Learning broken in ArasanX 9... versions?

[Guenther Simon]

I noticed several things with ArasanX 9.3 in a test tourney.

1. it doesn't create automatically a position.bin anymore,
when starting it or by double clicking the exe.
2. after taking a fresh one from version 8.4 I saw in the
logs that it mentioned learning values in the log but still
wasn't able to save them.
3. book learning also seems not to work anymore as the
book file also remained unchanged after 28 games?
4. After I noticed that the GUI exe! still is capable of producing
a fresh position file I tested with this one, but the same effect.

What's going on?

This message after game end looks like a saving error always?

Code: Select all

Learning result .. page = 0 index = 0 current value = 0


ArasanX log excerpts:

Code: Select all

Arasan version 9.3 game log
   move          time     depth   nodes   value   predicted

Arasan version 9.3 game log
   move          time     depth   nodes   value   predicted
1. e4         
1. c5           (book)
2. Nf3         
2. e6           (book)
3. d4         
3. cxd4         (book)
4. Nxd4       
4. a6           (book)
5. Nc3         
5. Qc7          (book)
6. Bd3         
6. Nf6          (book)
7. f4         
7. d6           (book)
8. Be3         
8. Be7            00:00:26    12   15116348    +0.00   Qf3 e5 Nf5 Bxf5 exf5 O-O O-O-O Nc6 g4
9. O-O         
9. O-O            00:00:26    12   14702563    +0.00   Qf3 Bd7 Nb3 Nc6 Bf2 Nb4 Bh4 Nxd3 Qxd3
10. f5         
10. Kh8           00:00:26    12   14222032    -0.06   fxe6 fxe6 Qf3 Bd7 Qg3 Rf7 Nb3 Bd8 Nd4
11. Qf3       
11. e5            00:00:26    13   14925382    +0.00   Nb3 Nbd7 Bg5 b5 Bxf6 Nxf6 Rad1 Qb6+ Kh1
12. Nb3       
12. Nbd7          00:00:26    13   14815675    +0.00   Bg5 b5 Rf2 Bb7 a3 Nc5 Rd1 Nxb3 cxb3
13. g4         
13. h6            00:00:26    13   15231567    +0.13   g5 hxg5 Bxg5 b5 Rad1 Qb6+ Kh1 Bb7 Qh3+
14. Qh3       
14. Kg8           00:00:26    12   15143417    +0.19   Qg2 b5 Rac1 Bb7 a3 Nc5 Rfd1 Nxd3 Rxd3
15. a4         
15. b6            00:00:26    13   14189597    +0.13   g5 hxg5 Bxg5 Bb7 Qg2 Qc6 Rad1 Bd8 Bxf6
16. g5         
16. hxg5          00:00:26    14   13846855    +0.19   Bxg5 Bb7 Qg2 Rfd8 Rad1 d5 Bh6 Bf8 exd5
17. Bxg5       
17. Bb7           00:00:26    12   13445852    -0.19   Qh4 d5 Nxd5 Bxd5 exd5 Bd6 c4 e4 Bxe4
learning position, score =  -0.19 depth = 12
18. Qg2       
18. Rfc8          00:00:26    11   13268151    +0.06   Rad1 Qc6 Rf3 Bf8 Bxf6 Nxf6 Re3 Rc7 Nd2
19. Bh6       
19. Bf8           00:00:26    12   14082851    +0.19   Bg5 Nh7 Rfd1 Kh8 Qh3 Bc6 a5 b5 Qh5 Re8
learning position, score =  +0.19 depth = 12
20. Be3       
20. Nh7           00:00:27    11   14197780    +0.19   Nd2 Be7 Bc4 Qd8 Nd5 Bh4 c3 Kh8 b4 Ng5
21. Rf3       
21. Nhf6          00:00:26    11   14038425    +0.19   Rff1 Bc6 Nd2 Be7 Bc4 Bb7 Rad1 Re8 Bg5
22. Rh3       
22. Nh7           00:00:26    11   13504998    -0.06   Bg5 Re8 Rg3 Ndf6 Bxf6 Nxf6 Rg5 Qc6 a5
23. Kh1       
23. Ndf6          00:00:53    13   27825564    -0.88   Rg1 Ne8 Qf2 Rcb8 Bxb6 Qd7 Be3 Bc6 Ra1
learning position, score =  -0.88 depth = 13
24. Rg1       
24. Ne8           00:00:25    13   13108179    -1.00   Rg3 Qd8 Bh6 Qf6 Bg5 Nxg5 Rxg5 Be7 Rg4
learning position, score =  -1.00 depth = 13
25. Nd2       
25. Rab8          00:00:25    13   12996313    -1.06   Nc4 Bc6 Qf2 Nef6 Nxb6 Rd8 Ncd5 Bxd5
26. Nc4       
26. Bc6           00:00:25    14   12554029    -1.19   Qf2 Nef6 Nxb6 Rd8 a5 d5 Ncxd5 Nxd5 exd5
27. Qf2       
27. Nef6          00:00:25    13   12529903    -1.50   Nxb6 Rd8
learning position, score =  -1.50 depth = 13
28. Nxb6       
28. Re8           00:00:50    13   24721649    -2.17   a5 Bb7 Qe2 Red8 Bh6 d5 Bxg7 Bxg7 Qg2
learning position, score =  -2.17 depth = 13
29. a5         
29. Bb7           00:00:46    12   23237716    -3.02   Qg2 Kh8 Bg5 Ng8 Qe2 d5 Ncxd5 Bxd5
learning position, score =  -3.02 depth = 12
30. Qh4       
30. Qc6           00:00:21    11   10574492   -10.91   Bg5 Be7 Bxf6 Bxf6 Qxh7+ Kf8 Ncd5 Qxd5 resigns
learning position, score = -10.91 depth = 11
Learning result .. page = 0 index = 0 current value = 0, new value = -1


