Re: brute force and knowledge

[Jeremy Thorpe <jthorpe@xxxxxxxxxxxx>]

Jeremy Thorpe (#455)                     "why is a rhyme the next best
Mech. Engineering undergrad at UCR                 thing to a reason?"
www.engr.ucr.edu/~jthorpe                               -jeremy thorpe

On Mon, 30 Nov 1998, John Clarke wrote:

> Hi
> 
> > >The brain isn't doing a lot of serial (eg. tree search) computation - since
> > >pros can play fast, so it must be a big parallel evaluation.
> > >
> > >You don't see the parallel search because it is subconscious.  That's why the
> > >pro's find it hard to say what made the move the best.  The search is almost
> > >certainly some kind of pattern matching since that's what we excel at.
> > 
> > Can you redefine what you say?
> > 
> > LIttle confused here what you mean to express. 
> > I see clear difference between:
> >   'pattern matching' and 'parallel search'
> 
> Please don't think that the only kind of search is alpha-beta pruned minimax.
> Certainly that's the best way for chess, but this is go.

this sounds better than before--i have to say that i thought you were
talking about minmax...

> 
> Think of search as just being "looking for the best move".
> 
> Possibly pattern matching could be coupled with trying lots of random (or
> somehow directed) moves to find good shapes on the board.  ("Wouldn't
> it be nice if it looked like this?")  Then one could play the move that
> appears in most of the good shapes.
> 
> The problem with a mechansitic pattern matching (one stimulus => one response)
> is that one can get trapped in local maxima.  The player may need to allow
> the situation on the board to appear to deteriorate for a while in order
> to get to something even better.

well, with a sharper evaluation function, it wouldn't actually appear to
deteriorate, would it.  i can't see how mechanistic pattern matching leads
to 'local maxima' at all, and in fact, i do think that very good
evaluation functions can be created based on nothing more than
mechanistic, efficient, pattern matching.

but i see this requires explanation:  i don't mean simply the sum of
evaluations of local patterns, but evaluations of larger patterns of
local patterns (for example, half connection to live stones here, 1 1/2
eyes there = good).  Anyway, this is the best way i can describe the way i
play--and i don't think that it involves much parallelism.  once a local
position is settled, i usually don't even think about it until either
something happens nearby or sente drops below a certain value.

> 
> So imagine something like a continous process of (1) matching patterns all
> over the board, (2) trying (sequences of) random moves all over the board,
> (3) finding consistently good moves (4) using results of (3) to direct the
> moves tried in (2).  This is a parallel pattern matching search.  (Very
> similar to the simulated annealing go program mentioned earlier in this
> thread.)

i find that (2), subconsciously trying sequences of random moves all over
the board, is highly implausible simply because there are way to many
variables for a subconcious thread to deal with.  i mean, if i just get up
from a game to go to the toilet, it's almost impossible for me to read
three or four moves into the last position that i was looking at, let
alone a position somewhere else on the board that i haven't been looking
at for 10 mins. 

> Notice there's no tree search here.  Just a process of finding moves likely
> to be good.  Those suggestions are then refered to the conscious brain for
> checking - by tree search.
> 
> Regards
> John Clarke
> johnc@xxxxxxxxxxxxxxxxx
> 
> 
>