Author: Moritz Berger
Date: 10:25:01 04/10/98
Go up one level in this thread
I will try to respond to a few issues Mats described in his posts below.
First, I want to thank him to have taken so much time and effort for his
article (including the games always helps to get a feeling for what
you're talking about, thanks again Mats).
[qoute]
Posted by Mats Winther on April 09, 1998 at 07:43:40:
I haven't followed the discussion but I was very suprised
when Fritz5 took the lead in SSDF's list (I had expected
more from MCPro7) since I have made several testgames when
testing my own variations (on a 686/200+ with 40m/2h)
and Fritz5 mostly plays ridiculous chess.
[endquote]
The Cyrix CPU might be responsible (or not) for certain oddities with
Fritz. I remember several complaints about crashes from Cyrix M5 users
on r.g.c.c. some months ago. Another think to invite trouble is running
Qemm products with Fritz 5 - they don't seem to be mutually compatible.
[quote]
This disappoints
me much since these games are useless as theoretical basis.
From this I conclude that Fritz5 may be overrated in comparision
to humans (but not necessarily in comparision to computers)
[endqoute]
Well said.
[quote]
since humans have a special skill in refuting positionally
bad play. Computers lack this skill. Fritz5 benefits from this
since he (more or less) simply skips the positional play - the
other computers can't refute the play anyway (not always the case,
of course). Instead Fritz5 counts longer variations and this is
useful against other computers.
[endqoute]
I don't think that it is that easy. Just compare the "mean and lean" 54
KB chess engine of the King 2.54 with Fritz' 105 KB "twice as much
knowledge" engine. Both programs play good chess. I know that some of
you will not agree with this quantitative approach (the Hiarcs 6 engine
is BTW about 200 KB big), but chess knowledge (whatever that might be)
has to be coded somewhere and counting engine sizes gives at least some
metrics to compare "amount of knowledge". Nodes per second (NPS) is
another quantum, useful when trying to compare "expensive" (i.e. higher
order) and "inexpensive" (e.g. root node only evaluated, preprocessed)
knowledge.
The essence of this argument IMHO is that there can't be *THAT* much a
difference in chess knowledge between engines that are of the same
magnitude in size and NPS (e.g. Fritz does only about 2-2.5 times as
many NPS as Rebel, even if NPS are not always comparable, but they are
within the same oder of magnitude).
My conclusion should be plausible also when observing the fact that all
computers tend to favour the same "typcial computer moves".
So for me, they are basically no "dumb" and "clever" programs. Most
notably in the endgame, knowledge becomes more important, but even there
e.g. Fritz doesn't perform too bad.
[qoute]
I am worried that this programming strategy may damage the
development of computer chess since other programmers will
concentrate on big hash tables and fast calculating instead of
developing the humanlike positional qualities.
[endquote]
Big words ... In the current CSS issue you will read the Kasparov
favours the Fritz 5 engine over Hiarcs for analysis (which he had
previously recommended as being better than the rest of the pack).
Even Fritz 3 scored well in human tournaments, you surely remember the
infamous "IM norm" on P90 hardware in 40/120 tournaments in 1994
(1993?).
[quote]
Actually, there
is very much to do in this field. The programs are still rather
dumb. The stupidity of Fritz5 damages computer chess since
the Fritz program is the only program that many chess players come
into contact with.
[endqoute]
Please differentiate: Fritz 4 is a different animal than Fritz 5. Fritz
5 OTOH on a P90 with 1MB hash tables is *much* weaker than Fritz 5 on my
P233MMX with 98MB hash tables and Powerbook.
[quote]
After playing against it they conclude that
it is boring because of its dumbness and then they loose interest
in this business not knowing that the other programs often produce
marvelous chess. I know that it works like this having talked
to people down at the club.
[endqoute]
Against humans, Rebel is still the best program *by far*. AEGON results
as well as my personal experiences seem to support this statement. But
Fritz is no slouch either. My father in law used to beat the hell out of
all Fritzes on my P166. This immediately stopped when I upgraded to a
P233MMX and Fritz 5. He has not been able to score a single draw ever
since (while having a comfortable 40% score vs. Hiarcs6/DOS on the same
machine).
[quote]
Here are two examples (among many) of lousy play by Fritz5 in variations
investigated by me which I wanted to test. But the
results are useless as theoretical basis:
[Event "test game1"]
[Site "?"]
