Timescale for computer world champion at go?
Tim Tyler
``At the recent FOST Cup I asked the participants two
questions. When do you think a computer will be shodan
level? and when do you think a computer will be able to beat
any human player?
Shodan level is 'international shodan', which is maybe 3 dan
in Japan, 2 dan in America, 1-dan in Europe. [...]
I've ordered the list with most optomistic first.
Professional players, and go programmers have been noted.
Shodan World Champion Name
1999 2005 Mei-Kou Tei 9-dan (programmer)
2000 2010 Darren Cook (programmer)
2000 2010 Chihiro Mizuuchi
2000 2010 Naritatsu Yamamoto
2002 2040 Martin Mueller (programmer)
2005 2050 Hirooka
2005 2100 Amano
2007 2097 Ken Chen (programmer)
2007 2097 Kojima 9-dan
2007 2197 Redmond 8-dan
2010 2023 Shinichi Sei (programmer)
2010 2030 Tristan Cazenave (programmer)
2010 2045 David Fotland (programmer)
2010 2050 Yung Jye Hunag (programmer)
2010 2100 Jun Saito
2010 2150 Kobayashii
2015 2035 David Keeble
2017 2090 Hiroyuki 8-dan
2020 2050 Mick Reiss (programmer)
2020 2100 Jay Burmeister
2020 2100 Chen Zhixing (programmer)
2020 2100 Prof. Hsu
2020 2100 Yoshikawa
2020 2200 Oyama
2030 2050 Masahiro Okazaki
2050 2100 Kim
2100 -- Izuka
2200 2500 Fujisawa
2500 3000 Oyaizu''
- http://www.computergo.net/forum/viewthread.php?tid=836&extra=page%3D5
Have there been any more recent estimates by those involved?
--
__________
|im |yler http://timtyler.org/ t...@tt1lock.org Remove lock to reply.
Hans-Georg Michna
very interesting, thanks!
Apparently almost all people who offered their opinion were far
over-pessimistic. Only a few appear to come close to the
probable truth.
Hans-Georg
--
No mail, please.
-
Hans-Georg Michna <hans-georgN...@michna.com> wrote:
> Apparently almost all people who offered their opinion were far
> over-pessimistic. Only a few appear to come close to the
> probable truth.
The question was ambiguous, speaking of "a computer..."
Distributed computing, multiprocessing, parallel cores, etc., is not
exactly what many people interpreted by the phrase "a computer"
in 1997. They were probably referring to a stand-alone portable
machine, easily afforded by average electronics consumers. This
programming challenge is cast in terms of bringing down operations
cost for software which may be easily sold and distributed. For the
Chess World this occurred in the wake of Deeper Blue: a good
number of reasonably-priced standalone chess programs. Most
Tournaments could not accept machine entries which must have
telephonic or internet connections elsewhere anymore than they
could accept human players who receive moves via telephone.
- regards
- jb
--------------------------------------------------------------
Web Video of the Week:
Zeitgeist - The Federal Reserve
The original two hour Zeitgeist movie was released in 2007, but its
provocative and insightful investigation of the history of the Federal Reserve
and the U.S. central banking system could not be more relevant than right now.
Every American needs to truly understand the background of this system. We
have posted the section of the movie focused on the Federal Reserve. It'll
make you think twice about everything you are hearing in the news.
Watch
http://www.organicconsumers.org/articles/article_14910.cfm
--------------------------------------------------------------
Tim Tyler
> Apparently almost all people who offered their opinion were far
> over-pessimistic. Only a few appear to come close to the
> probable truth.
The estimates were all over the place. The first ten seem
over-optimistic - at least when it comes to estimating
Shodan level. Eleven years have passed, though - it might be
time for another survey.
Batavia
mogotitan was eastimated at 4 or 5d. so 2007 for shodan wasn't too bad
Denis Feldmann
Really? It (he?) lost against a 8p on 7 stones, didn"t it?
Tim Tyler
> mogotitan was eastimated at 4 or 5d.
By whom?
http://idazuwaika.wordpress.com/2008/09/21/plenty-of-room-for-research-in-go-ai/
Michael Alford
> Batavia wrote:
>
>> mogotitan was eastimated at 4 or 5d.
>
> By whom?
>
> http://idazuwaika.wordpress.com/2008/09/21/plenty-of-room-for-research-in-go-ai/
>
I believe it was Myungwon Kim that said -some- of Mogo's moves were
amateur 5d level. This is discussed in the computer-go newsletter. While
I think it is possible for Mogo to play 5d moves in random game
positions, I doubt these programs will ever play a whole game at that
level. As we are all aware, at some point, to get better at this game,
you need a teacher. So it might just be my personal bias, but I simply
don't believe you can treat Go as a FSM, despite the confidence the
programmers seem to have in their playouts.
