[Computer-go] Combinatorics of Go

Erik van der Werf erikvanderwerf at gmail.com
Tue Jan 4 02:43:38 PST 2011


On Tue, Jan 4, 2011 at 7:49 AM, Robert Jasiek <jasiek at snafu.de> wrote:
> On 04.01.2011 00:30, Erik van der Werf wrote:
>>
>> The word 'almost' to me suggests that you would know for sure that
>> there exists an exception.
>
> So far you have given only heuristics hoping that your long cycles would
be
> detected somehow without comparing the positions.

I think here is our misunderstanding: When inferring the value of a
position, and the best move, the search does compare positions (it is
impossible to search meaningfully without somehow comparing positions).

The information that is used, e.g., to store & retrieve results from a hash
table, is the position plus the extra features I mentioned. What is not
needed to be stored in the hash is the position plus previous positions (and
that's why I say that the full history is not needed under Japanese ko
rules).


> Your program is White to move.
>
> . O . # O . #
> # # # # O O O
> . . # # O O .
> # # # . . O O
> # # # # # O O
> # # # O O O O
> # # # . . O O
>
>
> 1 O 2 # O 3 #    4 = pass
> # # # # O O O
> . . # # O O .
> # # # . 6 O O
> # # # # # O O
> # # # O O O O
> # # # 5 . O O
>
>
> With your program's status criteria (especially "number of consecutive
> passes", "number of Black/White captures", "number of Black/White passes
> (not necessarily consecutive)"), after move 4 your program is in danger of
> confusing the move-sequence with a sending-2-returning-1 cycle.

Once Black plays 2 (and White 3) the position has changed (in this case it's
even irreversible, assuming optimal play by at least one side); at this
point any previous position has become irrelevant for inferring the best
move (i.e., history can be discarded).

Best,
Erik


PS I just realized that the feature "number of consecutive passes" may be a
bit confusing; probably I should have stated "number of passes after the
last preceding play" (because there can also be a single pass).
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://computer-go.org/pipermail/computer-go/attachments/20110104/5edeeec6/attachment.html>


More information about the Computer-go mailing list