Re: [computer-go] Learning : was Chess programs versus go programs

["Frank de Groot" <frank@xxxxxxxxxxxxx>]

----- Original Message ----- 
From: "Matt Gokey" <mgokey@xxxxxxxxxxxxxxxxx>
Subject: Re: [computer-go] Learning : was Chess programs versus go programs


>Vincent, I don't necessarily disagree with you (about search).

I also fully agree with Vincent.
In fact he posts very interesting articles.
He just (and many others) seem to think that when a go programmer builds one
module using a certain approach, that because that approach is
non-conventional, that this programmer much be "naive" and will try to build
all the other modules using the same approach. In fact I will use the
appropriate techniques for each module, using extreme prejudice. When I
started this pattern system, everybody told me I should not do it (not even
for Joseki). They said things like: "Fusek and Joseki are not so important".
Imagine how much determination you must have to spend 2 years on a module
when the best Go programmers and very strong Go players tell you things
like: "This has no use, this will not work" etc.. :)

Of course, now my system predicts almost 50% of pro moves, I am a bit more
sceptical to those "experts" opinions. Especially because the best Go
programs are unable to play themselves out of a wet paper bag.

>There is something appealing about the objective and statistical nature of
the pattern harvesting technique that Frank has developed >over the ad-hoc
and "expert" tuned pattern/rule sets commonly used.  I would not discount it
as a useful part of a go-playing program.

I have fixed a bug (It saw only 10% of all positions fed) and even after
one-third of learning, the pro-prediction is above 48% instead of 46% on the
sample game I posted, meaning 4 more moves guessed correctly. I expect the
final system to attain an average of 50% on unseen pro games.

When modern Fuseki & Joseki deviates too much from the old, the system would
need a week of re-training. But computers are getting faster so in the
future this may be done overnight. Game records may be online by that time
(perhaps from my own website). O r it can be done incrementally, in that
case current PC's could do it overnight already.


>As Frank has said it's designed to be a better Joseki/Fuseki database, not
the holy grail.  There is nothing preventing combining it >with other
modules using other techniques for handling some tactical issues, local
search, ladders, ko fights, life/death analysis, etc.

As the pattern system is finished now (I have yesterday made a more
"scientific" way of displaying the pattern data), I will spend time on
evaluation & search. I wanted to do Fuseki & Joseki first, as people told me
"in the start of the game, the biggest moves are made". I also thought I
would be needing at least 2 years to familiarize myself with L&D issues.

My work now is to put as much domain-knowledge into semi-static evaluation
as possible. I am interested in methods that use very narrow (directed)
search to estimate whether a tactical situation is lost, won or balanced.
Those methods are described in beginners' books and locally, they are
correct, as opposed to the heuristics used in most other programs.

As a non-Go player, I am more interested in approaches that guarantee the
highest level of correctness, and I will speed those approaches up so that
they can still be used instead of heuristics.

_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/