because a chess
> > engine cannot do a 'brute force' search beyond five or six moves,
>
> A search engine usually tries to search deep, so tries to minimize the
> time spent on searching breadth. So move ordering always uses the
> probably best moves first in the hope that the move is so good it can
> stop searching the current node.

OK, "evaluation function" then. But I believe that "evaluation
function" is the key to having a PC search so good at blitz--because,
like you say, it does a depth first not breath first search, and given
that at blitz the PC only has five seconds to evaluate a chess tree
that has ~(32)^N positions to search, where N=3D8 or 10 (ie. ply), then
by necessity you are not going to search even four moves exhaustively
in five seconds, but have to rely on the evaluation function to find
the best move, I would imagine. At 180 seconds, with a fast PC I'm
sure you can do an exhaustive alpha-beta moves up to 4 moves (8 ply)
deep, but then beyond this, I would imagine again most chess software
relies on the evaluation function to 'order' the top 3 moves (from
which it picks its next move), and to search beyond 4 moves if it has
time.

RL

Date: 21 Oct 2008 12:57:42
From: raylopez99
Subject: Re: Why don't people use computers when annotating games? Flawed

On Oct 21, 7:21=A0am, David Richerby <[email protected] >
wrote:

> Yes but `Which of these World Champions plays most like a crippled
> version of Crafty?' really isn't a sensible way of evaluating
> anything. =A0I

We went over this last time, remember? The thread was well over
several hundred posts long. Can't remember the last time we had such
"unity" on this board! LOL, unless it was something that Sam Sloan
posts.

FYI, I 'respect' your (flawed) opinion and you should at least be
aware of mine: even in blundercheck you can check to see which
masters got 'closest' to the correct move, under the theory that
usually there is but a slight difference between an evaluation at five
seconds a move versus 180 seconds a move. And those times (the
turning points in the game) are not that critical for most games--
since, after all, these guys are all champions. Check it out
yourself, as I have: try evaluating master games at 5 seconds a move
on Fritz then go to 30 seconds a move, then 180 s/m, and see that the
evaluation for most moves does not really change the top three best
lines.

But, as in things mathematics, some people either get it or they never
do. Perhaps you're one of them...

RL

Date: 21 Oct 2008 06:23:36
From:
Subject: Re: Why don't people use computers when annotating games? Flawed

On Oct 20, 7:21=A0pm, raylopez99 <[email protected] > wrote:
> On Oct 19, 2:40=A0pm, [email protected] wrote:
>
> > =A0 It would be an interesting project to take masters who have been
> > considered great analysts =97 not necessarily the same thing as a great
> > player =97 e.g. Marco, Alapin, Capablanca, Alekhine, R=E9ti, Nimzovitch=
,
> > Botvinnik, Geller, et al, and subject good-size samples of their work
> > to computerized inspection, to see who had the least tendency to
> > error.
>
> You've probably seen this already, but they have done this for
> championship games--

That's not what I was talking about. I was referring to analytical
works =97 tournament and match books, best game collections, magazine
articles etc. =97 in which these men annotated games, either their own
or others'. Examples would be Alekhine's books on New York 1924 and
Nottingham 1936, Capablanca's "My Chess Career," R=E9ti's "Modern Ideas
in Chess," Lasker's Manual of Chess, Steinitz's annotations for
Hastings 1895 or his "Modern Chess Instructor," Marco's books on
Barmen 1905, Ostend 1906, and the Lasker-Tarrasch match, etc. As
opposed to how well a master performs under pressure in a game, I'm
interested in how accurately, objectively, and insightfully he can
dissect a game when analyzing at leisure.

> as opposed to 'good-size samples'-- and this group
> discussed this a while ago in a long thread (I started the thread):http:/=
/en.wikipedia.org/wiki/Comparing_top_chess_players_throughout_h...
> (see below)
>
> RL
>
> Actual moves played compared with computer choices
>
> One of the latest methods of analyzing chess abilities across history
> has come from Matej Guid and Ivan Bratko from the Department of
> Computer and Information Science of Ljubljana University.[19] The
> basis for their evaluation was the difference between the position
> values resulting from the moves played by the human chess player and
> the moves chosen as best by a chess program, Crafty. They also
> compared the average number of errors in the player's game. According
> to their analysis, the leader was Jos=E9 Ra=FAl Capablanca, followed
> closely by Vladimir Kramnik.
> The "Classical" World Chess Championship matches were analyzed, and
> the results for the fourteen Classical World Champions were presented.
> Players with fewest average errors:
> Jos=E9 Ra=FAl Capablanca
> Vladimir Kramnik
> Anatoly Karpov
> Garry Kasparov
> Boris Spassky
> Tigran Petrosian
> Emmanuel Lasker
> Bobby Fischer
> Alexander Alekhine
> Vassily Smyslov
> Mikhail Tal
> Mikhail Botvinnik
> Max Euwe
> Wilhelm Steinitz

Date: 21 Oct 2008 02:45:18
From: raylopez99
Subject: Re: Why don't people use computers when annotating games? Flawed

Offramp wrote:

> Which move? Do you mean the stunning 10...Bd6 after 8....Bd6?

