Re: [computer-go] Pattern Matcher

[Vincent Diepeveen <diep@xxxxxxxxx>]

At 13:54 7-11-2004 +0100, Frank de Groot wrote:
>> > I charge 1 dollarcent for a pattern, 8,388,608 patterns = 100,000 Euro.
>> >
>>
>> Hiihii!!! Sorry, I almost fell from my chair laughing when I read this!
>> :-)))
>
>But you laugh for the wrong reason (inability to understand humor).
>I would never sell my pattern system. I just explain why your "it ain't
>worth much" is ill-informed.
>When guys like you and Reiss mame hunderds of thousands of USD, both are
>well-educated and can make fortunes in any job, it means that any Go
>software system that solves a piece of the puzzle, is automatically worth a
>lot.

;)

>You simply have not understood what my pattern system is and what it does,
>and immediately saying that it is highly inferior in usefulness to a L&D
>solver is simply naive. Both are important. I am pretty sure that my system
>is much more important, as my system allows to predict 20% of all pro moves
>in unseen games and it takes a few microseconds per move to do that.  And,
>in a few weeks to months, it will predict 40% of pro moves, when things go
>well.
>
>
>> Keep generating patterns I'd say.
>
>The nice thing about my system that it has all relevant patterns already.
>That's another reason why it is valuable.
>I discovered that there isn't much more interesting beyond 8 million
>patterns.

Your problem is the statistical significance of a pattern. 8 million
patterns would require like 8 billion professional go-games to even have a
slight proof that a certain pattern works.

How many professional go-games do you have in your database?

