Re: computer-go: Pattern matching

["Gary Boos" <gboos@xxxxxxx>]

So you're working on a program ?  What language and platform ?
Why not joing the GNU-go effort ?  I'm sure they are willing to listen to
new ideas.
        Gary
--------
-------
-----Original Message-----
From: Patricia Hughes and David Elsdon <babel17@xxxxxxxxxxxxxxxxx>
To: computer-go@xxxxxxxxxxxxxxxxx <computer-go@xxxxxxxxxxxxxxxxx>
Date: Tuesday, November 09, 1999 12:43 PM
Subject: Re: computer-go: Pattern matching


>I am strongly in agreement with Tristan !!
>
>Tristan Cazenave wrote:
>
>>  I think we can expect a computer AI to learn better than a person.
>> As well as we can expect a computer AI to play chess better than a
person.
>>
>> In my opinion, working on making an AI learn better than a person is
>> fundamental, not only for Go, but for AI in general.
>> I wrote several papers related to this problem for Go, and I believe than
>> in
>> the future learning AI Go program will be among the strongest programs.
>> If search is what makes a chess or a traditional board game program
>> strong,
>> search AND good learning maybe what makes a Go program strong.
>>
>> There is so much knowledge involved in a Go program that mastering it is
>> not
>> a human task, it is an AI task. Working on Go programs using
>> metaknowledge,
>> that is knowledge to create and use knowledge is a research that can lead
>> to
>> a strong increase in the level of Go programs. Moreover it may enable to
>> master AI techniques that will enable AI to apply to more complex domains
>> than it presently does.
>
>And disagree with Moucheng:
>
>"Xu, Mousheng" wrote:
>
>> * Strongly doubt AI will work for Go programming. :)
>
>I don't want to get into too much of a discussion amount what counts as
>AI,
>but I think that a program using knowledge representatation and/or
>learning
>is an AI program.
>
>Deeper Blue uses a lot of chess knowledge as well as an extraordinary
>amount
>of search.
>The champion checker playing program uses checker knowledge plus search.
>The champion backgammon program "learned" to be that good !!
>
>So in Go where brute force search is not very effective even more AI
>will be required.
>
>> * We can't assume AI can do everything.
>
>Why not ?? Any knowledge which a person has can be represented in a
>computer. I don't say it will be easy.
>Give me an example of knowledge that cannot be represented on a machine
>!!
>
>> In some areas, AI does better than human, that's because people don't
>> understand the subject enough.
>
>I don't under this. The difficult areas for AI are those which people do
>not
>understand thoroughly - i.e. are unable to articulate. Common sense
>knowledge is notoriously hard to capture, though by no means impossible.
>
>Once knowledge it clearly and completely articulated it can be
>represented on
>a computer.
>
>> Go is probably different. Go knowledge can be very well explained by
>> "shape", which means human beings know Go very well, can translate the
>> knowledge very well, and can learn the knowledge very well. We know Go,
>> we know it!
>
>If we know it well, then we can articulate it and we can represent it on
>a
>computer. I am a knowledge engineer and I've been doing this sort of
>thing
>for the last 15 years !! The game of Go is "just" a bigger challenge.
>
>> The knowledge of "shape" can be somewhat understood by the
>> computer as well, but to let it master the knowledge has to do with
>> (huge) numbers.
>
>No bigger than the numbers (of patterns) that we humans have to cope
>with.
>
>> One can give so many "simple" questions to a computer to
>> learn, and it may progress 0 or a very little over years.
>
>Surely this is under-estimating the power of computers. Data-Mining
>software, for example, can trawl through gigabytes of data and find
>interesting patterns (using various machine learning techniques) where
>mere
>humans simply wouldn't have a chance.
>
>> I don't doubt, that after many many many years, an AI program will
>> eventually be a 9
>> dan pro, but that might take as long as to wait for Darwin to evolve to a
>> monkey.
>
>I confidently expect a computer to "beat" a 9 Dan Professional before
>the
>year 2010.
>
>> * AI still has a chance to win, but that would be AFTER we understand Go
>> better in a form a computer can understand.
>
>I think we will come to understand Go better by writing Go programs. A
>Go
>program is operationalised Go Knowledge. If you write a book about some
>aspect of Go you can get away with being quite sloppy. When you write a
>Go
>program you have to be very precise because your program will do
>"exactly"
>what you tell it to. This makes a Go program a very accurate
>representation
>of the whole body of Go knowledge. When we are able to combine this very
>accurate representation with state-of-the-art computer learning
>techniques
>we will start to get really strong Go programs.
>
>I am working on it !!
>
>But don't hold your breath !!
>
>Cheers
>
>David Elsdon
>