Arasan version 9.3 game log
   move          time     depth   nodes   value   predicted

Arasan version 9.3 game log
   move          time     depth   nodes   value   predicted
1. c4           (book)
1. Nf6         
2. Nc3          (book)
2. g6         
3. e4           (book)
3. d6         
4. d4           (book)
4. Bg7         
5. f3           (book)
5. O-O         
6. Be3          (book)
6. e5         
7. d5           (book)
7. Nh5         
8. Qd2          (book)
8. f5         
9. O-O-O        (book)
9. fxe4       
10. fxe4          00:00:28    13   17290681    +0.31   a5 Nf3 Bg4 Bg5 Nf6 Be2 Nbd7 Qe3
10. Na6       
11. Be2           00:00:28    13   16306775    +0.38   Nf6 Nf3 Qe8 Kb1 Qe7 Bf2 Nb4 Ng5 Bd7
11. Nf4       
12. Bf3           00:00:28    12   15318641    +0.38   Nb4 Kb1 Bd7 Ka1 Qe7 g3 Nh3 Rf1 Nxg1
12. Bd7       
13. g3            00:00:28    13   15568781    +0.50   Nh3 Rf1 Nb4 Kb1 Rf7 Nxh3 Bxh3 Bg2 Bxg2
13. Nh3       
14. Rf1           00:00:28    13   16332237    +0.50   Nb4 Kb1 a5 Nxh3 Bxh3 Bg2 Bxg2 Qxg2 Qe7
learning
page = 10 index = 1164 factor =    0.50
page = 8 index = 460 factor =    0.50
page = 10 index = 2276 factor =    0.25
page = 18 index = 1410 factor =    0.25
page = 0 index = 1786 factor =    0.25
page = 14 index = 1097 factor =    0.25
page = 18 index = 2211 factor =    0.08
page = 6 index = 1376 factor =    0.08
page = 14 index = 2443 factor =    0.03
14. Nb4       
15. Kb1           00:00:28    12   15201095    +0.50   a5 Nxh3 Bxh3 Bg2 Bxg2 Qxg2 Qe7 a3 Nd3
15. a5         
16. a3            00:00:28    12   15691806    +0.44   Na6 Nxh3 Bxh3 Rf2 Qe7 Rd1 b6 Bg5 Qf7
16. Nxg1       
17. Rhxg1         00:00:09    13   5298633    +0.63   Na6 Bg2 Nc5 Rxf8+ Qxf8 Qd1 Qe7 Bxc5
17. Na6       
18. h4            00:00:29    12   16217182    +0.63   Nc5 h5 Bh3 Rf2 Nb3 Qd1 Nd4 hxg6 hxg6
18. Bh3       
19. Rf2           00:00:29    12   16723973    +0.94   Qe7 Rh1 Qd7 h5 gxh5 Bxh5 Rxf2 Qxf2 Rf8
19. Rf7       
20. g4            00:00:29    13   16599187    +1.44   Qxh4 Bg5 Qxf2 Qxf2 Raf8 Rg3 Bxg4 Rxg4
learning position, score =  +1.44 depth = 13
20. Qe7       
21. Bg5           00:00:29    13   17471009    +2.38   Qd7 Qd1 h5 gxh5 gxh5 Be3 Raf8 Qe2 Bg4
learning position, score =  +2.38 depth = 13
21. Qf8       
22. Qe2           00:00:29    14   16890197    +2.78   Nc5 Rg3 Bxg4 Rxg4 Nb3 Qf1 Nd4 Rg3 a4
learning position, score =  +2.78 depth = 14
22. h6         
23. Be3           00:00:29    14   17071034    +3.14   Qe7 Rh2 Qxh4 Rgh1 Rxf3 Qxf3 Qxg4 Rxh3
learning position, score =  +3.14 depth = 14
23. h5         
24. gxh5          00:00:29    15   17535063    +4.13   gxh5 Nd1 Bg4 Bxg4 hxg4 Rfg2 Bf6 Qxg4+