>
>> > to play along with pro-level further-then-move #100 "after-Joseki". And
>it
>> > makes the correct decision as to which moves are more relevant than
>others
>> > (it will take care of pressing issues and then return to a Joseki etc.)
>No
>> > current systems are able or will be able to do that.
>> >
>>
>> "or will be able to do that."?!? Of a seldom arrogance, rarely seen even
>in
>> mailing-lists like these.
>
>
>Your jealousy has made you too quick to react. Obviously English is not your
>native language.
>You simply did not understand the sentence.
>I claim that current systems will not be able to do what my pattern system
>does.
>You interpret it as "future systems will not be able to do what my system
>does".
>I simply say that current systems, by their design, will never be able to do
>what my system does.
>I never said anything else and you interpreting it totally upside-down is of
>an arrogance rarely seen on mailing lists like this. It shows you have a
>very limited grasp of the English language, so limited that you really have
>to take great care in communicating with peers (you will easily offend them
>or say silly things).
>
>Let me give you some advice.
>Whenever you do not understand something, do not assume your proponent must
>automatically "not understand" and is "arrogant", "wrong" etc. And do not
>react with an insult. This is a typically Dutch thing to do, I am Dutch and
>I know all too wel how the Dutch are quick to insult. But the Dutch are
>armchair warriors, they talk big but they usually lack the staying power.
>
>I did not say that I laughed when I read your request to "send me the source
>of a much better pattern matcher for 1000 USD and I will throw it on the web
>for all to use". Because I'm a bit more polite :)
>You don't seem to understand that a top-quality pattern matcher is worth
>around 100,000 USD, as I tried to explain to you, obviously in vain.
>
>We have people like Kierulf who gave up Management-level jobs at Microsoft
>to start a Go software company, there are people investing millions in Go
>software and if you would understand the relevance of pattern matching in Go
>and what can be done with pattern matching in Go, you would not ridicule a
>hyperfast pattern matcher with 8 million patterns from 500,000 games.
>
>
>> use, as I'm of the opinion that a large pattern-database, no matter how
>> sophisticated, doesn't contain much fundamental Go knowledge.
>
>
>A L&D solver neither contains less Go knowledge.
>It makes it up on the spot.
>
>And your religious beliefs about what a pattern systemn can do, well,
>opinions are like ........, everbody has one.
>I explained that my system spent a hell of a lot of time correlating
>patterns, meaning it extracted Go knowledge.
>It has knowledge about the priority of those 8 million patterns. It is not
>just a matcher, it knows that pattern A is 3 times more urgent than Pattern
>B, etc. You can say that I am a liar but that would be rather desperate.
>
>
>> will therefore never be better than the kyu-player who inappropriately
>> mimics moves because he's seen them played by a pro,
>
>Are you suggesting that all Chess programmers should throw their opening
>books away and all the other pattern knowledge they use at all stages of the
>game and in all parts of their engine? Because "This only leads to beginner
>play"? Don't be ridiculous. Patterns are much more important in Go than in
>chess. Why on earth do you offer 1000 USD for a pattern matcher if you think
>it's a pretty irrelevant part?
>
>I start to believe you're just an asshole.
>I came along claiming to have a "special" pattern matcher, you checked it
>out and saw it was true, then you thought: "Let's fuck this guy by offering
>1000 USD for something just as good and then I'll throw it in the public
>domain to fuck this guy up, so that his work will be in vain".
>
>Well, buddy, tough luck but you'd be hard pressed to find anyone with a
>system comparable in power to mine give it to you for 1000 USD. Like I said,
>there are no such systems, currently and current systems can't be used to do
>what I do. Just the months of 24/7 number crunching and resulting huge
>databases ensures that it is a serious endeavor and no serious Go researcher
>will give away his work for 1000 USD to Mr. Boon. They either publish or use
>it in a product.
>
>Go programmers keep on telling us on how they spent years on laboriously
>entering tens of thousands of patterns, on how they enlisted the help of pro
>players to do it for them, I extracted and correlated more than one hundred
>billion patterns (from games I paid a lot of money for BTW, GoGoD is
>*totally* useless, in fact so useless I had to exclude it from the final
>learning cycles). I kept the 8 million best from 100 billion (125 to be more
>exact). So I don't think a cent per pattern is crazy, if I would be so
>stupid as to sell it.
>
>I discovered a few weeks ago that I get ridiculous pro-prediction (better
>than anything ever achieved) when I use my patterns as inputs in a 3-layer
>NN and add some more simple stuff, like how big the pattern is and at which
>move it is played, how often it occurs, the distance to the centre, the
>number of stones in it, the value of the pattern for the opponent, the
>distance from the last played move. Sometimes, almost an entire, never-seen
>pro game is predicted correctly (but that is rare). I am now improving the
>system by adding a 21/5-type input for the NN.
>
>You don't seem to be able to think beyond a mere matcher. Yes, a mere
>matcher is not worth more than 1000 USD.
>A matcher that comes with millions of cross-referenced patterns and packaged
>into a system that can be used inside a search engine or as input for a NN
>is a whole differrent ballgame. Perhaps you should read something about the
>current developments in computer Go.
>
>
>> understood why they're played. Without the underlying knowledge a pattern
>is
>> of limited use, and can even be counter-productive.
>
>You really have to read my posts if you want to comment on them.
>I explained that my system correlated all patterns.
>Meaning that when it sees a popular Joseki move, but there is a ladder
>threatened on the other side of the board, it will first do the most
>important move. My pattern system is bursting at the seams with knowledge.
>It knows the relative value of patterns. And a pattern is anything. "Defend
>chain in atari when there is not a longer chain in atari" is a pattern.
>"Play Tengen as long as there is not a XYZ Joseki" is a pattern. Because
>patterns arre correlated and cross-referenced and sorted into importance
>based on statical analysis.
>
>You keep saying that "without Go knowledge it's useless" and I keep saying
>that this Go knowledge is included in the "1 cent per pattern". By virtue of
>statistical analysis of the relative whole-board value of the patterns (I
>keep on saying it because you keep on ignoring it).
>
>When you look at the recent game on IGS "Mr. Popo vs. Gomonster" and you see
>that my system predicts just about EVERY MOVE up to somewhere around move
>120, and when you realize that my system has never seen that game before,
>you can't deny that it's a little more than a simple pattern matcher. It is
>an expert system.
>
>It is an expert system because it actually shows you where it has seen those
>patterns before. In a fraction of a second. The databases needed for that
>take up 500 MB. All this stuff took ages to extract and crunch.
>
>
>> would incorporate my pattern expert system", to quote you, will be
>trounced
>> by any of the top Go software that currently exists unless it also
>> incorporates a large part of the underlying knowledge that was the basis
>of
>> the pro moves in your game-database,
>
>So what? The point is that a strong Go program will become much stronger
>with my system.
>A chain is as strong as its weakest chain.
>
>
> in which case that program is centuries
>> ahead of everyone else already without the database. I don't claim to know
>> everything about computer-go, and I may be totally wrong. In fact, I'd be
>> happy to be proven wrong, it means I really learnt something. But you'll
>> have to prove it before I'll see the light. Go ahead, make a Go playing
>> program and see how it plays.
>
>The enormity of the work required on patterns (both in coding and number
>crunching) precludes much other work.
>But I will, in half a year hopefully, sell a study tool that used the
>pattern system. It will also suggest Go moves but only based on the pattern
>system and a simple neural network.
>
>> You'll have to, because I doubt any
>> go-programmer is going to give you the $100.000 you want for it. Until
>then
>> I'll consider all of this just a large amount of hot air.
>
>I don't have to do anything.
>My study/playing tool will demonstrate its worth and that 100,000 will come
>bit by bit :)
>When it predicts 85% of the first 120 moves in a never-seen-before semi-pro
>game, even I as a non- go player understand that perhaps this pattern
>matcher is not a pattern matcher but a high-powered expert system :)
>
>Sorry that you are unable to handle that :)
>
>> Oh, by the way. I agree that your pattern-matcher is probably a jewel. But
>> the value is not in the patterns it generates, it's in the games you feed
>> it. Did you know you have to pay royalties on pro games?
>
>Bullshit.
>I have won a copyright suit or two in my time.
>I know more about copyright law than the average lawer does.
>
>
>> You'd better be
>> careful about what you're trying to 'sell'.
>
>
>I will sell 52,000 pro games and hundreds of thousands of 6d* and 7d* games
>but not in the form of SGF files.
>The SGF files are contained in an encrypted, compressed SQL database and you
>can play them through and search throught them, but the pattern database is
>something alltogether different.
>
>Your outrageous claim that you "have to pay royalties" makes you ridiculous.
>Game records are recordings of events and can't be copyrighted.
>
>There are some funny laws in countries like Japan where they ignore the
>Berne convention and make up their own rules, but nobody outside Japan
>cares. GoGoD for example violates Japanese copyright law (AFAIK GoGoD
>contains Japanese games).
>
>_______________________________________________
>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/