I definitely see what you mean.

I'm not sure if this is a feasible idea, and perhaps it has been thought and discarded before. But perhaps one could use libkombilo (are there perhaps alternatives?) to do fuseki matching against a large pro database. Pretty soon there will be deviations from the pro-path and no hits returned. Then one could perhaps do half board searches, and when they dry out quarter board searches. I believe this would:

1. make a great fuseki book
2. find all josekis
3. find additional local pro strength moves.

Libkombilo uses sqlite under the hood and sqlite has the option to load the database in-memory instead of from disc every search. Perhaps this can be used to speed up things since such searches are pretty heavy. The memory usage would increase significantly. I'm just half-guessing but I think a pro database of 50000 games would not need more than 70Mb.

Regards
Johan




>Hello Johan,
>
>I guess the size of the book is the main problem. If you assume an
>opening of say 40 moves and maybe 20 reasonable moves on average, the
>size is 20^40. And there is no guarantee that the opponent will stick
>to one of the 20 moves that we think are reasonable.
>
>There is a large number of moves in the opening that are not hugely
>bad, just a little strange. Maybe lose 1-2 points. Whereas if you can
>improve the general fighting strength of the program you hope to gain
>many more points. I guess that's why I don't consider it high priority.
>
>On 9x9 things are different. Most games are decided by move 30, so we
>need to play only 15 moves until then. If 3 of them come out of the
>book, we save 20% of time (actually more since our time control
>decreases exponentially with move number).
>
>Also important on 9x9 is that I have a bigger chance of directing the
>game away from positions that Fuego does not play well. So I choose a
>small umber of openings and then add moves to those if the program
>loses games because of the opening
>
>For the big board, I think a joseki book (standard sequences in
>corner, maybe also on the side) would be useful. My old program
>Explorer has one. But you have to think about how to use it. E.g. give
>a bonus of some number of simulation wins to that move.
>
>Martin