Reply
 
LinkBack Thread Tools Display Modes
  #1   Report Post  
Old January 18th 08, 11:36 AM posted to rec.games.chess.computer, rec.games.chess.analysis
external usenet poster
 
First recorded activity by ChessBanter: Jul 2006
Posts: 1,015
Default Engine generation of an opening book

I would like to know if there are enough volunteers to help create a
pure engine based openings book ab initio.

I have tested a couple of engines so far with some curious results.
The act of running the CB interface in "infinite analysis mode" with
20 lines being evaluated for days is something of a stress test and I
have lost the results of some computations after the GUI fell over as
the final results were copied and pasted into the notation page. It is
necessary to save and restart immediately after pasting the results to
secure the data. For the initial exploration of principle lines I
suggest a working time of 48 hours on a moderately fast machine the
nominal ply reached depends on the engine Fritz8 (19), Rybka232 (22),
Shredder10 (24).

Stage 1 is to explore the very first move. There are considerable
differences in engines evaluations of best and worst moves, but after
a short while they all agree e4 is best and f3 is the worst.

I propose at stage 2 to analyse the principle variation after every
one of whites possible opening moves. There are 20 of these and so
there are 40 CPU days needed for this alone. And stage 3 is to explore
all 20 x 20 continuations for their principle lines (800 CPU days).
The final stage will be to run deep positional analysis with carefully
chosen settings to explore the most interesting lines in greater
breadth (4000 CPU days).

It has already thrown up some surprises for me. I have always defended
the latest engines ability to play the opening strongly in their own
right and with Shredder10 I have generally found this to be true. So
imagine my surprise when after 32h a stronger engine, Rybka 2.32 had
as its principal line after 32 hours of cogitation:

1. e4 Nc6
2. d4 d5
3. e5 Nh6?
4. Nc3

It takes a 17 ply search following 3. ... Nh6 for Rybka to see that
Bxh6 gxh6 doesn't make for a nice comfortable position. So there is a
clear example of an engine blindspot here (or a hash table quirk). My
hashtable was 384MB so if someone wants to reproduce this with a
different sized table that would be interesting.

The full analysis of the principal lines for white for every possible
opening move are below. Sorry but this will look a right mess. I'd be
interested in any other weak moves you can spot in the first 8-10 ply.

[White "New game"]
[Black "Rybka 2.3.2 32-bit"]
[Result "*"]
[ECO "B00"]
[Annotator "Brown,Martin"]
[PlyCount "14"]