learning position, score =  +4.13 depth = 15
24. gxh5       
25. Nd1           00:00:29    16   16717233    +4.06   Bg4 Bxg4 hxg4 Rfg2 Bf6 Qxg4+ Kh8 Nc3
learning position, score =  +4.06 depth = 16
25. Bg4       
26. Bxg4          00:00:29    15   17300026    +4.14   hxg4 Qxg4 Rxf2 Nxf2 Qf7 Bh6 Kh7 Bxg7
26. hxg4       
27. Qxg4          00:00:09    14   5665603    +4.22   Rxf2 Nxf2 Qf6 h5 Kh8 h6 Bxh6 Bxh6 Qxh6
27. Rxf2       
28. Nxf2          00:00:10    14   6059951    +4.28   Qf6 h5 Kh8 h6 Bxh6 Rh1 Rg8 Qe2 Kh7 Ng4
28. Re8       
29. h5            00:00:32    14   18581533    +5.13   Kh8 Qg6 Ra8 h6 Bf6 Ng4 Bh4 h7 Qg7 Nxe5
learning position, score =  +5.13 depth = 14
29. Kh8       
30. Qg6           00:00:32    15   19766949    +5.89   Re7 h6 Rf7 hxg7+ Qxg7 Qh5+ Qh7 Qe2 Rxf2
learning position, score =  +5.89 depth = 15
30. Re7       
31. h6            00:00:32    16   19745241    +6.50   Bf6 Ng4 Bh4 Qh5 Rf7 Qxh4 Rf1+ Kc2 Rxg1
31. Bf6       
32. Ng4           00:00:32    16   19932226    +6.86   Bh4 Qh5 Rf7 Qxh4 Rf1+ Rxf1 Qxf1+ Kc2
32. Bh4       
33. Qh5           00:00:32    15   19405728    +6.94   Rf7 Qxh4 Rf1+ Rxf1 Qxf1+ Kc2 Qf8 Qf6+
33. Nb8       
34. Qxh4          00:00:32    14   18302580    +7.38   Nd7 Qg5 Rh7 c5 Qe7 cxd6 Qxg5 Bxg5 cxd6
34. Nd7       
35. Qg5           00:00:33    13   18459653    +7.33   Qf7 Nf2 Qh7 Ng4 Rf7 Qd8+ Rf8 Qxc7 Qxe4+
35. a4         
36. c5            00:00:32    14   19222706    +7.72   Rf7 cxd6 cxd6 Ka2 Qc8 Bc1 Rh7 Qg6 Qf8
36. Rf7       
37. Ka2           00:00:33    14   19475811    +7.53   Kh7 cxd6 cxd6 Nf2 Kh8 Ng4 Kh7
37. Kh7       
38. Qh5           00:00:33    14   18132268    +7.69   Kh8 cxd6 cxd6 Qg6 b6 Rg2 Nc5 Bg5 b5
38. Re7       
39. Bg5           00:00:33    14   18284628    +7.75   Rf7 cxd6 cxd6 Rc1 Rf3 Rc7 Qf7 Qxf7+
39. Rf7       
40. Bd8           00:00:33    14   18765033    +8.97   Rf4 Nf2 Kh8 h7 Nf6 Bxf6+ Rxf6 Rg8+ Qxg8
40. Rf1       
41. Nf6+          00:00:22    15   13588739   +10.83   Rxf6 Rg7+ Kh8 Rxd7 Rf1 Bxc7 Ra1+ Kxa1
41. Rxf6       
42. Bxf6          00:00:22    16   13928736   +16.88   Nxf6
42. Nxf6       
43. Qg6+          00:00:08    14   5571797   +Mate9   Kh8 Rf1 b5 Rxf6 Qg8 Qf5 b4 Rf8 b3+ Ka1
43. Kh8       
44. Rf1           00:00:00    10   18029   +Mate9   b5
44. b5         
45. Rxf6          00:00:00    10   11573   +Mate9   Qg8
45. Qg8       
46. Qf5           00:00:00    10   149315   +Mate6   b4 Rf8 b3+ Ka1 c6 Qf6+ Kh7 Rxg8 Kxg8
Learning result .. page = 0 index = 0 current value = 0, new value = 1
Learning result .. page = 0 index = 0 current value = 0, new value = 1
Learning result .. page = 0 index = 0 current value = 0, new value = 1


