[Computer-go] Kaś Cup - results and prizes
pasky at ucw.cz
Tue Aug 7 03:35:29 PDT 2012
On Tue, Aug 07, 2012 at 02:08:49AM +0200, Łukasz Lew wrote:
> Notably at least one game (Round 2 Aya : Pachi) ended with premature
> resignation at final position.
Yes. This was a bug in Pachi - its dynamic komi state got reset when the
opponent played an unexpected move (that could therefore not be promoted
to the MCTS tree root); the move was unexpected because Pachi played
very fast at the game end, since its time controls are optimized for
shorter games (or won games with Pachi playing quickly at the end due to
very high winrate).
I have fixed the bug shortly after the game, alas the third place has
already escaped between Pachi's fingers. :-)
> Program authors / operators, please state in this thread the hardware used.
Intel(R) Core(TM) i7-2600 CPU @ 3.40GHz. It seems it is auto-overlocked
to 3.7GHz. Four cores, eight threads.
> If you wish, tell us whether and how you modified your programs.
Couple of things (summarized in new 'maximize_score' parameter in
* Pachi passes when status of all intersections is settled, not when
it wins subsequent counting.
* Dynamic komi - adaptive strategy as described in my paper, with the
exceptions that the komi ratchet is used even for losing extra komi
and dynamic komi is used to the very end of the game (normally its
use is discontinued near the end of the game to make sure it is
wrapped up properly).
* Value scaling - Pachi used to incorporate win margin in values
stored in the tree ([0,0.04],[0.96,1] instead of 0,1). We turned
that off recently since it started having a tiny negative Elo
contribution to Pachi's strength. For this tournament, I have
turned it on and added two enhancements; the win margin is
measured against average outcome of the game in this search episode,
not against jigo; and the margin contribution diminishes for balanced
tree branches (a=0.01 for values near 0.5, a=0.05 for values near
0 or 1). However, this has a fairly complex interaction with dynamic
komi which tries to keep the value around 0.5.
In my testing, this version of Pachi is about 60 Elo weaker than normal
Pachi against Fuego 1.1 with 500s/game (PachIV's rating drop against
humans on KGS is about 40 Elo). I did some rudimentary tuning but not
too much; perhaps I should have eventually gone with Pachi w/o dynamic
komi for the tournament like others... :) However, it seems to me
that even though it is weaker, playing Pachi with maximize_score
settings is more pleasurable for humans, so I also added a suggestion
about this to Pachi's README.
Congratulations to the winners as well! This was a fun exercise;
it's a pity that in the end, most programs competed unmodified.
If there will be a second installment, it would be great if it
would be announced more in advance (this was a very short call for
any modifications requiring serious testing and tuning) and perhaps
figure out a way to increase people's motivation not to use stock
versions of their programs.
(One random idea - discourage going for *too* big win margins,
e.g. reward starts decreasing with win margin 20 or more. This is
sort of Hikaru handicap game style challenge, going for close games.)
Petr "Pasky" Baudis
More information about the Computer-go