[computer-go] Re: Should 9x9 komi be 8.0 ?]

[Don Dailey]

Magnus Persson wrote:
> Quoting Mark Boon <tesujisoftware at gmail.com>:
>
>>
>> On 3-mrt-08, at 18:43, Don Dailey wrote:
>>
>>> I base that logic on my observations that once the score goes below 
>>> 10%
>>> for Lazarus, it is losing.   It's extremely rare that it salvages a 
>>> game
>>> once the score goes below even 20%.
>>
>> In which case I could argue that attempts at winning by playing
>> 'silly' moves are not working either. It hardly seems like an  argument
>> supporting your position. It may in fact make things worse.  This is a
>> feeling I get when I see MC programs play, as soon as one  falls behind
>> the loss is made irreversible by some extremely silly  moves. But this
>> is more the case early in the game, not by the end of  the game.
>
> The behavior of Valkyria as I see it is that it actually plays
> patiently up to the point where it knows for sure it has lost. When
> this happens it actually plays moves that makes the game as long as
> possible. These moves may look completely crazy, sometime it is just
> odd forcing moves. Against weaker programs this works quite often, and
> against stronger programs of course rarely.
>
> Often when I investigate "bad moves" in the endgame, it is clear that
> "normal" moves would had simplified the game and end in a clear loss.
> In those situations Valkyria should have started playing crazy even
> earlier.
>
> Also my impression is that MC-programs always plays "better" moves to
> a human eye if it searches deeper even in lost positions. It might
> still be crazy moves but better crazy moves.
Yes, I agree with this.    In fact, I have changed my mind about whether
a super-strong program would play badly if it was on the "wrong" side of
komi.     Let's imaging that a program plays with 7.5 komi and that the
opening position is a win for black,  just for the sake of argument.    
Current MC programs might see the score as being pretty much even.   But
we can imagine a super MC program thinking that black has a modest
advantage, and an even stronger program might think black has a BIG
advantage and so on.   

However, these programs are constructed probabilistically,  which means
they will never see the opening positions as 1.0, but instead will see
it as 0.99...  ,  and it will still choose the moves that give it the
best chances.     Deeper searches will cause it to better discriminate
between tiny fractions that represent signficant differences in winning
chances.    For instance 0.001 chance of winning is 10X greater than
0.0001 chance of winning.  (Of course these numbers don't really
represent  winning chances, but represent the number of lines that
resulted in a win during the exploratory process,  but that's a
critically important number.)  

In some sense, this is superior to alpha/beta.   A perfect alpha/beta
player would exhibit the bad behavior with a vengeance if it could see
the end of the game,  but a UCT player I believe would still play
reasonably.

- Don


>
> For me it feels like a waste of time experimenting with playing style
> in the end of the game of lost positions. At least it comes really low
> on my priority list.
>
> -Magnus
> _______________________________________________
> computer-go mailing list
> computer-go at computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/
>