{384MB, Power06.ctg, AOI}
{Rybka 2.3.2 32-bit : 1)} 1. e4
({Rybka 2.3.2 32-bit : 2)} 1. Nf3 Nf6 2. d4 g6 3. c4 Bg7 4. Nc3 O-O 5.
e4 c5 6. d5 Qa5 7. Bd2 Na6 {0.12/22})
({Rybka 2.3.2 32-bit : 3)} 1. d4 Nf6 2. Nf3 g6 3. c4 Bg7 4. Nc3 O-O 5.
e4 c5 6. d5 Qa5 7. Bd2 Na6 {0.12/22})
({Rybka 2.3.2 32-bit : 4)} 1. Nc3 d5 2. Nf3 Nf6 3. d4 g6 4. g3 Bg7 5.
Bg2 O-O 6. O-O Bf5 7. Bf4 Nh5 {0.02/22})
({Rybka 2.3.2 32-bit : 5)} 1. g3 Nf6 2. Nf3 g6 3. d4 Bg7 4. Bg2 d5 5.
O-O O-O 6. Nc3 Bf5 7. Bf4 Nh5 {0.02/22})
({Rybka 2.3.2 32-bit : 6)} 1. c4 e5 2. Nc3 Nf6 3. g3 Nc6 4. Bg2 Bc5 5.
Nf3 O-O 6. O-O d6 7. d3 Bg4 {0.00/20}) ({Rybka 2.3.2 32-bit : 7)} 1.
e3 Nf6 2. Nf3 g6 3. c4 Bg7 4. Nc3 O-O 5. Be2 d5 6. cxd5 Nxd5 7. d4 Nc6
{0.00/21})
({Rybka 2.3.2 32-bit : 8)} 1. d3 Nf6 2. g3 Nc6 3. Bg2 d5 4. Nf3 g6 5.
O-O Bg7 6. c4 O-O 7. Nc3 d4 {-0.02/22})
({Rybka 2.3.2 32-bit : 9)} 1. h3 Nf6 2. Nc3 d5 3. d4 c5 4. e3 Nc6 5.
Nf3 e6 6. Bd2 Bd6 7. Bd3 O-O {-0.02/22})
({Rybka 2.3.2 32-bit : 10)} 1. a3 d5 2. d4 g6 3. Nf3 Bg7 4. Nc3 Nf6 5.
g3 Nc6 6. Bg2 Ne4 7. O-O O-O {-0.06/22})
({Rybka 2.3.2 32-bit : 11)} 1. f4 Nf6 2. Nf3 g6 3. Nc3 d5 4. e3 Bg7 5.
d4 O-O 6. a3 a6 7. Ne5 Qd6 {-0.12/22})
({Rybka 2.3.2 32-bit : 12)} 1. a4 e5 2. e4 Nf6 3. Nc3 Bb4 4. Bc4 O-O
5. Nf3 Nc6 6. O-O Nxe4 7. Nxe4 d5 {-0.13/22})
({Rybka 2.3.2 32-bit : 13)} 1. b3 Nf6 2. Nf3 g6 3. Nc3 Bg7 4. e3 O-O
5. Bd3 Na6 6. Bxa6 bxa6 7. O-O c5 {-0.13/22})
({Rybka 2.3.2 32-bit : 14)} 1. c3 Nf6 2. Nf3 Nc6 3. d4 d5 4. g3 g6 5.
Bg2 Bg7 6. Be3 O-O 7. O-O Bf5 {-0.14/22})
({Rybka 2.3.2 32-bit : 15)} 1. h4 e5 2. c4 Nf6 3. Nc3 Bb4 4. e3 Nc6 5.
Nd5 O-O 6. Nf3 Bc5 7. Ng5 h6 {-0.18/22})
({Rybka 2.3.2 32-bit : 16)} 1. Nh3 e5 2. e3 Nf6 3. d4 Nc6 4. d5 Nb4 5.
c4 Be7 6. Be2 O-O 7. O-O d6 {-0.22/22})
({Rybka 2.3.2 32-bit : 17)} 1. Na3 e5 2. d3 Nc6 3. Nf3 d5 4. g3 Nf6 5.
Bg2 Be7 6. O-O O-O 7. c4 e4 {-0.24/22})
({Rybka 2.3.2 32-bit : 18)} 1. g4 d5 2. h3 e5 3. Nf3 e4 4. Nd4 h5 5.
g5 Qxg5 6. d3 Qe5 7. dxe4 dxe4 {-0.27/22})
({Rybka 2.3.2 32-bit : 19)} 1. b4 e5 2. b5 Nf6 3. Bb2 Bc5 4. Bxe5
Bxf2+ 5. Kxf2 Ng4+ 6. Ke1 Nxe5 7. Nf3 Nxf3+ {-0.27/22})
({Rybka 2.3.2 32-bit : 20)} 1. f3 e5 2. Nc3 Nf6 3. e3 Nc6 4. Bd3 d5 5.
Nge2 Bd6 6. O-O O-O 7. Ng3 Be6 {-0.33/22})
1... Nc6 (1... c5 2. Nf3 d6 3. d4) 2. d4 d5 3. e5 Nh6 4. Nc3 g6 5. h3
Bg7 6. Nf3 O-O 7. Be2 f6 {0.17/23} *

The other engines I have tried so far are Fritz & Shredder. I haven't
spotted either of them doing funny looking moves yet. It is also clear
that they have an exaggerated sense of good opening play cf Rybka.

