Re: [computer-go] Pattern Matcher

[Vincent Diepeveen <diep@xxxxxxxxx>]

Frank,

Let's summarize: you work on a module that can order moves hopefully better
as it regurarly picks the correct moves.

The advantage of a move ordering is that you can afford to try some real
stupid move first, hoping it is a brilliant attacking move.

In alfabeta with nullmove and a deep search, it is very important to pick
the first move a reasonable move.

Lucky we have hashtables for that and internal iterative deepening,
killermoves and history moves are very popular with most programs too (not
working for DIEP somehow).

Even after that a good move ordering is very interesting. We can for
example throw out history table and use your code there.

My question is how much system time it costs to select a candidate out of
the legal move list using your pattern matching.

If i want to do it say 100000 times a second at an opteron 2.0Ghz, can i do
it with your software?

I hope you see the *need* for doing it quickly. Because a better move
ordering might sooner give a cutoff, it means a deeper search. The tradeoff
is that calling your function costs system time. I want to know how much
and it is no problem to mention things you guess can get speed optimized.

I also hope that if at a spot which normally gets called 50000 times a
second, that if your move ordering can deliver it only 50000 times a
second, that the program slows down 2 times, which is a severe penalty.

Vincent

At 15:55 8-11-2004 +0100, Frank de Groot wrote:
>Mark Boon wrote:
>
>> $100,000 or any other amount. I've only expressed my opinion that it's
>less
>> valuable to a Go playing program than a good life-and-death module. This
>> because I think it's harder and more work to make.
>
>This is an illogical statement.
>The value of a software module to a Go program depends on the increase in
>playing strength it adds to that program, not how hard it is to make.
>
>
>> games, is the answer. Instead I think smaller sets of specialised
>patterns,
>> used for specialised tasks, will work much better.
>
>No, they are *also* needed.
>
>But your hand-entered patterns will never be able to know all contemporary
>Joseki.
>You would need to enter 20 times Kogo's in order to get even close.
>
>
>> The only way your approach is going to be better is if your patterns
>include
>> Go knowledge such that you can reliably throw away the bad candidates and
>> not look at them at all. And do this without evaluation. Correct me if I'm
>> wrong, but I understood that this is exactly what you claim. But if that's
>> the case, then I don't see why you need evaluation and lookahead at all.
>You
>> pattern-matcher will always only suggest pro-quality moves. Didn't you say
>> you were able to do this reliably for the first 100 moves or more?
>
>Never.
>
>
>> I suggest a simple experiment? Play a game against another program, say
>GNU
>> Go. Let your pattern-matcher suggest the best move in each position and
>play
>> it against GNU Go. If you're right, then after 100 moves your program
>should
>> be miles ahead. If your Go level is not sufficient for such an experiment,
>> ask a strong Go player for help.
>
>You came with the belief that a pattern matcher is useful to play Go.
>I never said that.
>But you made it up, and then you accuse me of saying that a pattern matcher
>is useful to play Go.
>
>
>> Now if your program indeed shows such good play during the first 100
>moves,
>> then I have no problem to admit I was wrong. And in that case I think your
>> program is worth much more than $100,000.
>
>
>You keep on referring to my "program", thereby spreading the misinformation
>that my "program" is based on pattern recognition.
>
>I have said time and time again: I do not have a Go program and I do not
>intend to make a Go program based on pattern recognition. Please stop
>spreading this nonsense as it may damage the reputation of my future Go
>program.
>
>The only thing I have at the moment is a pattern expert system. I intend to
>use it for Fuseki, Joseki and good shape and to help sort moves for search.
>
>
>
>> equal after 100 moves, then I think it doesn't mean much yet apart from
>that
>> your claim seems wildly exaggerated.
>
>The only things I claim are my results.
>
>> worse ater 100 moves, then I think you'll have to admit your program is
>> probably not worth more than any other plain pattern-matcher  when it
>comes
>> to a Go playing program. It could still have good value as a study tool,
>but
>> that's not what we're discussing here.
>
>You are blabbering nonsense. Please cease doing that.
>
>
>> And please refrain from saying people here are just jealous. It's childish
>> and only makes us think you are the jealous one of the money that some of
>us
>> here seem to have made with their software.
>
>I do not care about money. If I cared, I would not have quit my job.
>If I cared about money I would have worked on commercially much more
>interesting software.
>It's madness to bet that one will ever make money with a Go program.
>In fact I had no idea that people made quite some money with their very bad
>Go programs.
>It's nice but I could care less. I have not long to live most likely and by
>the time my program is ready to be sold, I will be dead anyway.
>
>_______________________________________________
>computer-go mailing list
>computer-go@xxxxxxxxxxxxxxxxx
>http://www.computer-go.org/mailman/listinfo/computer-go/
>
>
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/