Re: [computer-go] Pattern matching - example play

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

----- Original Message ----- 
From: "Darren Cook" <darren@xxxxxxxxxxxxxxxxx>
Subject: Re: [computer-go] Pattern matching - example play


>
> That is very interesting; do you have the graph online anywhere?

The learning process is not even finished yet (It will take 36 more hours or
so).
I rememebr from last year that the graph was a bathtub, but my
pro-prediction result was 17.5% as opposed to 46%..

Still I find much better prediction during the start so I assume it's the
same, a bathtub only with less steep sides.
I could make a graph available when I have the time but my work is not
geared towards publication, it is intended to help pay medical bills very
soon :)
So I prefer to do stuff that will yield a commercial result first.


> I mentioned a few weeks back how I thought 46% was impossible - there are
> just so many equally good moves on the board at any time. So it seemed
like
> a bug or over-training. Frank didn't follow-up; perhaps I'm in his kill
list?


When I saw the 45% figure I almost fell off my chair but I didn't believe it
at all, so I did manual checking of games, move by move and wrote the
results down on paper and then calculated the percentages with a calculator.

Overtraining was a problem with my previous approach (if I could get it
trained at all, it took weeks to train) but the current method works much
better and even when I feed it 450,000 ama moves, the overall pro prediction
stays almost exactly the same.

It means that a lot of pro patterns are different from ama patterns.
Pro's use Joseki and go very far with standardized variations.

I have found that my pattern database includes one million patterns that do
not occur at all in 52,000 pro games (I consider a 7d* a pro).


> However then I realized there is probably an urgent move most of the time
in
> the middlegame, so I realize a good performance in the middle game would
> bring up the overall average.


It is the worst in the middle game bec. it is very bad at tactics.
It really is a very good module for the first 120 moves.


> So the prediction graph (defined as number of times they guessed the next
> move correct on their first try) I would expect for a pro on a pro game
> would be 20% or so in the first few moves,

No it's ridiculously high at the first moves.
It hardly ever gets below 50% but often it goes much higher, like 70%.
The real strength is that it follows Joseki to a point that it's already the
beginning of mid-game.


> Have you noticed any correlations between game style and the prediction
rate
> in each phase of the game? E.g. by fuseki, by country of the players, by
> time limit, etc.

It predicts fuseki well because it acts as a real fuseki library, and it
always chooses the most requent fuseki continuation used.
Same with Joseki.

I am making automatic Joseki & Fuseki annotation based on referring to pro
games statistics (looking up patterns is so fast that I can annotate any
"joseki" position with "preferred patterns" and a bunch of diagrams and
winning / rank statistics in a few seconds).

About which nationalities etc. of players - This a topic I consider
classified :)
Well, seriously, I can tell you that the GoGoD dictionary of games is pretty
crappy for what I do with patterns.
Their collection does not contain "state-of-the-art" games enough.

I haven't tested on nationalities.
When my stuff becomes available I'm sure people will fool around with it
though..

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