Dan wrote: > (1) a tendency to make bad tenuki's. This problem is > not special to the opening, but it is often in the > opening that GNU Go plays away from an urgent position > to take a big point somewhere else.
But it's worth noting that GNU Go doesn't play tenuki as such, in the sense that it decides that it doesn't want to respond to the opponent's last move. In fact the location of the opponent's last move is not taken into consideration at all, except to some extent in the generation of moves to subdue a thrashing dragon. Of course when GNU Go misevaluates a position and incorrectly thinks that a move somewhere else than near the opponent's last move is bigger, it will in effect be an inappropriate tenuki. > (2) GNU Go doesn't really understand the importance of > making a base. (That is, an extension along the side > where eyes are most easily gotten.) If a stone > is approached from one side, extending on the other > is reflexive but GNU Go doesn't understand this > principle. The only problem is that this cannot be too strong a reflex but has to be quantified accurately. > In case it is not clear what I mean, in the > following example, moves 6, 9 and 12 are important > moves because they are extensions from a stone that > is approached on one side to create a base. Of > course GNU Go would play 9 (because it is Joseki). > But that is beside the point. GNU Go doesn't really > understand this principle. In move 6 (response to low kakari against hoshi) there are of course also the options of denying the opponent a base and play a pincer, or tenuki and allow a double kakari if there is sufficient gain elsewhere. The fuseki database contains two extensions, four pincers and no tenuki in this full board position. Move 9 is found in the fuseki database, as well as being joseki. Without fuseki and joseki databases GNU Go would prefer F5, with the joseki move J3 as second option. Move 12 has left the fuseki database: A B C D E F G H J K L M N O P Q R S T 19 . . . . . . . . . . . . . . . . . . . 19 18 . . . . . . . . . . . . . . . . . . . 18 17 . . . . . . . . . . . . . . . . . . . 17 16 . . . O . . . . . + . . . . . X . . . 16 15 . . . . . . . . . . . . . . . . . . . 15 14 . . . . . . . . . . . . . . . . . . . 14 13 . . . . . . . . . . . . . . . . . . . 13 12 . . . . . . . . . . . . . . . . . . . 12 11 . . . . . . . . . . . . . . . . . . . 11 10 . . . + . . . . . + . . . . . + O . . 10 9 . . . . . . . . . . . . . . . . . . . 9 8 . . . . . . . . . . . . . . . . . . . 8 7 . . . . . . . . . . . . . . . . X . . 7 6 . . O . . . . . . . . . . . . . . . . 6 5 . . . . . . . . . . . . . . . . . . . 5 4 . . . O . . . . . + . . . . . X . . . 4 3 . . O . . X . . X . . . . . . . . . . 3 2 . . . X . . . . . . . . . . . . . . . 2 1 . . . . . . . . . . . . . . . . . . . 1 A B C D E F G H J K L M N O P Q R S T To some extent GNU Go does understand this position. R13 is matched by pattern F401 which adds a strategical defense of R10 move reason. The generated move at N4 is thought to erase much of black's moyo but is also given a penalty for being an invasion. The problem is with the move valuation, in particular the strategical values, e.g. the value R13 is awarded for strengthening R10. One problem is that the weakness values associated with dragons are not very reliable in early positions. For example both D16 and R10 have the same weakness value 0.39 but clearly R10 is in more danger than D16. Another problem, and this goes way back to GNU Go 3.0, is that the effectiveness of a strategical attack or defense move is not estimated. The contribution to the move value is only dependent on dragon size and dragon weakness. This is one of the most urgent areas for improvement of GNU Go (not only with respect to fuseki). /Gunnar _______________________________________________ gnugo-devel mailing list [email protected] http://lists.gnu.org/mailman/listinfo/gnugo-devel
