[Computer-go] Semeais

[terry mcintyre]

From: Kahn Jonas <jonas.kahn at math.u-psud.fr>

>I mean: play out the semeai before the MC playout. At least till the
>point the MC playout always gets it right (say 2 libs).
> Of course this biases the simulations in favour of the winner of the
> semeai, since it adds stones, but I wonder if the bias is very strong.


There are games where it is entirely reasonable to give up one semeai in favor 
of some other benefit, such as a ko or semeai. 

Wouldn't the "play out the semeai before the MC playout" strategy fail in those 
circumstances?

Thinking about the "favor moves which appropriately defend the semeai" strategy 
- which I have advocated - it seems that such bias should not be a move-by-move 
bias, but a once-per-playout or once-per-chain-of-exchanges bias. 

That is, some of the playouts might randomly choose not to defend the semeai - 
and consistently follow that policy. This is is more reasonable ( more 
reflective of the likely outcome ) than a playout which ignores the first 
critical approach move, then foolishly defends a group which is lost. Modulo ko 
threats, of course.

No wonder this problem is tough to solve. I am going to make a meta-suggestion: 
lock your best programmers in a room with a very strong player and a select 
sample of games where your program was swindled out of victory. Send in their 
favorite caffienated beverages and takeout food until they beat this problem 
into submission. You'll know you've won when your program stops making kyu-level 
mistakes and the go player is sweating.


      
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://computer-go.org/pipermail/computer-go/attachments/20110115/32efc0fd/attachment.html>