Re: [gnugo-devel] A few questions

[Paul Pogonyshev]

address@hidden wrote:
> [setup description]
>     Now that all of the preliminary information is out of the way, here is
> the question. Does increasing the computational settings cause GG to look
> at more and or different initial moves? If not, why isn't a basic tesuji
> play in the list? I think it takes 33 moves to capture the group although
> it is obviously dead after five moves. I just started a last test run that
> I expect to take three days with settings that I think should enable GG to
> compute the death of the group to the end. But that is only a guess based
> upon the documentation and my feeble Go skills.

No, raising the level will never make GNU Go a 9 dan program, not even shodan.
GNU Go picks moves based on coded human knowledge.  It is present in two forms:
patterns (i.e. "in this or similar positions try this move") and heuristical
algorithms (e.g. "find such and such string and try to attack at that liberty".)
If neither of them finds the first move of the sequence, the move is just never
tried, no matter how high the level is.

Various settings cause GNU Go search deeper and wider (think like in chess
here.)  However, if in given position it finds 3 candidate moves with patterns
and algorithm, it will never consider more than 3, even if branching limit is
10.  This description is simplified, but it should explain why increasing level
doesn't solve all problems.  If you told GNU Go that move was worth trying, it
would likely find that the move was indeed good and pick it.

If you can send us the game as an SGF file, we can probably try to teach GNU Go
this particular tesuji.

Paul


P.S.  Please don't post to this or other mailing lists in HTML.