[Computer-go] Should playout patterns be gathered from high-dan games only?

Brian Sheppard sheppardco at aol.com
Wed Jan 16 13:03:43 PST 2013


> Is it enough to use big number of hign-dan games to build a strong 
> pattern library for random playouts?

When I address large-scale patterns, I will use the following principles:

1) Select examples only from the program's own games, to ensure that
everything that is learned is actually relevant.

2) Select examples only from the moves of the side that won, to ensure that
the lessons learned are asymptotically consistent.

3) Select examples only from the moves where the existing rules predicted
something else, to ensure that we don't overemphasize obvious moves.

4) Continue to sample indefinitely, to ensure that learning always
continues.

Best,
Brian






More information about the Computer-go mailing list