>
>
> MC is playing most "goal-directed" ("zielgerichtet"
> in German) when the position is balanced or when
> the side of MC is slightly behind. However, when
> MC is clearly ahead or clearly behind it is playing rather
> lazy.
At some point we were investigating that here, but only on small sets of
games (we have essentially games with no handicap
between versions which are close to each other - for example, the conclusion
according to which white has the advantage when using komi 7.5 in 9x9 is
more documented than questions around handicap). I find the following
elements interesting:
1) MoGo was seemingly weaker in handicap games ("was" because we did not
have a look at that recently - perhaps it's always true). This was nearly
established by comparing the success rate against people of level X with
handicap n and people of level Y with handicap p. Of course, this was not
very scientific - just an element.
2) MoGo has become stronger (and the difference is huge for long time
settings) with more randomization; I guess this is because with more
randomization, you are less likely to have a constant result as you increase
the number of simulations. With more randomization you are more likely to
see that the x<<50 % of probability of loosing becomes y<<x<<50% with a
given move than if you have a too much deterministic simulations such that
y=x because the few simulations in which the move introduces a difference
occur with very very very samll probability in your too deterministic
simulations. By the way, this reasonning is also intuitively satisfactory
for explaining that in the MC simulations uniformity on reasonnable moves is
often better than non-uniform probabilities - with uniform probabilities,
the less likely event is more likely than with non-uniform probabilities.
This reasonning is not mathematically sound, of course, it's not a theorem,
but intuitively this looks reasonnable (at least for me :-) ). This strong
improvement with randomization might be also an improvement in the case of
handicap games, but I've not checked that.
3) By the way, we have positive results (small improvements, but
improvements) by using different Monte-Carlo simulations for different
cores. This is consistent with 2) above - by performing N simulations of
type 1 and M simulations of type 2, you cover more
markov models than with one Markov model - you can cover different
scenarios. I'd like to extend this idea - for the moment
we have only two Monte-Carlo simulators and the improvement is small, I hope
we can define 3,4,5,etc Monte-Carlo
simulators and have a strong improvement.
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