Michael
Ted S.
> I believe it was Myungwon Kim that said -some- of Mogo's moves were
> amateur 5d level.
Hell, I can play hoshi on my opening move. That's a 9p move right
there. :-)
--
Ted S.
fedya at hughes dot net
Now blogging at http://justacineast.blogspot.com
Hans-Georg Michna
>Hans-Georg Michna wrote:
>> Apparently almost all people who offered their opinion were far
>> over-pessimistic. Only a few appear to come close to the
>> probable truth.
>The estimates were all over the place. The first ten seem
>over-optimistic - at least when it comes to estimating
>Shodan level. Eleven years have passed, though - it might be
>time for another survey.
Tim,
these surveys are near-useless, because if most of the surveyed
people are clueless, then the result of the survey will be
nonsense.
It would be more interesting to survey renowned futurologists.
Futurology is a difficult and very peculiar science. Even
technology specialists can guess the future totally wrongly.
Joel Olson
"Hans-Georg Michna" <hans-georgN...@michna.com> wrote in message
news:f7lee4dlkomr4pmst...@4ax.com...
I quit on them when they came up with this weird version of Bayes -
"Pick the future you want & then figure out how to get there."
Tim Tyler
> these surveys are near-useless, because if most of the surveyed
> people are clueless, then the result of the survey will be
> nonsense.
These guys were the 3rd FOST Cup (1997) participants -
including many computer go / AI experts.
> It would be more interesting to survey renowned futurologists.
> Futurology is a difficult and very peculiar science. Even
> technology specialists can guess the future totally wrongly.
Maybe - one would need to find the ones that know something
about both AI and go. The World Open Computer Go Championship
seems like a fairly reasonable place to ask the question.
Hans-Georg Michna
>Hans-Georg Michna wrote:
>> these surveys are near-useless, because if most of the surveyed
>> people are clueless, then the result of the survey will be
>> nonsense.
>These guys were the 3rd FOST Cup (1997) participants -
>including many computer go / AI experts.
>> It would be more interesting to survey renowned futurologists.
>> Futurology is a difficult and very peculiar science. Even
>> technology specialists can guess the future totally wrongly.
>Maybe - one would need to find the ones that know something
>about both AI and go. The World Open Computer Go Championship
>seems like a fairly reasonable place to ask the question.
Tim,
indeed it does. It is just that the results of the first survey
are, for the most part, so ridiculously far off that I'd take
that as proof that those people were totally clueless when it
came to answering this particular question.
I guess the result of any similar survey would be similarly far
off and not helpful at all in predicting the future.
Tim Tyler
> indeed it does. It is just that the results of the first survey
> are, for the most part, so ridiculously far off that I'd take
> that as proof that those people were totally clueless when it
> came to answering this particular question.
>
> I guess the result of any similar survey would be similarly far
> off and not helpful at all in predicting the future.
It sounds as though you have your own assessment of the issue...?
Or maybe you are just looking at the variance in the data ;-)
If the former, I'd be delighted if you shared your estimates.
Hans-Georg Michna
>Hans-Georg Michna wrote:
>> indeed it does. It is just that the results of the first survey
>> are, for the most part, so ridiculously far off that I'd take
>> that as proof that those people were totally clueless when it
>> came to answering this particular question.
>>
>> I guess the result of any similar survey would be similarly far
>> off and not helpful at all in predicting the future.
>It sounds as though you have your own assessment of the issue...?
>
>Or maybe you are just looking at the variance in the data ;-)
>
>If the former, I'd be delighted if you shared your estimates.
Tim,
I'm just looking at the current crop of computer go bots and see
that they are very close to dan level. LeelaBot, for example,
plays as 1k on KGS.
It is obvious to me that it won't take another decade to reach
dan level. The strength of the programs will increase, even if
the programmers do nothing, just because processor performance
rises.
-
Hans-Georg Michna <hans-georgN...@michna.com> wrote:
> The strength of the programs will increase, even if the programmers
> do nothing, just because processor performance rises.
Why does everyone neglect to mention heuristics ?
- regards
- jb
------------------------------------------------------------
The Federal Reserve: A Corrupt Institution
http://www.wvwnews.net/story.php?id=5738
------------------------------------------------------------
ADL Reports Surge in Anti-Semitic Messages on Online Finance Sites
http://www.wvwnews.net/story.php?id=5736
------------------------------------------------------------
Hans-Georg Michna
>Hans-Georg Michna <hans-georgN...@michna.com> wrote:
>> The strength of the programs will increase, even if the programmers
>> do nothing, just because processor performance rises.
> Why does everyone neglect to mention heuristics ?
Perhaps because it isn't considered important?
What exactly are you aiming for?
