Re: Using partial plys

[Jeremy Thorpe <jthorpe@xxxxxxxxxxxx>]

On Wed, 23 Dec 1998, David Mechner wrote:

> Jeremy Thorpe wrote:
> > there is a simple mistake here. [...]
> > if the value of the next moves are distributed
> > fairly nicely (no huge gaps), the value of sente 
> > should be about half the biggest move [...]
> 
> But there /are/ often gaps in the distribution -- for example a lot of
> maneuvering goes on to get the "last big move" of the opening, middle
> game, endgame. I guess it's an empirical question which approximation
> would work better in practice.

you're right about maneuvering, and it's another interesting topic.
however, i wouldn't even have brought it up, but your approximation isn't
even an approximation at all--here's why:  the average value of sente in
the first 100 moves is like 5.5 to 8.0 points (the value of the 'fair'
komi).  but the average difference between the best and second best moves
is like small fractions of a point (this i say without strong proof, but
consider a position where there are a dosen or so moves that a strong
player (even professional, but certainly amateur dan) might make.  the
difference in value of the best to the worst is far, far less than 12.).
anyways, it's no big deal, i'm just saying.

> > -snip-
>
> I agree, this is a smart thing to do (that's what we do :). But I think
> this is what most traditional programs do, in effect, with their group
> strength. Group strength may not always take on values from 0..1 or be
> explicitly interpreted as probability of surviving, but the effect is
> probably similar: the weaker and the bigger the group the bigger the
> evaluation penalty.

yes, agreed, and it's a good point that 'strength' should not necessarily 
be interpreted as probability of surviving.  it also can have something to
do with sacrifices the killer must make to kill, or the bad (from the
killer's perspective) aji that's left after death is assured.

> 
> Unfortunately knowing that a group is alive with probability .75 is not
> very useful for deciding how to play in a particular instance.

but i disagree strongly.  always, no matter the strength of the player,
there comes a time when life is uncertain.  these are not the closed
positions of the corner (a reasonably strong human player can figure these
out right every time, and it is reasonable to want a computer program to
do the same), but rather open positions where the only thing professional
commenters can say is 'wow, black has chosen to pick a massive fight
here... let's see what happens.'

but i see that your real argument is about the important 'next step' where
the computer finds the fine line that we humans learned when we passed
20k, which is that there's some times that some stones are simply going to
die, and there's nothing to do about it.

my argument is that there's no reason that we can't simply extend the
binary classification of 'alive' and 'dead' to include some things in the
middle, and that this extension will never go away when our program learns
to play like a dan, and then like a pro.  it will simply get better at
classifying positions more strongly.

> what you really need to play above the double-digit kyu level is a
> module that:
> 1) says 'alive,' 'dead,' or 'undecided' 
> 2) is right 90% of the time when it's confident,
> 3) tells you when it's not confident, and
> 4) is confident in a large proportion of positions of normal 
>    complexity.
> 
> This isn't impossible, it's just hard.
> Life and death is the critical subproblem of computer go.
> 
> > or--another thought: let's say you know that your stones can 
> > be killed if you don't play, but you think that 75% of 
> > programs out there can't find the right move.  
>   [...]
> > what do you think?
> 
> I think that's as self-destructive a way to think for a go program as it
> is for a human player -- your mileage may vary.

alright, i'll take that... my comment withdrawn.  just for the record, i
wasn't thinking of a decision that a programmer makes when he writes the
program about how strong the other programs will be, but rather about a
population of evolving agents each of which learns nothing more than how
to survive in the immediate population.
 
> -David
> 
> -- 
> David A. Mechner            Center for Neural Science
> mechner@xxxxxxxxxxxxxxxxx         4 Washington Place, New York, NY 10003
> 212.998.3580                http://cns.nyu.edu/~mechner/
> 

-jeremy