[Computer-go] Are 4 'easy to avoid errors' common to all MC programs?

[terry mcintyre]

Testing thousands of random games is not enough; one must test games where the 
new feature makes a difference - both pro and con.

A feature (such as some versions of dynamic komi) which applies to handicap 
games should be tested on handicap games, not on even games. It should, of 
course, also be designed to have no effect, or at least no negative effect on 
even games. 

Likewise with features designed to apply to even games, which are meant to 
improve endgame play - they can be designed to yield "no worse" results than the 
existing state of art, and should be tested against many endgame situations. 
Random games are fine, but the testing regime should include specific game 
positions which are known to break other programs, including previous versions 
of the same program. 


      
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://computer-go.org/pipermail/computer-go/attachments/20110124/213ce7e2/attachment.html>