[gnugo-devel] level of play

[david doshay]

Hello and a Happy New Year to all.

Before the holiday season I reported on using a cluster to play
GnuGo with look-ahead against GnuGo in its standard form.

Gunnar replied:
**********
> Game 1: W+48.5
> Game 2: W+14.5
> Game 3: B+36.5
> Game 4: W+11.5
> Game 5: W+24.5
> Game 6: B+16.5
> Game 7: W+5.5
> Game 8: W+21.5
> Game 9: W+5.5
> Game 10: B+0.5

Looks okay but of course way too few samples to be statistically
significant. Also an interesting question is how your lookahead
compares to a single GNU Go spending a comparable amount of
computations by playing at a higher level.
**********

And there have been other emails about the level of play between
Dan Bump and Hiroshi Yamashita:
**********
> I suggested in my earlier e-mail that you could run GNU
> Go at level 12, which is what we did in Europe. Arend
> pointed out that on an 850Mh machine, level 12 could
> be too slow. So maybe you should either run it at
> level 10 or do a test run first to see if it seems
> too slow on your hardware.

I tested GnuGo 3.5.3 Level 12 games with Aya and Katsunari.
It spent 8 minutes in Aya's game and 54 minutes in
 Katsunari's game on Celeron 850MHz machine.

So I'm going to use Level 10.
**********


However, in the code it says:
--level <amount>  strength (default %d, up to 10 supported)\n\

#define DEFAULT_LEVEL 10

So, we thought we were running at the highest level possible
when playing the above games.

Just to let you know about some other games we have run on our
cluster, when we throw in lots of resources, we managed to win 4
out of 8 doing a 16 move lookahead against a 4 stone handicap.

Game 1: B+42.5
Game 2: W+14.5
Game 3: W+4.5
Game 4: W+10.5
Game 5: W+53.5
Game 6: B+87.5
Game 7: B+24.5
Game 8: B+21.5

This was using 6 look-ahead paths, 4 based upon GNU Go's
estimate of our best move, and 2 paths based upon Gnu Go's
estimate of our opponent's best move.

Cheers,
David