[computer-go] More statistics and conclusions from CGOS data

[Gunnar Farnebäck]

Christoph Birk wrote:
> On Thu, 14 Feb 2008, Gunnar Farnebäck wrote:
>>> Interesting. If I do the same with MonteGNU's fuseki database, which
>>> is based on online learning from own CGOS games, and cut off at 200
>>> samples I get:
>> E5  8101
>> |  C3  2950
>> |   |  G5  1798
>> |   |   |  G3  1145 (A)
> 
> And the results (win/loss %) ?

Due to the way the learning is done and because wins against weaker 
opponents are excluded, the win/loss statistics don't mean anything and 
are in fact close to 50% for most popular moves. Look at the numbers 
above as indications of how often the corresponding move historically 
has seemed like a good idea to play for the fuseki algorithm.

/Gunnar