'castrobot' needs some more training lessons...

Looks like 'castrobot' needs some more training lessons,

I just won against him straightforward (http://www.littlegolem.net/jsp/game/game.jsp?gid=1356217&nmove=11). He did not at all realize that that ring wins, even though he tried to make rings himself later.

Probably that is very easy to repeat?

That's a clear weakness indeed. It still could have cut the six 'frame' from within, but that also escaped its algorithm. Timo just implemented improvements in the realm of rings, and they are noticable if it attacks, but defensively it doesn't show the same behaviour yet. I won base-7, -8 and -9 today.

"It still could have cut the six 'frame' from within, but that also escaped its algorithm."

Not that that's so surprising in an algorithm that is based on random play-outs. Though the order isn't relevant, getting 5 white stones inside the 6 black ones isn't a frequent occurence in random play.

That is a known weakness. It tends to think you're trying to setup a whole bunch of small ring threats, hence move 14 in the center. Frames are easy for humans to recognize, but hard for computers to recognize. The current version is experimental. I've been trying to get it to stop playing so many ring threats and instead go for the fork and bridge wins, but it's hard to get it to recognize what is a worthy ring threat and what's a stupid ring threat.

I just noticed, Castro just won a game by a framed ring: http://www.littlegolem.net/jsp/game/game.jsp?gid=1356219&nmove=36 , so despite making a ton of pointless ring threats, it managed to use them to travel across the board and even construct a real ring. Some reasonable counting shows Castro was a similar distance to a win despite his opponent having most of a frame setup already.