-
>> Hans-Georg Michna <hans-georgN...@michna.com> wrote:
>>> The strength of the programs will increase, even if the programmers
>>> do nothing, just because processor performance rises.
> "-" wrote:
>> Why does everyone neglect to mention heuristics ?
Hans-Georg Michna <hans-georgN...@michna.com> wrote:
> Perhaps because it isn't considered important?
Would you similarly conclude that statistical analysis is unimportant?
While unproven, many assumptions occur as the consequence of
a "learning" activity. Despite being assumptions perhaps a strong
likelihood exists owing to the fact they may later be proved correct.
> What exactly are you aiming for?
10 dim A%(5000):T=20:' ----------------- prmsp.ub prime spacing
20 for I=1 to 50000:J=4
30 inc A%(prm(I+1)-prm(I))
40 print I;prm(I);
42 while A%(J):print using(1,2),A%(J)/A%(J-2);:J+=2:wend:print
45 if I@T=0 then input T
50 next
Note that after many iterations, pr(A%(2)) = pr(A%(4)), suggesting
that the conjecture of infinite primes differenced by 2 is identical to
the conjecture of inifinite primes differenced by 4. Values for A%
reach a maximum at A%(6), where pr(A%(6)) = 8/5 * pr(A%(2))
and pr(A%(8)) = 2/5 * pr(A%(6)) = (4/5)^2 * pr(A%(2)) . Owing
to the "square" term relating pr(A%(8)) with pr(A%(2)) this may
establish the necessary and sufficient conditions for two-eye life
in the game of Go. Surely a smart algorithm could benefit by
making use of -some- inferences derived from heuristic results.
- regards
- jb
----------------------------------------------------------------
Results 1 - 9 of 9 for statistic "difference of primes". (0.19 seconds)
http://www.google.com/search?hl=en&as_q=statistic&as_epq=difference+of+primes&as_oq=&as_eq=&num=100&lr=&as_filetype=&ft=i&as_sitesearch=&as_qdr=all&as_rights=&as_occt=any&cr=&as_nlo=&as_nhi=&safe=images
----------------------------------------------------------------
-
jazze...@hotmail.com (-) wrote:
> ... Values for A% reach a maximum at A%(6) ...
In turn, heuristic methods require additional heuristic
temperments, i.e. "Jumping champions" (Experiment. Math.
Volume 8, Issue 2 (1999), 107-118, by Andrew Odlyzko, Michael
Rubinstein, and Marek Wolf), see:
http://projecteuclid.org/DPubS?service=UI&version=1.0&verb=Display&handle=euclid.em/1047477055
& http://www.dtc.umn.edu/~odlyzko/doc/cnt.html
- regards
- jb
------------------------------------------------------------------
Hidden structure in the randomness of the prime number sequence?
http://www.sciencedirect.com/science?_ob=ArticleURL&_udi=B6TVG-4GP2726-2&_user=10&_rdoc=1&_fmt=&_orig=search&_sort=d&view=c&_version=1&_urlVersion=0&_userid=10&md5=f1c55c9495fa0d20a7b3580db61ed80f
------------------------------------------------------------------
-
> "-" wrote:
>> Why does everyone neglect to mention heuristics ?
Hans-Georg Michna <hans-georgN...@michna.com> wrote:
> Perhaps because it isn't considered important?
"Delayed this year while creator David Fotland rewrote
his engine to use a combination of Monte Carlo Tree Search
and its existing go knowledge, a prerelease of The Many Faces
of Go version 12 -- which just won the International Computer
Game Association's annual world championship ... The MFOG
version 12's engine "combines the new Monte Carlo Tree
Search algorithm with Many Faces' go knowledge, to give
something stronger than either," Fotland tells the E-Journal.
"It plays a more natural style than other MCTS programs, which
seem to try to win by exactly 0.5 points, and is approximately
five stones stronger than Version 11."
"Many Faces was running on a 32 core cluster provided by
Microsoft, and Mogo was running on up to several hundred
cores provided by the Dutch Huygens supercomputer."
( Text material published in the AMERICAN GO E JOURNAL may be
reproduced by any recipient: please credit the AGEJ as the source.
-- October 6, 2008; Volume 9, #52 )
http://www.grappa.univ-lille3.fr/icga/tournament.php?id=181
- regards
- jb
--------------------------------------------------------------
Kobayashi Maru
http://en.wikipedia.org/wiki/Kobayashi_Maru
--------------------------------------------------------------
Tim Tyler
> "Delayed this year while creator David Fotland rewrote
> his engine to use a combination of Monte Carlo Tree Search
> and its existing go knowledge, a prerelease of The Many Faces
> of Go version 12 -- which just won the International Computer
> Game Association's annual world championship ... The MFOG
> version 12's engine "combines the new Monte Carlo Tree
> Search algorithm with Many Faces' go knowledge, to give
> something stronger than either," Fotland tells the E-Journal.
