[Computer-go] level of MCTS programs depending on the number of sims in 9x9 (unlimited time setting for humans)

"Ingo Althöfer" 3-Hirn-Verlag at gmx.de
Thu Nov 17 23:06:40 PST 2011


> Von: Olivier Teytaud <olivier.teytaud at lri.fr>
> Does anyone roughly know the level of a MCTS program in 9x9 depending on
> the number of simulations per move ?
> 
> I would roughly say:
> 500 simulations = KGS 10 kyu
> ...
> (please don't spend too much time on the 1000 reasons for which such a
> correspondance does not make too much sense,
> we all know it :-) )

I am interested in such data for very very small number of simulations.
In some artificial games and also in some natural games I found a basin 
structure when playing matches with pure Monte Carlo: Let MC(t) be
pure Monte Carlo which makes t runs for each candidate move. I look(ed)
at matches between MC(t) and MC(2t) for "all possible" values of t. Typically
there is some value t* "in the middle" where MC(t) performs worst against
MC(2t). There is a report on this at
http://www.althofer.de/monte-carlo-basins-althoefer.pdf

I tried to get it published. But some of the referees were very negative,
using the following arguments:
(i) There ae much better (and more refined) algorithms than pure Monte Carlo.
(ii) Small thinking times (where the basin typically occurs) are not
interesting for practice.
They did not understand the nature of basic research ...

And pure MC is sort of a building block for MCTS-type algorithms. So
it is important to know how pure MC behaves for small parameters.

Ingo.

 



-- 
Empfehlen Sie GMX DSL Ihren Freunden und Bekannten und wir
belohnen Sie mit bis zu 50,- Euro! https://freundschaftswerbung.gmx.de



More information about the Computer-go mailing list