Sorry, that should read 10...Be6

RL

Date: 21 Oct 2008 00:35:06
From: Offramp
Subject: Re: Why don't people use computers when annotating games? Flawed

On Oct 21, 12:18 am, raylopez99 <[email protected] > wrote:
> On Oct 19, 2:40 pm, [email protected] wrote:

> As for a computer finding new moves, this is brought out in a book on
> openings by Neil McDonald ("Concise Chess Openings" --I love this book
> because it's so small), to wit, on the McDonnell Gambit of the King's
> Gambit, McDonald says that Fritz found this move: 1.e4 e5 2.f4 exf4
> 3.Nf3 g5 4.Bc4 g4 5. Nc3 gxf3 6.Qxf3 d5 7.Nd5 Nc6 8.O-O Bd6 9.d4 Nxd4
> 10.Qh5 Bd6 11.Bxf4 Bxf4 12.Nxf4 Bxc4? 13.Qe5+ wins rook, but here
> Fritz found 12...Nf3+!!, which refutes the variation.
>
> I defy a human to find such a counterintuitive move.

Which move? Do you mean the stunning 10...Bd6 after 8....Bd6?

Date: 21 Oct 2008 00:30:54
From: Offramp
Subject: Re: Why don't people use computers when annotating games? Flawed

On Oct 21, 12:21 am, raylopez99 <[email protected] > wrote:
> On Oct 19, 2:40 pm, [email protected] wrote:

> The "Classical" World Chess Championship matches were analyzed, and
> the results for the fourteen Classical World Champions were presented.
> Players with fewest average errors:
> Jos=E9 Ra=FAl Capablanca
> Vladimir Kramnik
> Anatoly Karpov
> Garry Kasparov
> Boris Spassky

I can only imagine Spassky being in 5th if the first sentence had
been,
"The "Classical" World Chess Championship matches were analyzed, and
the results for the fourteen Classical World Champions OVER ALL THE
GAMES IN THEIR CAREERS were presented."
That would then include the thousands and thousands and thousands and
thousands and thousands and thousands and thousands of very quick
draws that litter Spassky's career after 1977.

Date: 20 Oct 2008 16:21:12
From: raylopez99
Subject: Re: Why don't people use computers when annotating games? Flawed

On Oct 19, 2:40=A0pm, [email protected] wrote:

> =A0 It would be an interesting project to take masters who have been
> considered great analysts =97 not necessarily the same thing as a great
> player =97 e.g. Marco, Alapin, Capablanca, Alekhine, R=E9ti, Nimzovitch,
> Botvinnik, Geller, et al, and subject good-size samples of their work
> to computerized inspection, to see who had the least tendency to
> error.

You've probably seen this already, but they have done this for
championship games--as opposed to 'good-size samples'-- and this group
discussed this a while ago in a long thread (I started the thread):
http://en.wikipedia.org/wiki/Comparing_top_chess_players_throughout_history
(see below)

RL

Actual moves played compared with computer choices

One of the latest methods of analyzing chess abilities across history
has come from Matej Guid and Ivan Bratko from the Department of
Computer and Information Science of Ljubljana University.[19] The
basis for their evaluation was the difference between the position
values resulting from the moves played by the human chess player and
the moves chosen as best by a chess program, Crafty. They also
compared the average number of errors in the player's game. According
to their analysis, the leader was Jos=E9 Ra=FAl Capablanca, followed
closely by Vladimir Kramnik.
The "Classical" World Chess Championship matches were analyzed, and
the results for the fourteen Classical World Champions were presented.
Players with fewest average errors:
Jos=E9 Ra=FAl Capablanca
Vladimir Kramnik
Anatoly Karpov
Garry Kasparov
Boris Spassky
Tigran Petrosian
Emmanuel Lasker
Bobby Fischer
Alexander Alekhine
Vassily Smyslov
Mikhail Tal
Mikhail Botvinnik
Max Euwe
Wilhelm Steinitz

Date: 21 Oct 2008 15:21:29
From: David Richerby
Subject: Re: Why don't people use computers when annotating games? Flawed

raylopez99 <[email protected] > wrote:
> One of the latest methods of analyzing chess abilities across
> history has come from Matej Guid and Ivan Bratko from the Department
> of Computer and Information Science of Ljubljana University.[19] The
> basis for their evaluation was the difference between the position
> values resulting from the moves played by the human chess player and
> the moves chosen as best by a chess program, Crafty.

