[Computer-go] ambient temperature through the game.

Nick Wedd nick at maproom.co.uk
Sat Apr 10 04:36:27 PDT 2010

In message <E02C6EE5407D4D01BA9B42E34A9CD9A5 at drecksack>, Stefan 
Kaitschick <stefan.kaitschick at hamburg.de> writes
>> In the simplest model, ambient never increases. But in a case such as 
>>this, it would make sense to "restart".
>> The typical model in Go seems that there is large-scale fighting with 
>>very  high temperatures in the middle game, then it sooner or later 
>>settles into  the large endgame with a few moves at temperature of 
>>maybe 10, then drops  quickly. It would be interesting to do a 
>>large-scale study but we cannot  do this automatically yet.
>> Here is what I know about real-life endgame analysis using temperatures:
>> Berlekamp, Spight and their students have analyzed a number of 
>>difficult  human endgames by using combinatorial game theory. Their 
>>analyses usually  uncover a number of mistakes even in top 
>>professional play.
>> In Berlekamp's "environmental Go", the values of moves on the board 
>>must  constantly be compared to a stack of "coupons", simple gote 
>>plays worth a  decreasing number of points. There have been several 
>>matches where  professionals play environmental Go. One such endgame 
>>is analyzed in great  depth by Spight at
>> http://www.msri.org/publications/books/Book42/files/spight.pdf

It is possible to calculate the importance of tedomari.

Suppose we play a game ignoring tedomari.  As the value of sente drops 
from (about) 14 to 0, which will each get about seven points' worth of 
tedomari effect, though we won't be aware of it.  Now suppose we play 
another game, in which you manage the tedomari perfectly, while I ignore 
it.  You will gain all 14 points' worth.  Thus you will win the second 
game by 14 more points than the first.  So the value of understanding 
and applying tedomari perfectly is twice the correct value of komi, or 
the marginal value of one handicap stone.

Nick Wedd    nick at maproom.co.uk

More information about the Computer-go mailing list