[computer-go] thoughts on 100,000 cgos games

[David G Doshay]

On 8, May 2006, at 8:33 AM, steve uurtamo wrote:

>> Thus my comment that
>> I agree
>> that this can be used to point out a likely place to
>> inspect ...
>> but it still must be manually by a human.
>
> hm.  well, given that it can be noticed during
> a game, it seems at least that it'd be a great
> opportunity to throw some extra clock at the
> evaluation of the situation -- i mean, the move
> was unexpected and perhaps not in our favor,
> so ... pause to reflect.  that's what most people
> do, anyway.

I think I see your point now: if we thought we were
OK but now our evaluation shows that we have
fallen behind, spend more time on this or the next
few moves. If we had statistics that showed that
more cpu time or deeper search or more branching
worked better, then it would be easy to agree.

The very strange thing we are facing is that it does
not look like any of those things are true with the
evaluation functions we are using. What we really
seem to need is either tuning of our functions or
a whole different set of evaluation functions.

>> Also in your assumption is pruning of the lookahead
>> tree.
>
> my bad, i assumed that branches got trimmed if they
> fell below a certain value.

The right assumption to make in a serial alpha/beta
search. But on the cluster we have each lookahead on
a different cpu, all happening at the same time, and thus
we don't know what the pruning value would be until
they all return their answers. We could do multiple stage
searching that used the results from earlier stages, but
we have not tried that yet.

>> But in as far as your "with enough depth" comment, I
>> can assure
>> you that our results show very clearly that deeper
>> search results
>> in weaker play with the evaluation functions we are
>> using.
>
> this must imply that the likelihood that you can
> predict the response move of the opponent declines
> as the depth increases, which would make sense.

Yes, this seems to be the reason ... deeper search
makes decisions about moves based upon boards with
vanishing probability of really happening ... unless we
are playing against GNU Go, our underlying engine, and
in that case the deeper we look the greater our winning
percentage. We call this the "evil twin" effect.



Cheers,
David