[Computer-go] How many probes down the tree are necessary for a "good" bot?

David Fotland fotland at smart-games.com
Sat Nov 16 00:19:49 PST 2013


Bot strength is still progressing, but now it is more a matter of
engineering than grand new ideas.  It's not so interesting to say something
like, "I fixed 5 bugs this weekend and got 30 ELO", or "I used CLOP to tune
30 parameters for a week and got 50 ELO".

I think all the strong programs use some form of dynamic Komi.  There are a
few published papers on it, but I suspect the strong implementations are
rather different from the papers.  Mine is.  Nothing secret, just more
complex.

On-line learning in the playout policy is another thing that is worth doing,
but needs considerable engineering to a good strength boost.  Several papers
talk about this, so the new ideas are not secret, just the details of the
implementations.

Since each program has different go knowledge, any sharing of tuning
parameters or small bits of knowledge would not be very interesting to the
larger community.

I think many of us are happy to answer questions, so feel free to ask.

Lately I've been working on making better use of Many Face's patterns in the
tree.  I do not use large harvested diamond-shaped patterns in the tree like
most programs.  I use the old hand-entered pattern knowledge I put into Many
faces over the last 20 years.  These patterns are irregular shapes, so more
general, but there are many missing patterns, so the knowledge has more
holes.  The old many Faces engine with all the knowledge is single threaded,
which is a pretty severe bottleneck on modern machines.  I pulled the
pattern matcher out of the old engine and made it thread safe so I can use
it as an independent module in the tree.  Right now it is still too slow,
but once it is working well, I plan to do automatic tuning of the pattern
weights to replace the values I created by hand over the years.  Hopefully
that will give a jump in strength, since I can do large pattern matching at
all nodes in the tree.  Today I only run the old engine (which includes the
patterns) on nodes with at least 30 visits.

Is this the kind of discussion you were looking for?

David

> -----Original Message-----
> From: computer-go-bounces at dvandva.org [mailto:computer-go-
> bounces at dvandva.org] On Behalf Of Stefan Kaitschick
> Sent: Friday, November 15, 2013 1:13 AM
> To: computer-go at dvandva.org
> Subject: Re: [Computer-go] How many probes down the tree are necessary
> for a "good" bot?
> 
> When you do only a single playout, you dont even need a playout, just
> generate the first move candidate - voilá. :-)
> 
> Closed sources are indeed regretable at this point.
> Before, it sparked a kind of bot war, and the greatest technical
> advances are always made at war time.
> But now that the top bots are kind of treading water, looking for the
> next breakthrough, it's truly harmful.
> And my feeling is, that the attitude has also changed on the "academic",
> open source side.
> I only know that this forum used to present really interesting ideas,
> but now that is seldom the case.
> Possibly the problems have become too technical for general discussion.
> But more plausible to me is, that  programmers are trying to avoid a
> onesided disadvantage, and are discussing their ideas in private.
> Well, no use crying over spilt milk.(As I, admittedly, just have)
> 
> 
> Stefan
> 
> 
> 
> On Fri, Nov 15, 2013 at 12:59 AM, Darren Cook <darren at dcook.org> wrote:
> >>> ... I'm going by the "Measuring program strength" thread (Aug
> >>> 3013) where people are getting 50% against GnuGo with 1K to 10K
> >>> playouts.)
> >>
> >> Actually, in that thread Hiroshi was saying he only needs 350
> >> playouts, and Detlef was at 700. So it seems you're off by roughly a
> >> factor 10.
> >
> > Yes, I was surprised just how heavy the Aya (and Zen) playouts must
> be.
> > We've almost come full loop, and soon the programs will be traditional
> > computer go programs doing a single playout ;-)
> >
> > It is a shame the developer's skill is all dead knowledge (not open
> > source, no papers). :-(
> >
> > Darren
> >
> > _______________________________________________
> > Computer-go mailing list
> > Computer-go at dvandva.org
> > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
> _______________________________________________
> Computer-go mailing list
> Computer-go at dvandva.org
> http://dvandva.org/cgi-bin/mailman/listinfo/computer-go




More information about the Computer-go mailing list