[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [computer-go] Pattern matching - example play
At 11:59 AM 12/3/04, you wrote:
----- Original Message -----
From: "Ray Tayek" <rtayek@xxxxxxxxxxxxxxxxx>
Subject: Re: [computer-go] Pattern matching - example play
> your prediction rate sounds great. does it vary much between the beginning,
> middle and end stages of the game?
Yes.
The prediction graph looks like a bathtub.
At the start, the system's Fuseki patterns are used.
Then the Joskie patterns are used and a whole lot of patterns that are not
striclt Joseki but large patterns that are common in the first 100 moves.
this makes sense. but i would expect a measurably higher prediction rate
for pro's as opposed to amateurs (i think the pro's are better at the
"timing" - but in this case it would be when one could leave a joseki in
one corner to make an "asking" move in another corner joseki where you had
left a move unanswered.
The more "tactical" a game is with (for me) an "erratic" game development (I
mean games that not are about a slow building up of equilibrium
i think i would call these "small group to small group" style - the
traditional japanese style (as opposed to some of the "cosmic" "moyo" or
large scale style or fighting style)
but are more
like one big exploding fight), the less well the pattern system predicts. I
have seen prediction rates as low as 19% for such games.
that makes sense as big fights early in the game have a high variance of
outcomes.
At the end, the pattern system also works well because end patterns are
about filling dame etc. Well-established ways to finish up things. So I
often see a 60% prediction rate in the first 100 moves or so.
this makes sense also, as endgame tesugi are fairly standard.
the bathtub makes sense also as nobody understands the mid-game.
> sure. but i suspect that most of them were trying to be helpful.
Nope. Most of them are trying to be asses. ...
The helpful people can be counted on the fingers of one hand, the rest has
made it a mission to tell me that I need to shut up until I am a good Go
player, or that I will be a failure with a Go program because I have made
the world's best Joseki expert system.
i can understand that and would probably have given you the same advice had
i not jumped into the middle of this thread.
but there is one aspect of go that might be useful and that is timing and
asking moves. let me try and draw an analogy from chess (i know how the
pieces move, but i don't know how to play, so we're sort of in the same
boat in this respect.
suppose there is a sequence on one side of the chess board where the person
(say white) who initiates a trade is has a choice of two branches, both
yielding about the same value. and the opponent (black) has some choice as
to how to answer the sequence (but can't avoid either one). (i might call
this an "asking" sequence"). also let's suppose that the initiator can
move a rook (or something) to one spot or another (both of which are big
threats and must be answered immediately). so white would start the
sequence until black made one of his choices. then white would move his
rook to the most advantageous place, bkack would be forced to respond and
white finished the sequence with his rook in a better place. this sort of
thing has to occur often enough in chess that there must be a name for this
in chess, but i don't know what it is.
moving to the go board. suppose we start a game where black makes the first
4 moves at the 3-4 points in the corner and white immediately approaches
that corner with the standard knight's move approach. now black can start
up 4 corner joseki's. the choice of which joseki to choose and when to
switch to another corner is complicated. (this is why i asked the original
question about the order of moves).
i have heard that the go pro's that play shogi (japanese chess) say that
playing go is about like playing 6 games of shogi at once. so the frequency
of these "asking" moves/sequence and their timing is extremely important in go.
i wonder how your predictor would do if it ordered the moves first by the
corner area and then by the original move order. things like "ladder
breakers" http://senseis.xmp.net/?LadderBreaker would mess it up (but maybe
you could identify them?) but it (the predictor) might do much better in
certain classes of games.
i''m not an ai person (just a recycled mathematician who has become an old
programmer :), but i wonder if the prediction rate of your system could
classify a game into one of "small group to small group", "cosmic/moyo" or
"big fight from the get go" or maybe "many fights left unsettled". this
style should be mostly apparent in the first few moves ( say from a small
number 5-10 to a larger number based on how many josekis; get played out
and to what extent?) this might aid some heuristics? (i hope to write a go
program some day driven mostly by some heuristics during the beginning and
early midgame).
also, one thing i learned *really* late in my go playing about joseki's is
that if there is a stone on the side (say at 3-10 point or even a little
farther away and a standard corner approach move is made against a (say a
3-4) corner point (from the opposite side of the stone at the 3-10 point).
and the defender of the corner (or the attacker) plays out a joseki, then
this is usually wrong as it will turn out to be a very obviously unequal
exchange or somebody will have one really badly placed stone (probably
overconcentrated). so i wonder if the presence of a stone near the 3-10
point or not would be a predictor of how well your predictor works in that
corner? i also suspect that the color of the stones in the adjacient
corners that are closest to the corner in which the joseki is being played
out would also affect your predictor.
thanks
---
ray tayek <http://tayek.com/>, co-chair <http://www.ocjug.org/>, actively
seeking telecommuting work.
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/