Re: computer-go: Pattern matching

[Patricia Hughes and David Elsdon <babel17@xxxxxxxxxx>]

"Hans F. Zschintzsch" wrote:

> > I confidently expect a computer to "beat" a 9 Dan Professional before
> > the year 2010.
>
> How many humans you know who only needed 10y form 5k to 9d ?

This really is not the point !! I have written an expert system which
successfully captured the knowledge of a particular area of optical design
from an optical design expert. I did not have to become an optical design
expert in order to do this. I spent about half a day a week with the expert
for some 40 weeks eliciting the knowledge and structuring it. There were
two of us who were involved with the coding. The resulting system, running
on a Micro-Vax, did in half an hour what took the expert hours or days to
achieve.

The system we built knows a lot more about optical design than I do !!

The resulting system also codified the experts knowledge. This codification
looks nothing like a text book on optical design. Instead it captures all
the heuristics the expert actually uses to do his job.

> I tried too, but my first program played Go so incredibly
> bad - after a lot of effort - that i'm convinced a "break
> through in computing technology" is needed to get near
> 5 handicap stones of pro level.

> My theory: Everybody who tried in theory only and not really
> made a program play, severly underestimates the difficulties
> to get as high (or low) as 5k, which todays best Go playing
> programs have achieved (as far as i know).

I have written a program in Prolog which uses about 10 Go concepts
(strings, influence, cutting-points, etc) and 30 rules and no lookahead at
all. It takes less than a second to make a move and plays quite well - for
a program. It beat Plodder on the 9x9 go-ladder. It can also beat Many
Faces on a 9x9 board when playing white, but loses by about 20 points when
playing black. Clearly my program is not robust, but quite good considering
I have only spent a few months developing it.

> PS. Please show us what your program does today.

Certainly, I can send you the Prolog code or we can arrange to meet on one
of the Go Servers where you can play against my program. It only plays on a
9x9 board since it has no rules to play 19x19 fuseki yet.