Yes but `Which of these World Champions plays most like a crippled
version of Crafty?' really isn't a sensible way of evaluating
anything. If, say, Kasparov were to play a match against Crafty, he
would surely win it convinicingly. Yet Guid and Bratko would look at
the games and, every time Kasparov played a winning move, they would
say `That move was bad because Crafty was expecting something else.'

Dave.

--
David Richerby Chocolate Painting (TM): it's like a
www.chiark.greenend.org.uk/~davidr/ Renaissance masterpiece that's made
of chocolate!

Date: 20 Oct 2008 16:18:04
From: raylopez99
Subject: Re: Why don't people use computers when annotating games? Flawed

On Oct 19, 2:40=A0pm, [email protected] wrote:
> Tal's annotations
> of his 1960 match with Botvinnik were also quite good, especially
> considering that he dictated them verbally without sight of any board,
> replaying the games in his head from memory.

That is wicked! Tal is one "bad" chess player!

I agree with Taylor Kingston here, and point out that a lot of people
don't realize that a chess engine has two components: a brute force
portion, and a "move ordering" portion that picks the most likely
moves to search first. This latter component is because a chess
engine cannot do a 'brute force' search beyond five or six moves, so,
it must order the 'most likely best moves' beyond move five (actually
more like move 4 for most PCs). Thus, there is real "artificial
intelligence" involved in picking the likeliest 'best moves'. Typical
candidates for 'best moves' are the usual ones from the rules of
chess: pick moves that support other pieces (like two knights
reinforcing each other), rooks behind pawns, rooks in open files (or
soon to be open), move the king in the endgame not in the beginning
(when out of book), etc. Chess engines, using these rules and others,
are surprisingly "human like"--Fruit, designed by a chess player, is
one example.

As for a computer finding new moves, this is brought out in a book on
openings by Neil McDonald ("Concise Chess Openings" --I love this book
because it's so small), to wit, on the McDonnell Gambit of the King's
Gambit, McDonald says that Fritz found this move: 1.e4 e5 2.f4 exf4
3.Nf3 g5 4.Bc4 g4 5. Nc3 gxf3 6.Qxf3 d5 7.Nd5 Nc6 8.O-O Bd6 9.d4 Nxd4
10.Qh5 Bd6 11.Bxf4 Bxf4 12.Nxf4 Bxc4? 13.Qe5+ wins rook, but here
Fritz found 12...Nf3+!!, which refutes the variation.

I defy a human to find such a counterintuitive move.

RL

Date: 19 Oct 2008 14:40:05
From:
Subject: Re: Why don't people use computers when annotating games? Flawed

On Oct 19, 4:40=A0pm, Offramp <[email protected] > wrote:
> On Oct 19, 3:05 pm, [email protected] wrote:
>
> > "Lasker's Manual of Chess," published by Russell Enterprises, which
> > will be out in a few weeks.
>
> In that go-hick I believe that Lasker's book of the 1909 St Petersburg
> tournament, the one where he says the analysis "is accurate" has held
> up unbelievably well. I remember having a quick look at some of his
> analysis with a reasonable chess program and it all stood up well.

I reviewed an e-book version of that some years ago:

http://www.chesscafe.com/text/review312.pdf

On the whole Lasker's analysis there did stand up quite well to
computer analysis. The same is not true of his "Common Sense In
Chess," where I discovered quite a few errors in editing a recent new
edition.(http://www.chesscafe.com/text/review625.pdf).

> I read somewhere that the analyst whose analysis stands up best to
> computer analysis is ... Steinitz.

I couldn't say; I've never put him under silicon-based scrutiny.
Among those I have, I'd say Reuben Fine was surprisingly inaccurate,
while Botvinnik and/or Fischer was perhaps the best. Tal's annotations
of his 1960 match with Botvinnik were also quite good, especially
considering that he dictated them verbally without sight of any board,
replaying the games in his head from memory. Editing Russell
Enterprise's 5th edition of that book, I found one or two whoppers,
and a few lesser errors, but overall it was amazingly clean. A
surprisingly error-prone book was Timman's on Cura=E7ao 1962 (http://
www.chesscafe.com/text/review495.pdf).
It would be an interesting project to take masters who have been
considered great analysts =97 not necessarily the same thing as a great
player =97 e.g. Marco, Alapin, Capablanca, Alekhine, R=E9ti, Nimzovitch,
Botvinnik, Geller, et al, and subject good-size samples of their work
to computerized inspection, to see who had the least tendency to
error.

Date: 19 Oct 2008 13:40:50
From: Offramp
Subject: Re: Why don't people use computers when annotating games? Flawed

On Oct 19, 3:05 pm, [email protected] wrote:

> "Lasker's Manual of Chess," published by Russell Enterprises, which
> will be out in a few weeks.

In that go-hick I believe that Lasker's book of the 1909 St Petersburg
tournament, the one where he says the analysis "is accurate" has held
up unbelievably well. I remember having a quick look at some of his
analysis with a reasonable chess program and it all stood up well.

I read somewhere that the analyst whose analysis stands up best to
computer analysis is ... Steinitz.

Date: 19 Oct 2008 13:19:30
From:
Subject: Re: Why don't people use computers when annotating games? Flawed

On Oct 19, 2:15=A0pm, [email protected] (John Savard)
wrote:
> On Sun, 19 Oct 2008 07:05:55 -0700 (PDT), [email protected]
> wrote, in part:
>
> > =A0A strange attitude. I would be very reluctant to produce any game
> >collection without computer analysis today.
>
> True, but I can see what is behind it. Tactical phenomena that only a
> computer would be able to detect could be claimed to be _irrelevant_ to
> human over-the-board play.

I can't accept that argument at all. Good moves are good, and bad
moves are bad, whether discovered by human brains or computer
programs. It's an annotator's job to point out both, with the greatest
accuracy possible.

> I, too, think it makes more sense to use the
> computers - but take what they say with the appropriate grain of salt,
> in the fashion you've noted.
>
> John Savardhttp://www.quadibloc.com/index.html

Date: 19 Oct 2008 07:05:55
From:
Subject: Re: Why don't people use computers when annotating games? Flawed

On Oct 19, 9:03=A0am, Offramp <[email protected] > wrote:
> On Oct 18, 2:05 pm, raylopez99 <[email protected]> wrote:
>
> > I'm looking at an analysis of the recent third game of the Kramnik -
> > Anand game on Chessbase (http://www.chessbase.com/newsdetail.asp?
> > newsid=3D4963) and am shocked that the annotator apparently did not use
> > chess software when analyzing the game.
>
> The recent Keene-Simpole book Petrosian v the Elite says that they
> specifically did not use a computer when analysing the games.
>
> "We used our own brains for our comments and cannot blame a computer
> program - which we rarely employed - for any blunders which have
> persisted." Page 13.

A strange attitude. I would be very reluctant to produce any game
collection without computer analysis today. Even the great world
champions erred in their analysis. I recently edited a new edition of
"Lasker's Manual of Chess," published by Russell Enterprises, which
will be out in a few weeks. There were enough significant analytical
errors, of both omission and commission, that I added a 24-page
appendix of analytical endnotes to correct them.
If an all-time great like Lasker can make that many mistakes, how
much more a lesser player like Keene, whose analytical style already
tends to the superficial?

> I think computers are still a bit shaky when it comes to very long or
> deep sacrifices.

They are shaky in the sense that they may not see the value of a
long-term sacrifice, especially if the compensation is positional
rather than tactical, and so they won't make the sac in the first
place. A human GM, using general principles, will sense the potential
of the sacrifice, whereas for the computer the ultimate payoff will be
beyond its analytical horizon, and so it will not consider the sac
worthwhile. However, once the sac is made, the computer will often
handle the resulting variations more accurately than the GM.

Date: 19 Oct 2008 06:03:11
From: Offramp
Subject: Re: Why don't people use computers when annotating games? Flawed

On Oct 18, 2:05 pm, raylopez99 <[email protected] > wrote:

> I'm looking at an analysis of the recent third game of the Kramnik -
> Anand game on Chessbase (http://www.chessbase.com/newsdetail.asp?
> newsid=4963) and am shocked that the annotator apparently did not use
> chess software when analyzing the game.

The recent Keene-Simpole book Petrosian v the Elite says that they
specifically did not use a computer when analysing the games.

"We used our own brains for our comments and cannot blame a computer
program - which we rarely employed - for any blunders which have
persisted." Page 13.

I think computers are still a bit shaky when it comes to very long or
deep sacrifices.

I agree with you that they should be used, though.

Date: 19 Oct 2008 00:30:32
From: Sanny
Subject: Re: Why don't people use computers when annotating games? Flawed

GetClub Easy Level plays as good as International Master.

So It need a IM to beat Easy Level at GetClub.

GetClub uses applet which is 3-4 times slower than downloaded programs
Still it plays very strong games.

So want a game with IM then play with Easy Level at GetClub.

Bye
Sanny

Play Chess at: http://www.GetClub.com/Chess.html

Date: 18 Oct 2008 07:17:53
From: SAT W-7
Subject: Re: Why don't people use computers when annotating games? Flawed

A very interesting read , thanks ..Do this for all the games ...My webtv
sucks so i can not watch the games live ..I think the 4th game is
starting now Germany time ....