Thanks to someone pointing out the info source for UCT method. I did some 
reading.
 
UCT is a width first search algorithm. Some of the MoGo's plays are not bad. 
Especially its cohesion and emphasizing the center are good, which are the 
common weaknesses of most Go programs.
 
To my understanding UCT-MC is a statistical method. The reason for this is that 
the evaluation method is statistical, the score propagation up the nodes is 
statistical, and the choice of next node is statistical. Nothing wrong with 
statistical methods, many of which can give very accurate answers. What one 
need to pay attention to a statistical method is its accuracy. Since it's a 
statistical method, one has to take a grain of salt with the statement that 
with enough given computing power and time it will found the eaxct right answer.
 
Vital to the UCT method is the ability of evaluate the board at any given move. 
Here it comes the importance of the MC score method. It's a statistical method. 
Its answer has random errors. To measure its error I think one way is to 
specify the number of stones on the board. Write this number as n. Then the MC 
score can be written as M(n). It makes sense to measure the error of M(n) for 
each value of n. For different n the errors and their distributions are 
apparently different.  The error of M(n) comes from two sources. One is 
intrinsic, due to its random number generation. The other one is due to the 
board position variations for a given n. The distribution of the first error is 
gaussian. The distriution the second error is unknown,but can be assumed as 
gaussian. Even we know it's gaussian, we still need to know it's variance. or 
standard deviation. Once the error distributions are known, it can help to 
evaluate and fine tuning the UCT method.
 
I suspect the accuracy of M(n) increase quickly with decreasing n. This makes 
the UCT method very effective for localized problems, such as tsumigo and local 
tactical readings. However, we do know there are heuristics that can solve the 
tsumigo very quickly.
 
The accuracy of M(n) (at least for some n) can be evaluated using the database 
of pro player games. 
 
 
 
Daniel Liu
 
 
________________________________________________________________________
AOL now offers free email to everyone.  Find out more about what's free from 
AOL at AOL.com.
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/

Reply via email to