Re: [computer-go] Territory & nfluence

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

> Once you have your current round of learning complete, I have another
> stat for you to calculate that I think would be interesting.  When
> ordering your moves, how many moves must be checked for each turn to
> ENSURE that the pro's move is always considered?


I am not sure I understand the question, but 90% of the pro's moves are
within the top-25 of moves.
All legal moves are always considered. There are 8 patterns around each
move, making 361 x 8 = 2888 patterns max. per position.

I just see that it will take 4 days for the new learned stuff to be
complete, if there is no crash.
The process generates, looks up, and processes statistics of 100,000
patterns/sec. of which 12,500 patterns are the size of the entire board,
12,500 are the size of half the board, etc. So 4 days is a bargain for
500,000 games :)



> Also, how many moves
> on average must be checked?

All legal moves on the board are always checked.
I could not even change that if I wanted to, and it would not matter much in
speed anyway.


> You will probably have some cases when
> the pro's move is ordered quite low by your matcher.

Yes, I am pretty sure this is mainly in case of complex tactical fights.
I will, in 4 days, publish the whole graph, a ranking from 1 to 361.


> Could you maybe
> post a couple of these positions?

Good idea, yes, thanks.
I will do that after 4 days and perhaps people can tell me whether it is
because of "deep tactics" or something else.
Very interesting.

>  My guess is that they would be
> interesting, unique tactical situations.

Yes. But you know, this pattern system has no concept *whatsoever* of even
the simplest tactics except of near-atari chains that are directly adjacent
to the point to play on.
I had to limit this due to "explosion"-avoidance and to keep the time needed
under control, and to have a statistically significant number of equal
patterns. It seems a lot but half a million games is not too much. The
difference in pro-prediction between 200,000 and 500,000 was 7% or
thereabouts. Has to do with:

- Very rare "joseki" than only becomes statistically significant after many
games
- The fact that pro's have different patterns than amateurs (to predict
medium-strong amateur games well, it is cneccessary to train with many of
those games)
- The fact that even when you include extremely rare patterns (1::100,000)
the statistics only become apparent after half a million or more games. Such
extremely rare patterns have a VAST influence on the performance as on
average, 40 moves per game are in that class!

Very simple: I have in tha database 4 million move patterns that occur only
once per 100,000 games.
That implies that the average game contains 4 million / 100,000 = 40 of such
patterns!

(This is a "secret"..)

And when you see such a pattern 5 times in half a million games, you know,
statistically speaking, enough about that pattern to be able to say with
some confidence what its value is.

This is very counter-intuitive, n'est-ce pas? You would never believe that
half of my pattern module's pro-prediction is due to ultra-rare patterns
that occur only once every 100,000 games.

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