[computer-go] Selectivity

[Łukasz Lew]

We all know Remi's selective search algorithm.

Again, I would like to thank him for sharing his ideas.
It's easy to see how deep impact it had on our community.

Then UCT came, and everybody with a MC Go program advanced enough, tried it.
Maybe MoGo is using it also.

Remi told us (thanks again) that he increased exploration parameter 10
times and it was good enough to beat his previous algorithm.

But looking at CGOS results one can see that there are many variants
with different performance. (for instance Valkyria (go! Valkyria! go!
:) ) )

I can tell you only about level-1 selectivity as I do not have a deep
search in my program.
(BTW Talking about 1 year old program - LeGoBot, now abandoned)
The best results were achieved not by any combination of  number of
simulations played and current value, but explicitly pruning worst
moves depending on time left.
I.e. after 1/3 of time I was sampling 40 best moves
1/2 of time per move I was sampling only from 15 best moves,
a lot of time at the end was spent on 3 and 2 best moves.

This worked really well with a small resources.

So let's discuss our variants of search. And ways to implement selectivity!

Best Regards,
Łukasz Lew