[Computer-go] Refutation of useless ko threats during random playout

Alexander Kozlovsky alexander.kozlovsky at gmail.com
Wed Jan 16 07:53:43 PST 2013


Hi everybody!

Some times ago the was a thread about bots predisposition
to silly ko threats
http://dvandva.org/pipermail/computer-go/2012-March/004860.html

and the common opinions were:

> Whether a move is a ko threat and whether it is played as a
> ko threat needs to be detected presumably by expert reading.

and

> You almost need to determine in the playout how to refute a
> stupid ko threat but I don't think that is simple at all.

Can anybody explain why it is so hard to find good response
to stupid ko thread during playout?

I think the next may be enough:

1) Analyze large corpus of professional and semi-professional
games and found all move sequences of length 4 with the type of
"ko - move1 - move2 - ko". With such sequence is is very probably
the move1 is ko-threat and move2 is proper refutation.

2) Include patterns for all move1-type moves in patterns library
used for random playouts. Set flag for all such moves
"this is ko-threat".

3) Remember offset of move2 in pattern data for move1.
Or, as alternative, add all move2-type moves into pattern
library and set flag "this is ko-thread refutation"

4) During random playout, when next candidate move is
analyzed, reduce probability of moves when they match
"ko-threat" patterns and previous move was not ko.
Also, increase probability of moves when they match
"ko-threat refutation" pattern

Any opinions, is this plan viable?



More information about the Computer-go mailing list