Page 1 of 1

Toga II 1.3 with HT=80

Posted: Tue May 15, 2007 9:54 pm
by Jouni
According to http://www.computerchess.org.uk/ccrl/404/
HT=80 is the best Toga version. I played 6*20 games against same engines to check. First default (HT=75):

Code: Select all


Toga II 1.3   - Rybka 2.3.1 32-bit     7.0 - 13.0    +5/-11/=4    35.00%
Toga II 1.3   - Fritz 10               7.5 - 12.5    +5/-10/=5    37.50%
Toga II 1.3   - Deep Sjeng 2.5         13.0 - 7.0    +10/-4/=6    65.00%
Toga II 1.3   - Strelka                10.5 - 9.5    +8/-7/=5    52.50%
Toga II 1.3   - Spike 1.2 Turin        9.5 - 10.5    +7/-8/=5    47.50%
Toga II 1.3   - Glaurung 2             15.5 - 4.5    +12/-1/=7    77.50%

Total = 63.

With HT = 80:

Code: Select all


Toga II 1.3   - Rybka 2.3.1 32-bit     7.0 - 13.0    +3/-9/=8    35.00%
Toga II 1.3   - Fritz 10               12.0 - 8.0    +9/-5/=6    60.00%
Toga II 1.3   - Deep Sjeng 2.5         11.5 - 8.5    +6/-3/=11    57.50%
Toga II 1.3   - Strelka                11.0 - 9.0    +6/-4/=10    55.00%
Toga II 1.3   - Spike 1.2 Turin        10.5 - 9.5    +8/-7/=5    52.50%
Toga II 1.3   - Glaurung 2             14.0 - 6.0    +11/-3/=6    70.00%

Total = 66. It's not worse as default I think. BTW what is this history threshold meaning?

Jouni

Re: Toga II 1.3 with HT=80

Posted: Wed May 16, 2007 12:02 pm
by mjlef
Jouni wrote:According to http://www.computerchess.org.uk/ccrl/404/
HT=80 is the best Toga version. I played 6*20 games against same engines to check. First default (HT=75):

Code: Select all


Toga II 1.3   - Rybka 2.3.1 32-bit     7.0 - 13.0    +5/-11/=4    35.00%
Toga II 1.3   - Fritz 10               7.5 - 12.5    +5/-10/=5    37.50%
Toga II 1.3   - Deep Sjeng 2.5         13.0 - 7.0    +10/-4/=6    65.00%
Toga II 1.3   - Strelka                10.5 - 9.5    +8/-7/=5    52.50%
Toga II 1.3   - Spike 1.2 Turin        9.5 - 10.5    +7/-8/=5    47.50%
Toga II 1.3   - Glaurung 2             15.5 - 4.5    +12/-1/=7    77.50%

Total = 63.

With HT = 80:

Code: Select all


Toga II 1.3   - Rybka 2.3.1 32-bit     7.0 - 13.0    +3/-9/=8    35.00%
Toga II 1.3   - Fritz 10               12.0 - 8.0    +9/-5/=6    60.00%
Toga II 1.3   - Deep Sjeng 2.5         11.5 - 8.5    +6/-3/=11    57.50%
Toga II 1.3   - Strelka                11.0 - 9.0    +6/-4/=10    55.00%
Toga II 1.3   - Spike 1.2 Turin        10.5 - 9.5    +8/-7/=5    52.50%
Toga II 1.3   - Glaurung 2             14.0 - 6.0    +11/-3/=6    70.00%

Total = 66. It's not worse as default I think. BTW what is this history threshold meaning?

Jouni
Hi Jouni,

It is the percentage of times a move being considered for reduction has actually been found to be "good" (cause a fail high when it was played before in the search). So 80% would mean that 4 out of 5 times this move was found to be good. The lower this value is, the lower the chance a good move will be reduced, but the larger the overall tree will be.

Mark

Re: Toga II 1.3 with HT=80

Posted: Wed May 16, 2007 12:16 pm
by Vempele
mjlef wrote:
Jouni wrote: Total = 66. It's not worse as default I think. BTW what is this history threshold meaning?

Jouni
Hi Jouni,

It is the percentage of times a move being considered for reduction has actually been found to be "good" (cause a fail high when it was played before in the search). So 80% would mean that 4 out of 5 times this move was found to be good. The lower this value is, the lower the chance a good move will be reduced, but the larger the overall tree will be.

Mark
It's also used in Toga's extended history pruning, which in 1.3X4 means using half the threshold in PV nodes.

Toga also outright prunes some moves because of bad history; the values used are HT/2 for depths 1 and 2, HT/3 for the rest, with the additional condition that the remaining depth must be equal or lower than root depth/2.

Note that re-search after Extended history pruning is buggy: it first adjusts the bounds (and searches again) and if the score continues to be above alpha, it does another research with the non-reduced depth. Sort of like closing your eyes and hoping real hard search instability will make the problem go away. :roll: