computer-go: Results on self play vs recorded games

[Peter Drake <drake@xxxxxxxxxx>]

We've got some results for 9x9 Go, trained through combinations of 3000 
games of self play and the 3205 non-handicapped recorded games from a 
set provided by Nici Shraudolph.

After training, each network played one game as black and one as white 
against each of the other four.  Here are the total Chinese scores 
(with 0 being a tie):

Untrained network:  -276
Self play only: -216
Recorded games only: 38
Self play followed by recorded games: 148
Recorded games followed by self play: 306

I believe recorded games push the network into the right global region 
in the space of possible strategies, while self play provides fine 
tuning.


Peter Drake
Assistant Professor of Computer Science
Lewis & Clark College
http://www.lclark.edu/~drake/