Relevant learning options in the Arasan.rc file:
#
# Learning
#
# true to enable book learning based on score out of book, false to disable
learning.score_learning=true
#
# Number of moves out of book at which learning occurs:
learning.score_learning.horizon=5
#
# Score threshold at which learning occurs (in units of 1 Pawn=100)
learning.score_learning.threshold=33
#
# Minimum search depth for score learning
learning.score_learning.minDepth=7
#
# True to enable use of book learning info based on game results
learning.result_learning=true
#
# True to enable storage & use of position values, false to disable
learning.position_learning=true
#
# Position learning threshold (in units of 1 Pawn=100)
learning.position_learning.threshold=50
#
# Minimum depth for position learning
learning.position_learning.minDepth=7
#

Guenther

[jdart]

I'll take a look at it. I have verified that it does not seem to create position.bin.

--Jon

[jdart]

Actually, the learning file (position.bin) is created not at program start, but when the "new" command is received from Winboard. So it should appear, but only after a game has been started. If you just run "arasanx" from the console and then type "new" you can observe this.

As for the messages after the game related to result learning: this is a minor bug - the location of the learning info is not written to the log correctly, but the info is written to the book file (note: result learning goes into book.bin, not position.bin). So learning should be working ok, it is only the log output that is incorrect.


--Jon

[Guenther Simon]

Actually, the learning file (position.bin) is created not at program start, but when the "new" command is received from Winboard. So it should appear, but only after a game has been started. If you just run "arasanx" from the console and then type "new" you can observe this.

As for the messages after the game related to result learning: this is a minor bug - the location of the learning info is not written to the log correctly, but the info is written to the book file (note: result learning goes into book.bin, not position.bin). So learning should be working ok, it is only the log output that is incorrect.


--Jon

Hi Jon,

The problem is as mentioned above, the book file did _not_ change
after 28 games (also position bin remained _unchanged_ after
28 games)
The games were played in WB 4.27x at 40/15.
All debugs and logs are saved.

Best regards and thanks for the quick response,
Guenther

[Guenther Simon]

It would be nice if a mod reading that thread would move
it to the main forum
Sorry for accidentally missplacing...

Regards,
Guenther

[Richard Pijl]

done

[jdart]

I will take a further look at the learning behavior.

--Jon

[jdart]

It looks to me like the book.bin and position.bin files are changing - if I run a md5 hash over them before and after a match I see changes - but the file modification time does not change, for some reason.

--Jon

[Guenther Simon]

It looks to me like the book.bin and position.bin files are changing - if I run a md5 hash over them before and after a match I see changes - but the file modification time does not change, for some reason.

--Jon

That's strange and interesting In previous versions before
ArasanX 9... versions the time stamp changed and I believe
also I know of no learning files from whatever programs which
wouldn't change timestamps if learning happened.
I will do some binary comparisons too before and after
some games.
Thanks for the first infos. Do you have any idea about
the unchanged timestamps so far?

Best regards,
Guenther

[Guenther Simon]

Hi Jon,

I can verify that after a testmatch of 4 games both files
,book and position bin, contained some binary changes.
In my example 343 and 340 bytes were different despite
the unchanged timestamp.
This means learning obviously works even when the timestamp
for Arasans book/learn files for unknown reasons never change
since version 9x.

Best regards,
Guenther