[Date "1998.03.29"]
[White "THiarcs6"]
[Black "TFritz5"]
[Result "1-0"]
1. e4 d6 2. d4 Nf6 3. Nc3 c6 4. f4 g6 5. Nf3 d5 6. e5 Nh5
{
1. r n b q k b . r
2. p p . . p p . p
3. . . p . . . p .
4. . . . p P . . n
5. . . . P . P . .
6. . . N . . N . .
7. P P P . . . P P
8. R . B Q K B . R
A B C D E F G H
Fritz5 rates this position as completely equal (0.00) which
is ridiculous since black has no compensation for the space
deficit. Nevertheless black's position is very stable and he
can place the knight (not the bishop) on g7, play e6 and place
the bishop on e7. Then he can play on the queen wing. But
Fritz5 manages to loose this position in a couple of moves (on a
686/200+) - an amazing feat!}
7. Bd3 Bg7 8. f5 c5 9. dxc5 d4 10. Bb5+ Bd7 11. Qxd4 Nc6
12. Qe4 Bxf5 13. Bxc6+ bxc6 14. Qxc6+ {with a winning position}
1-0
[endquote]
********* BIG problem
I have tried to reproduce this evaluation and move with my copy of
Fritz5.
The trouble is: My Fritz doesn't chose neither 6...Nh6 nor 7.Bg7. This
makes me wonder if there's some problem with your hardware (maybe it's
also due to the fact that I used the Fritz 5.01 engine from the Junior
CD???).
Here's the game from the same starting position (move 7, white to move
as first calculated move by Hiarcs). One *very nice* game no human 2560
ELO player would need to be ashamed of. Please note Hiarcs' evaluation
at move 36, Hiarcs is tactically faster in this position than Fritz!
(Hiarcs: -2.60, Fritz at next move -0.88).
{I just see that Fritz dropped some evaluations when converting the game
to PGN. I will be happy to send the CBV file to everybody interested in
the complete evals. Eval for move 7.Bf5 was: 0.13/12 after 281 seconds,
Hiarcs eval after 7.Bd3: 0.67/10 after 286 seconds}
Fritz scored a convincing win with the strong novelty 7...Bf5!
FERNANDO: Please replay and comment on this game!
[Event "test game"]
[Site "?"]
[Date "1998.03.29"]
[Round "?"]
[White "THiarcs6"]
[Black "TFritz5"]
[Result "0-1"]
[ECO "B06"]
[PlyCount "90"]
1. e4 d6 2. d4 Nf6 3. Nc3 c6 4. f4 g6 5. Nf3 d5 6. e5 Nh5
7. Bd3 Bf5 8. O-O Bxd3 {(e6) -0.16/13 278} 9. Qxd3 Qd7 {(e6) 0.03/12
276} 10.
Bd2 {(Rb1) 0.67/9 314} 10... Na6 11. Ne2 Nc7 {(e6) 0.00/11 202} 12. a4 {
(Rae1) 0.77/9 207} 12... Ng7 {(Bh6) -0.03/12 212} 13. Qb3 {(Nh4) 0.68/9
276}
13... Rb8 {(0-0-0) 0.00/12 211} 14. Ba5 {(Bb4) 0.77/9 182} 14... Nce6 {
(e6) -0.03/13 251} 15. Ng5 {(Nh4) 0.82/8 166} 15... Nxg5 {(c5) -0.03/13
266}
16. fxg5 Ne6 {(b6) -0.06/12 188} 17. Qf3 Nxg5 18. Qf4 Qf5 19. Qxf5 gxf5
20.
Rxf5 Rg8 21. h4 {(Bd2) 0.50/9 98} 21... Ne4 {(b6) -0.03/12 242} 22. Raf1
b5 23.
axb5 Rxb5 24. b4 {(Bc3) 0.16/10 290} 24... Nd2 {(f6) -0.16/12 230} 25.
Ra1 Bh6
26. g3 {(Kf2) -0.20/9 161} 26... Nc4 27. Kh1 {(Kf2) -0.23/10 330} 27...
Bd2 {
(Be3) -0.19/11 198} 28. c3 h6 29. Bc7 {(Kh2) -0.09/10 159} 29... Rb7 30.
Ba5
Rg4 31. Rf3 f6 {(Re4) -0.31/11 207} 32. exf6 exf6 33. Raf1 Re4 {
(Re7) -0.44/11 125} 34. Nf4 Ne3 35. Nh5 Nxf1 36. Nxf6+ Ke7 37. Ng8+ Kd6
38.
Rxf1 Bxc3 39. Nxh6 Ke6 40. Nf5 Rf7 {(Re1) -2.34/11 29} 41. Rc1 Kxf5 42.
