Re: burte force and knowledge

[David Mechner <mechner@xxxxxxxxxxx>]

John Clarke wrote:
> By "search" I mean "looking for the best move". 
> [...]
> Some people want to reserve "search" for minimax search, which 
> seems a bit stingy to me.

It's not that we're stingy, "search" happens to be a technical term in
AI, referring to state space search (not necessarily minimax). You seem
to intend the colloquial sense of just looking for something. 

A few observations about how humans play (fast)...

1. Human pattern matching in go is not parallel, but rather requires a
(fast) scanning of attention (and the eyes) over the board. 
Neuropsychologists use the term "parallel" to describe processes that
operate (in the case of vision) over multiple elements in the visual
field simultaneously, without having to shift attention (or the eyes) to
each in turn. For example, to decide whether there is a red element
among a field of green elements, you don't have to consider each element
individually. In contrast, if you are searching for a T in a field of
L's, you will have to scan the display. Similarly in go, if you're
looking for a shape, say a crosscut, you'll have to actively scan the
board until you find one (or finish scanning the board and don't find
one). (The fact that many neurons are working in parallel doesn't make
the /process/ parallel - if it did everything the brain does would be
parallel, and the term would be meaningless.)

2. when humans recognize a part of a position (a "pattern"), they are
simultaneously aware of associated information - for example good places
to play in the area, what goal would be served by each such move, likely
countermoves, how urgent or unstable the situation is, and (depending on
the pattern and the strength of the player) what the final result of the
local goal will be with best play, or is likely to be.

3. using this knowledge, strong players can read extremely fast and
accurately: their extensive knowledge of pat positions, and accurate
judgment of which moves will and won't work allows them to read/evaluate
deep sequences in seconds. In my opinion, the distinction between
knowledge and search breaks down to some degree, and our ability to
report precisely what we're doing by introspection breaks down too. 


So, in some sense John Clark is right - pattern matching is "repeated
over the whole board" but "simple" and "brute force" don't seem to
characterize the process well - a huge amount of domian-specific (go)
knowledge and expertise at evaluating novel positions are integral to
the process, even in speed go.

I agree that this knowledge and expertise is accessed unconsciously,
certainly as a result of the simultaneous operation of billions of
neurons in many parts of the brain. 

But that doesn't mean that any random stupid operation repeated many
times will result in strong play, which seems to be the suggestion. 

-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/