[Computer-go] Narrow wins
fotland at smart-games.com
Mon Jun 10 23:48:01 PDT 2013
Sorry, secret. It took me quite a while to find something that worked well.
I think you are right that most of the benefit comes from keeping the win
rate close to 50%. I think Pachi found it was ideal to keep the win rate a
little below 50%. I try to keep it in the range of 45% to 55%. I think
Pachi's approach has been published. Mine is rather different.
From: computer-go-bounces at dvandva.org
[mailto:computer-go-bounces at dvandva.org] On Behalf Of Don Dailey
Sent: Monday, June 10, 2013 9:02 AM
To: computer-go at dvandva.org
Subject: Re: [Computer-go] Narrow wins
On Mon, Jun 10, 2013 at 11:18 AM, David Fotland <fotland at smart-games.com>
The dynamic komi can be pretty large. I think I limit it to 30 points or
so. If the komi were only a point or two you would see a lot of games won
by 2.5 points or so. You can see from the results that the margins are
often pretty large.
A big issue which dynamic komi tries to solve is the difficulty in resolving
close scores. For example in a won position the playouts might return
99.8% wins and a move which wins by larger margins might still only return
99.8 or even 99.7 or something just due to statistical noise.
In computer chess testing we have a similar issue - especially as our
program has gotten so strong. If you test against a weak program, it's
almost impossible to resolve small differences in strength. Imagine a
weak club player playing Go against the world champion in order to resolve
the relative ELO difference (assuming no handicapping.) The pro is likely
to win thousands of games before losing a single one and there would be no
basis for determining where this player stands with any precision.
So I assume that by far the primary benefit of dynamic komi is to bring the
statistics back into the sweet spot so that it's easier to resolve the
quality of move from one to the other.
I have not kept up with this, but is your dynamic komi implementation a
carefully guarded secret? I would like to know more. It seems to me that
there would be some serious efficiency losses in "fishing" for the right
value to use - or is this arrived at pretty quickly?
> -----Original Message-----
> From: computer-go-bounces at dvandva.org [mailto:computer-go-
> bounces at dvandva.org] On Behalf Of Stefan Kaitschick
> Sent: Monday, June 10, 2013 1:24 AM
> To: computer-go at dvandva.org
> Subject: Re: [Computer-go] Narrow wins
> On Sun, Jun 9, 2013 at 9:23 PM, David Fotland <fotland at smart-
> games.com> wrote:
> > Dynamic komi and some other tricks work quite well. Thanks to
> > for pushing dynamic komi until I figured out how to make it work
> > Often the playout have some bias due to a misread in a fight, so
> > important for the bot to keep its lead.
> Is that a small, fixed bonus of 1 or 2 points?
> A large dynamic komi in the endgame would really surprise me.
> But to actually compensate a misread fight, that komi would have to
> be substantial ...
> Like Don, I don't think that much good could come of it.
> Isn't it superior to award a point bonus in the playouts?
> Here's another idea: find a smart way to average the playout results
> into an expected result.
> If the bot is leading, try to maintain that lead in the next move.
> Computer-go mailing list
> Computer-go at dvandva.org
Computer-go mailing list
Computer-go at dvandva.org
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Computer-go