Yamato wrote:
> Hi all.
>
> I'd like to share my test set for MC programs.
> This is a gnugo-style "regression test" file. To use it, you need
> to implement 2 GTP commands, loadsgf and reg_genmove.
> Then, run your program like this:
>
> gnugo --mode gtp < mctest.tst | awk -f regress.awk tst=mctest.tst
>
> This set contains 50 problems. They are relatively easy to human
> players and perhaps classic Go programs, but hard to MC programs.
>
> Here are my results.
>
> Zen-1.5 (1k playouts) : 15/50
> Zen-1.5 (10k playouts) : 20/50
> Zen-1.5 (100k playouts) : 22/50
> gnugo-3.7.10 (Level 0) : 22/50
> gnugo-3.7.10 (Level 10) : 32/50
>
> I want to know the results of other programs. Please post here.
GNU Go 3.7.12 level 0 : 21/50
GNU Go 3.7.12 level 10 : 32/50
GNU Go 3.7.12 mc, 1k (*1) : 28/50
GNU Go 3.7.12 mc, 10k (*2) : 28/50
GNU Go 3.7.12 mc, 100k (*3) : 35/50
GNU Go 3.7.12 mc, light, 1k (*4) : 30/50
GNU Go 3.7.12 mc, light, 10k (*5) : 27/50
GNU Go 3.7.12 mc, light, 100k (*6): 22/50
GNU Go 3.7.12 mc, mogo, 1k (*7) : 32/50
GNU Go 3.7.12 mc, mogo, 10k (*8) : 28/50
GNU Go 3.7.12 mc, mogo, 100k (*9) : 30/50
(*1) gnugo --monte-carlo --mc-games-per-level 100
(*2) gnugo --monte-carlo --mc-games-per-level 1000
(*3) gnugo --monte-carlo --mc-games-per-level 10000
(*4) gnugo --monte-carlo --mc-patterns uniform --mc-games-per-level 100
(*5) gnugo --monte-carlo --mc-patterns uniform --mc-games-per-level 1000
(*6) gnugo --monte-carlo --mc-patterns uniform --mc-games-per-level 10000
(*7) gnugo --monte-carlo --mc-patterns mogo_classic --mc-games-per-level 100
(*8) gnugo --monte-carlo --mc-patterns mogo_classic --mc-games-per-level
1000
(*9) gnugo --monte-carlo --mc-patterns mogo_classic --mc-games-per-level
10000
--mc-patterns uniform is uniform distribution except for filling eyes,
i.e. so called light playouts.
--mc-patterns mogo_classic is an approximation of the simulation
policy from the report "Modification of UCT with Patterns in
Monte-Carlo Go" by the MoGo team.
The reason why light playouts do better with 1k simulations than
100k simulations is that the --monte-carlo move generation is a hybrid
between UCT search and normal move generation and with few
simulations, especially light playouts, the influence of normal move
generation (in this case at default level 10) tends to be stronger.
/Gunnar
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