Rxc3
Kg4 43. Kg2 {(h5) -4.43/10 1060} 43... Re2+ 44. Kg1 Ref2 {(Rfe7)
-4.19/12 35}
45. Rc1 Kh3 {adjud. -5.72/14 162} 0-1
Here's the other game Mats' quoted:
[quote]
[Event "test game2"]
[Site "?"]
[Date "1998.03.22"]
[White "THiarcs6"]
[Black "TFritz5"]
[Result "1-0"]
1. d4 d5 2. c4 dxc4 3. Nf3 Nf6 4. Nc3 a6 5. e4 b5 6. e5 Nd5
7. a4 Nxc3 8. bxc3 Qd5 9. Be2 Bb7 10. 0-0 e6 11. Ne1 Qd7
12. Bf3 Nc6 13. Nc2 Rb8 14. Qe2 Be7 15. Rd1 0-0 {theory ends}
16. Bf4
{
1. . r . . . r k .
2. . b p q b p p p
3. p . n . p . . .
4. . p . . P . . .
5. P . p P . B . .
6. . . P . . B . .
7. . . N . Q P P P
8. R . . R . . K .
A B C D E F G H
}
16. ...Na5? ( 16. ...b4 {The King instead suggests this plausible
continuation which leads to equality. The King obviously has a
much better understanding of the position although its rating is lower.}
17. cxb4 Nxb4 18. Bxb7 Rxb7 19. Qxc4 Nd5 )
17. axb5 Bxf3 18. Qxf3 Rxb5 19. Ne3 Rfb8 20. Ra4 Bf8 21. h4 c6
22. h5 R8b7 23. h6 Qc7 24. hxg7 Bxg7 25. Qg4 Kh8 26. Bg5 Rb8
27. Qh4 Qb6 28. Be7 Rc8 29. Kh2 c5 30. Bf6 Qc6 31. Ng4 Rg8
32. Rxa5 Rbb8 33. dxc5 Rbf8 34. Rd6 Qe4 35. Rd4 Qg6 36. Rxa6 h5
37. c6 Rc8 38. Bxg7+ Rxg7 39. Rd8+
1-0
[endqoute]
My "version" (a beautiful struggle!):
[Event "test game1"]
[Site "?"]
[Date "1998.03.22"]
[Round "?"]
[White "THiarcs6"]
[Black "TFritz5"]
[Result "1/2-1/2"]
[ECO "D24"]
[Annotator "Berger,M"]
[PlyCount "121"]
1. d4 d5 2. c4 dxc4 3. Nf3 Nf6 4. Nc3 a6 5. e4 b5 6. e5 Nd5 7. a4 Nxc3
8. bxc3
Qd5 9. Be2 Bb7 10. O-O e6 11. Ne1 Qd7 12. Bf3 Nc6 13. Nc2 Rb8 14. Qe2
Be7 15.
Rd1 O-O {theory ends} 16. Bf4 Na5 17. axb5 Bxf3 18. Qxf3 Rxb5 {
(Qxb5) -0.53/13 376} 19. Ne3 {(Qg4) -0.25/9 153} 19... Rfb8 20. Ra4 {
(Qg4) -0.21/8 137} 20... Bf8 {(c5) -0.69/11 239} 21. h4 Kh8 {(Be7)
-0.56/12 370
} 22. h5 {(Kh2) 0.14/8 140} 22... h6 23. Nxc4 Nxc4 24. Rxc4 R8b6 {
(a5) -0.09/12 352} 25. Qa8 {(Ra1) 0.48/9 156} 25... Kg8 26. Ra1 Qe7 {
(a5) -0.09/12 337} 27. Qe4 {(g3) 0.42/9 288} 27... Qh4 {(c6) -0.19/12
190} 28.
Rxc7 {(g3) 0.55/8 217} 28... f5 {(Rb2) 0.00/12 334} 29. exf6 Rxh5 30.
Kf1 Rb2 {
(Qh1+) 0.09/11 114} 31. Qxe6+ Kh7 32. Qe4+ Kg8 33. f7+ Kh8 34. g3 Qh1+
35. Qxh1
Rxh1+ 36. Kg2 Rxa1 37. Rc8 Kh7 38. Rxf8 Kg6 39. d5 {(c4) -0.15/10 305}
39...
Rd1 {(Raa2) -0.06/13 634} 40. d6 Rb7 41. Ra8 {(c4) 0.28/11 311} 41...
Kxf7 42.