Rybka best 1. Nc3 +0.16, 1. e4, 1.d4 +0.12, 1. Nf3, 1. g3 +0.02, 1.
e3, 1. d3 0.0 worst 1.f3 -0.33
Fritz8 best 1. e4 +0.22, 1. d4, 1. Nf3, 1. e3, 1.c4 +0.06, 1. Nc3
+0.03, worst 1. f3 -0.44 1. g4 -0.53
Shredder10 analysis lost in crash best 1. e4 +0.40 worst 1. f3 -0.93
(recorded by hand)

I am inclined to think Rybkas scores are closer to the truth, but that
Shredder is better at picking its way through the opening without the
help of a book. Fritz8 is a bit too slow to be interesting for this
task.

Anyone interested in running some of these please drop me a line - the
odd looking email address is valid as it is.

Regards,
Martin Brown
  #2   Report Post  
Old January 18th 08, 12:41 PM posted to rec.games.chess.computer, rec.games.chess.analysis
external usenet poster
 
First recorded activity by ChessBanter: Sep 2006
Posts: 28
Default Engine generation of an opening book

Martin, this seems the most inefficient conceivable way to find
answers. Much better to collect a vast quantity of engine games and
see what they empirically tell you as a first step. THEN you can work
on evaluative analysis.

You are treading on ground that others have already plowed, you know.
Practically all computer chess experts already have very formidable
books--large, heavily analyzed and customized. None are in public
domain as they are used competitively and nothing defeats competitive
effectiveness than letting your opponents see your book. So the only
real answer is to find those sources of games and, using your own
criteria and methods, build something that works for you.
  #3   Report Post  
Old January 21st 08, 12:46 PM posted to rec.games.chess.computer, rec.games.chess.analysis
external usenet poster
 
First recorded activity by ChessBanter: Jul 2006
Posts: 1,015
Default Engine generation of an opening book

On Jan 18, 12:41*pm, wrote:
Martin, this seems the most inefficient conceivable way to find
answers.


It isn't meant to be an efficient way, and it certainly will not be
the strongest possible book. But it is intended to address the
question about whether or not some modern engines can find adequate
opening theory for themselves given enough time and without any human
intervention (beyond choosing some magic numbers for search time/depth/
width).

*Much better to collect a vast quantity of engine games and
see what they empirically tell you as a first step. *THEN you can work
on evaluative analysis.


The trouble is that most engine-engine matches will have books enabled
and so cannot be considered pure. I want to see how well (or badly)
the engine can do in discovering the opening lines with a significant
search depth (way beyond what would be realistic in tournament play).
This stems from previous discussions with ChessOne about the validity
or otherwise of opening books in chess engine games against humans.

You are treading on ground that others have already plowed, you know.
Practically all computer chess experts already have very formidable
books--large, heavily analyzed and customized. *None are in public
domain as they are used competitively and nothing defeats competitive
effectiveness than letting your opponents see your book. *So the only
real answer is to find those sources of games and, using your own
criteria and methods, build something that works for you.


What additional tools are available for opening book preparation? I
have CB9.

Regards,
Martin Brown
Reply
Thread Tools
Display Modes

Posting Rules

Smilies are On
[IMG] code is Off
HTML code is Off
Trackbacks are On
Pingbacks are On
Refbacks are On


Similar Threads
Thread Thread Starter Forum Replies Last Post
Books? Frenki rec.games.chess.misc (Chess General) 4 December 4th 07 03:26 PM
How to get good at chess? inquisitor rec.games.chess.misc (Chess General) 33 August 24th 06 11:02 AM
Books for beginning player? [email protected] rec.games.chess.misc (Chess General) 7 August 23rd 06 03:06 AM
Chess Book Artificer rec.games.chess.misc (Chess General) 6 March 13th 06 12:46 AM
Computer Chess; Chess Books Sanford rec.games.chess.misc (Chess General) 2 December 6th 05 08:34 PM


All times are GMT +1. The time now is 02:10 AM.

Powered by vBulletin® Copyright ©2000 - 2019, Jelsoft Enterprises Ltd.
Copyright 2004-2019 ChessBanter.
The comments are property of their posters.
 

About Us

"It's about Chess"

 

Copyright © 2017