I have similar experiences.Two liberty blunder appears after fixing one liberty blunder. And 3-lib blunder appears after fixing two liberty blunder.
I have one test case similar as following diagram. All other places are mostly settled, except the bottom three string. O O X X X O O O X X O O O X . . . X . O O . X . - - - - - - - - In most cases, StoneGrid's playout ends up with the left bottom corner being killed. With self-atari analysis, either side will not play self atari moves, which is right behavior. However, without analysis on the 3-lib string, the chance for the black to find the right move in the random playout is very small. It is far less than 1/2. As when black plays randomly at one of its two outside libs, it will bring the white to play around it, and end up with black being killed. On Fri, Dec 3, 2010 at 10:45 AM, Brian Sheppard <[email protected]> wrote: > >Is the problem as bad as in Fuego? > > Maybe not this specific problem, but similar problems. > > Approaches like adding a 2-liberty feature helped Pebbles. But the > difference is not as large as you would like. > > For example, your program will shred other programs that lack a 2-liberty > feature. But humans have a 2-liberty feature, and a 3-liberty feature, and > so on. And if the human is of sufficient rank, then his error rate on those > features is quite small. > > So you still run into problem cases. It just happens one move later than > before. > > Note that the combinatorics suggest that 3-liberty features will be very > hard to write correctly. A 2-liberty feature is mostly branch-free. E.g., > one player ataris, and the response is to capture if possible and escape > otherwise. When you have 3-liberty features, there are now 2 ways to > escape. > Additionally, you can defend by capturing or by reducing an opponent to 1 > liberty. > > Even the 2-liberty feature is hard to get right. I expect to be fixing mine > for many years. > > Brian > > > _______________________________________________ > Computer-go mailing list > [email protected] > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go >
_______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