Rxa6 g5 43. Be3 {(Be5) 0.00/11 323} 43... g4 44. Ra4 h5 45. Rd4 {
(Rb4) -0.03/9 82} 45... Rxd4 46. Bxd4 Rb2 {(Ke6) -0.69/14 150} 47. Bc5
Ke6 48.
Bb4 Kd7 49. Ba3 {(Kg1) -0.21/11 49} 49... Ra2 {(Rd2) -0.50/16 215} 50.
Bb4 Ra1
{(Rc2) -0.63/15 132} 51. f4 {(Bc5) 0.00/10 78} 51... Ra2+ {(Ke6)
-0.78/14 127}
52. Kf1 {(Kg1) 0.00/10 64} 52... Kc6 {(Ra1+) -0.66/14 167} 53. Kg1 Rd2 {
(Kd7) -0.88/15 249} 54. Kh1 {(Kf1) 0.00/10 124} 54... Rc2 {(Ra2)
-0.84/14 150}
55. Kg1 Re2 {(Rd2) -0.84/15 199} 56. Kf1 Rd2 57. Kg1 Kd7 {(Re2) -0.72/14
166}
58. Kh1 {(Kf1) 0.00/10 73} 58... Ke6 {(Rd1+) -0.69/15 245} 59. Kg1 Rc2 {
(Kd7) -0.75/14 173} 60. Kf1 Rb2 61. Bc5 {Draw accepted adjud. 0.00/11
156}
1/2-1/2
[quote]
Subject: Re: The Fritz5 discussion
Posted by Mats Winther on April 10, 1998 at 04:56:26:
In Reply to: Re: The Fritz5 discussion posted by Scott Carmichael on
April 09, 1998 at 11:10:43:
Thanks for investigating my results, but I am amazed to see that
some people think I cheated the results.
[endqoute]
I hope you didn't get the impression that I wouldn't take your
statements serious. As you see from this long reply, quite the contrary
is true.
[quote]
I'm glad to see that at
least Fernando understood what I was talking about. Even ChessBase
GmbH admits that Fritz is rather stupid but instead fast.
[endqoute]
But they don't claim that it isn't strong ;-))
[quote]
And they say that it is suitable as an analysing module, and they
may be right. But I tried to show that it is not necessarily suitable
for engine-engine games.
[endqoute]
Please note that I played the games with 98304 KB hash tables for Fritz
and 32768 KB hash tables for Hiarcs. Also, Fritz really runs best on
Intel CPUs (e.g. for Fritz 5, the AMD K6/200 is about 25%-30% slower
than an Intel P5MMX/200).
[quote]
I also stated that most other programs
are more fun to play against. How can anybody think differently
having played against the other top programs? Or do grown up adults
fall in love with a chess program and cannot stand it being critizised?
[endqoute]
The fun levels of Fritz are really fun. The tree is very nice for
opening studies. I get 3 extremely strong chess engines (Fritz, Hiarcs,
Junior). I'm not complaining...
[quote]
I was invited to a Fritz5 discussion by e-mail and I did not expect
cry-babies to be around. Let us have a serious discussion between
adults.
[endquote]
Please ...
[quote]
The conditions for the two test games were these: Cyrix-686/200
with 48 Mb RAM. I used engine-engine in Fritz5. Since the Fritz
module starts swapping if more than half of the RAM is used for
hash (something I don't understand and cannot fix) I set the
hash sizes for each module to 11328 Mb. This is consequently
about the maximum size on my computer (remember that most PC
owners seldom have more than 32 Mb RAM). The clock settings for
the first game was 120´/40+60´/20+30´. The clock settings for
the second game was 90´/30+60´/20+30´ (consequently about three
minutes per move in both cases). No Power Books were used.
[endquote]
I'm curious: The move 6...Nh5 is not in my power book. Where did you get
this move from? A novelty of your own? Fritz really hates this move ;-)
[quote]
The thing is that Fritz5 plays completely horrible chess in both
these instances. I know that the hash size is too small for Fritz
but a human with 2560 would never play like this. Such bad play
should not occur with a top program even with small hash. Or does
Fritz5 play badly only with this hardware configuration? I am
bewildered since so many get different results.
Mats Winther
[endquote]
A fair statement. Please don't underestimate the difficulty all computer
programs have with certain types of positions which are quite common in
human master chess. The idea of starting many computer-computer matches
from the same starting position is quite interesting, John Nunn and Dirk
Frickenschmidt are 2 notable proponents of this way of testing.
Please continue your contributions here on CCC! (and maybe try to be
even more diplomatic in the future? ;-))
Moritz
This page took 0.01 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.