Re: computer-go: Results on self play vs recorded games

[Markus Enzenberger <compgo@xxxxxxxxxxxxxxxxxxxxx>]

On Wednesday 20 August 2003 10:16, Peter Drake wrote:
> We've got some results for 9x9 Go, trained through
> combinations of 3000 games of self play and the 3205

self-play usually takes at least 1-2 orders of magnitude 
more games than learning from existing games.

Did your learning converge? In my experiments the playing 
strength of the networks (tested against GnuGo) is still 
increasing even if the final and temporal errors do not 
change much any more. So it is useful to check the playing 
strength at regular intervals during training to decide 
when to stop the training.

What exploration strategy did you use?

Also 4 games is surely not enough statistics. In my tests I 
usually play 100 games on 9x9 to estimate the playing 
strength and the standard error of the score after sorting 
out duplicates and calculating the variance of the results 
is still about 5 points. Probably even more, because of 
partial correlataions in the games.

- Markus