[Computer-go] 7.0 Komi and weird deep search result

Erik van der Werf erikvanderwerf at gmail.com
Mon Apr 4 05:48:42 PDT 2011


Even without exploration. With N going to infinity and, I suppose, the
implicitly assumed infinite memory to store results with infinite
precision, thus eventually representing the complete tree, all you
have to do is not expand solved branches. I don't see how that would
be different for 3-valued spaces. Anyway, that's just plain old
minimax; such arguments probably have very little to do with actual
performance.

I don't know about other programs but Steenvreter does have an
exploration term. My impression is that ignoring the fact that for
integer komi the distribution is not binomial makes it explore too
much rather than too little.

Erik


On Mon, Apr 4, 2011 at 2:01 PM, Michael Williams
<michaelwilliams75 at gmail.com> wrote:
> Yeah.  As N goes to infinity, MCTS+RAVE goes to MCTS.  So it sems like it
> would be guaranteed to converge only if you had an exploration term in the
> MCTS.
>
>
> On Sun, Apr 3, 2011 at 10:53 PM, Petr Baudis <pasky at ucw.cz> wrote:
>>
>> On Sun, Apr 03, 2011 at 04:21:10PM -0400, Brian Sheppard wrote:
>> > >AIUI, RAVE without special modifications (like those done in Mogo
>> > > later)
>> > > does not have any convergence guarantees either.
>> >
>> > MCTS using RAVE prioritization *does* converge to game theoretic values
>> > in a
>> > binary-valued space.
>>
>> I do not think this is true, at least for the general space of
>> simulation functions, as they may entirely shun the winning move.
>> Can you reference some more detailed analysis claiming this?
>>
>> --
>>                                Petr "Pasky" Baudis
>> UNIX is user friendly, it's just picky about who its friends are.
>> _______________________________________________
>> Computer-go mailing list
>> Computer-go at dvandva.org
>> http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
>
>
> _______________________________________________
> Computer-go mailing list
> Computer-go at dvandva.org
> http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
>



More information about the Computer-go mailing list