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

["Eric Boesch" <ericboesch@xxxxxxxxxxx>]

Just to elaborate unnecessarily on a point that has been beaten to death 
already --

In chess, doing global search while pruning moves that are not severe 
threats may work.  In go, you have to be able to ignore severe threats that 
are not locally relevant, or else you face an unnecessary combinatorial 
explosion.  The go board gets littered with threats that demand an answer 
but where the exchange is not necessarily profitable -- that is, large ko 
threats.  If you can't even "search" a trivial three-ply capture without 
being distracted by the possibility that your opponent will interject a 
30-ply sequence of large ko threats, then you can't play on the 19x19 board. 
 Yet playing such ko threats is your opponent's prerogative, and failing to 
answer every one could lead to a lost position.  Of course, your opponent 
will not waste his 15 large ko threats (not so large a number for 19x19 go) 
just to delay the inevitable, so you will have the same problem next turn, 
and the turn after that.

In chess, you generally take profit as soon as you see it, because it's 
better to take the profit than to keep your pieces tied down threatening to 
do so.  In go, the pieces don't move anyway, so often you should spend your 
moves developing elsewhere.  With the ko rule and the immobility of go 
stones, the attacker and defender each have their own reasons to let ko 
threats accumulate.

High branching factors and persistent severe threats only aid humans in 
comparison to programs if the humans can handle them better.  They do -- 
successful programs on the 19x19 board all localize their search somehow, 
but I doubt any program does it as well as a typical 15-kyu beginner.    (If 
the program wins, it is because of the beginner's weakness in other areas.)

Ladders are the only global tactic in go, and even if theorists have proven 
that ladders can be very complex, centuries of go play have proven that 
practical ladders are almost always tactically simple.  Humans also try to 
localize their search in chess, but bishops, rooks, and queens can attacik 
four corners or sides of the board at once.  Even pawns, knights and kings 
are somewhat non-local in a negative sense, because (excluding pawn 
promotions) there is a limited number of them -- there are no immediate 
queen-side knight tactics if the knights are all on the king side.

_________________________________________________________________
Express yourself instantly with MSN Messenger! Download today - it's FREE! 
http://messenger.msn.click-url.com/go/onm00200471ave/direct/01/

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