Some sloppy results

[Uri Blass]

Next time you release a version, I will test it with the new book. However, I cant change my time control. Do you have any plans to fix incremental time controls?

I already have "sort of" fixed the time management. The fix will be in the next version of Sloppy. I say "sort of" because Sloppy will still perform better at longer time controls, but that has very little to do with time management.

I find that the difference between ccrl 40/4 and 40/40 is usually extremely small and I rarely find cases of difference of more than 50 elo between programs with many games.

I wonder if sloppy is going to be different.

Here are some example when I chose only programs who have at least 400 games in CCRL 40/40 and rating behind 2700

Only in one case I found difference of more than 50 elo(in frenzee3.00 and I suspect something is wrong with testing of frenzee3.00 because the 64 bit version of frenzee3.00 is more than 100 elo better in blitz and this seem too much gain from 64 bits)

CCRL 40/40
53 Amyan 1.597 2570 +27 −27 44.5% +38.9 31.7% 467
50.1%

CCRL 40/4

54 Amyan 1.597 2562 +14 −14 46.2% +25.9 26.8% 1910
56.5%

CCRL 40/40

48‑49 AnMon 5.60 2610 +22 −22 36.3% +96.2 29.3% 743(-7 elo)
50.1%

CCRL 40/4

43‑44 AnMon 5.60 2617 +14 −14 48.3% +12.7 26.6% 1796
49.9%


CCRL 40/40

47 Little Goliath Evolution 3.12 2612 +25 −25 42.4% +51.4 32.1% 532
53.8%

CCRL 40/4

46 Little Goliath Evolution 3.12 (TB On) 2613 +18 −18 48.6% +9.1 29.0% 1055
53.4%
Little Goliath Evolution 3.12 2610 +13 −13 52.3% −18.1 28.2% 2142
61.0%

CCRL 40/40

41 Ufim 8.02 2643 +28 −28 49.7% −1.9 29.6% 436
50.8%

CCRL 40/4
45 Ufim 8.02 2615 +12 −12 50.9% −9.2 28.7% 2460
58.4%

CCRL 40/40

Colossus 2006f 2660 +23 −23 47.4% +17.1 31.5% 660
55.7%

CCRL 40/4

Colossus 2006f 2644 +17 −17 46.2% +26.5 30.3% 1213
51.5%

CCRL 40/40

38‑39 SOS 5.1 2661 +21 −21 45.3% +29.8 36.0% 744
52.9%

CCRL 40/4

32 SOS 5.1 2682 +12 −12 46.8% +23.0 28.9% 2505
55.6%

CCRL 40/40

Frenzee 3.0 2665 +28 −28 51.6% −17.9 33.8% 438
57.8%

CCRL 40/4
Frenzee 3.0 2602 +24 −24 48.0% +14.8 27.1% 616
51.1%
Frenzee 3.0 64-bit 2710 +15 −16 49.4% +4.2 29.5% 1476
53.9%

CCRL 40/40

35 Jonny 2.83 32-bit 2678 +22 −22 47.8% +13.6 30.9% 712
49.7%

CCRL 40/4

38 Jonny 2.83 32-bit 2651 +14 −14 46.2% +29.6 25.9% 1980
51.3%


CCRL 40/40

32 Pseudo 0.7c 2691 +19 −19 49.8% +0.1 38.7% 860
54.3%

CCRL 40/4

33 Pseudo 0.7c 2676 +11 −11 50.7% −8.3 30.1% 2997
55.5%

CCRL 40/40

WildCat 6 2696 +22 −22 47.1% +18.9 37.5% 680
53.0%

CCRL 40/4

WildCat 6 2659 +17 −17 53.5% −35.1 25.4% 1305
54.6%


CCRL 40/40

30 Aristarch 4.50 2697 +15 −15 43.9% +44.5 34.1% 1523
52.6%

CCRL 40/4

35 Aristarch 4.50 2668 +12 −12 46.0% +30.8 27.3% 2707
48.4%

Note that sloppy started very well in the blitz CCRL list by beating bugchess 20-12

http://computerchess.org.uk/ccrl/404/cg ... oppy_0_1_1

Uri

[Tony Thomas]

Uri, when he said longer time controls, it is possible that he was talking about something that is longer than my time control.

[Alessandro Scotti]

I find that the difference between ccrl 40/4 and 40/40 is usually extremely small and I rarely find cases of difference of more than 50 elo between programs with many games.

Both Hamsters and Kiwi seem to perform considerably worse at blitz... would love to get those 50 elo back!

[Uri Blass]

I find that the difference between ccrl 40/4 and 40/40 is usually extremely small and I rarely find cases of difference of more than 50 elo between programs with many games.

Both Hamsters and Kiwi seem to perform considerably worse at blitz... would love to get those 50 elo back!

My guess is that one of the secrets of better results at longer time control is simply having a slower engine and if you find some trick to make Hamsters 10 time faster without more changes the advantage of hamsters at longer time control is going to disappear.

It is possible to test it if CCRL test hamsters with more time(I think that testing always with equal time control is a bad idea and we should find testers also for not equal time control when arena can be used to give hamsters 40/40 against 40/4 and 40/400 against 40/40).

Uri

[Dann Corbit]

I ran over 100 games at fast time control and 30 games at 30 min base + 5 sec/move.

[ilari]

I find that the difference between ccrl 40/4 and 40/40 is usually extremely small and I rarely find cases of difference of more than 50 elo between programs with many games.

I wonder if sloppy is going to be different.

As Tony guessed, by "longer time controls" I meant longer than the one Tony uses.

These results are very different:

40 moves/2 minutes:
Sloppy-0.1.1 vs. BugChess2 V1.5.2: +55 -25 =20 65%

1m+1sec (what Tony uses):
Sloppy-0.1.1 vs. BugChess2 V1.5.2: +38 -43 =19 47.5%


The current development version does better at 1m+1sec, but there's still a noticeable difference.

[Jim Ablett]

Hi Alessandro,

Both Hamsters and Kiwi seem to perform considerably worse at blitz... would love to get those 50 elo back!

Maybe this will help a lttle bit.

Code: Select all

Kiwi 0.6d:

Allesandro's:

perft: depth=6, FEN=rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - - -
  depth=1, nodes=20
  depth=2, nodes=400
  depth=3, nodes=8902
  depth=4, nodes=197281
  depth=5, nodes=4865609
  depth=6, nodes=119060324
perft complete: total=124132536 nodes in 15.125 seconds (8207 KNps)


Mine:

perft: depth=6, FEN=rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - - -
  depth=1, nodes=20
  depth=2, nodes=400
  depth=3, nodes=8902
  depth=4, nodes=197281
  depth=5, nodes=4865609
  depth=6, nodes=119060324
perft complete: total=124132536 nodes in 14.219 seconds (8730 KNps)

Kiwi 0.6d win32 Intel compiler 10 p.g.o (includes all processor and Intel® Core2™ Duo, Quad & Xeon(R) only builds)


Download:

http://www.zshare.net/download/48184466d0a7de/
Mirror:
http://mihd.net/lx9fyw
Mirror:
http://w14.easy-share.com/9846601.html

rgds,
Jim.