> "It plays a more natural style than other MCTS programs, which
> seem to try to win by exactly 0.5 points, and is approximately
> five stones stronger than Version 11."
Pretty impressive. Version 11 was estimated as being 7kyu
by its author - that would make Version 11 around 2kyu -
though you /might/ need specialised hardware for that.
It Wasn't Me!!!
news:xfDGk.3198$iL2...@newsfe30.ams2...
> - wrote:
>
> > "Delayed this year while creator David Fotland rewrote
> > his engine to use a combination of Monte Carlo Tree Search
> > and its existing go knowledge, a prerelease of The Many Faces
> > of Go version 12 -- which just won the International Computer
> > Game Association's annual world championship ... The MFOG
> > version 12's engine "combines the new Monte Carlo Tree
> > Search algorithm with Many Faces' go knowledge, to give
> > something stronger than either," Fotland tells the E-Journal.
> > "It plays a more natural style than other MCTS programs, which
> > seem to try to win by exactly 0.5 points, and is approximately
> > five stones stronger than Version 11."
>
> Pretty impressive.
Not really. Consider MFOG traditional arch-rivals (those pre-MCTS era) - who was
it equipped with better go knowledge in past competitions? And let hypotheses
those rivals now applied the same MCTS technique - where would MFOG be by then?
> Version 11 was estimated as being 7kyu
> by its author - that would make Version 11 around 2kyu -
> though you /might/ need specialised hardware for that.
That's for sure...yes indeed.
Tim Tyler
> "Tim Tyler" <seem...@googlemail.com> wrote [...]
>>> "Fotland tells the E-Journal.
>>> "It plays a more natural style than other MCTS programs, which
>>> seem to try to win by exactly 0.5 points, and is approximately
>>> five stones stronger than Version 11."
>> Pretty impressive.
>
> Not really.
I was impressed. When was the last time there was a 5 stone
strength jump between consecutive version numbers?
It Wasn't Me!!!
> It Wasn't Me!!! wrote:
> > "Tim Tyler" <seem...@googlemail.com> wrote [...]
>
> >>> "Fotland tells the E-Journal.
> >>> "It plays a more natural style than other MCTS programs, which
> >>> seem to try to win by exactly 0.5 points, and is approximately
> >>> five stones stronger than Version 11."
> >> Pretty impressive.
> >
> > Not really.
>
> I was impressed. When was the last time there was a 5 stone
> strength jump between consecutive version numbers?
> --
Wait and I'm sure you be impressed still when other traditional go software
(with better go knowledge) applied the same MCTS technique...ten stones stronger
perhap than their previous version.
I think there is a syndicate behind all these just to after your money.
Gian-Carlo Pascutto
> Pretty impressive. Version 11 was estimated as being 7kyu
> by its author - that would make Version 11 around 2kyu -
> though you /might/ need specialised hardware for that.
Leela maintains a stable 1 kyu rating on an Intel Core 2 Quad on KGS.
This is pretty cheap hardware by now, <500 USD for a complete system.
MFGO claims 2 kyu on an Intel Core 2 Duo, so it is similar.
So at least 1-2 kyu level can be gotten by commodity hardware now.
--
GCP
-
jazze...@hotmail.com (-) wrote:
> 10 dim A%(5000):T=20:' ----------------- prmsp.ub prime spacing
> 20 for I=1 to 50000:J=4
> 30 inc A%(prm(I+1)-prm(I))
> 40 print I;prm(I);
> 42 while A%(J):print using(1,2),A%(J)/A%(J-2);:J+=2:wend:print
> 45 if I@T=0 then input T
> 50 next
>
> Note that after many iterations, pr(A%(2)) = pr(A%(4)), suggesting
> that the conjecture of infinite primes differenced by 2 is identical to
> the conjecture of inifinite primes differenced by 4. Values for A%
>reach a maximum at A%(6), where pr(A%(6)) = 8/5 * pr(A%(2))
Proving to be specious with additional calculation. Using Pari with:
a=vector(5000);n=2;
for(t=2,99999,p=nextprime(n+1);a[p-n]=a[p-n]+1;
for(p=2,50,printp1(a[p]," ");p=p+1)
above p > 10^10 still obtaining near equality for average
pr(gap=2) = pr(gap=4), however pr(gap=6)/pr(gap=2) is
approaching 9/5 instead of 8/5. Difficult to understand is
pr(gap=6) > pr(gap=12) > pr(gap=18) > pr(gap=10) > pr(gap=2).
- regards
- jb
-----------------------------------------------------
http://en.wikipedia.org/wiki/Liberalism
-----------------------------------------------------
-
jazze...@hotmail.com (-) wrote:
> Using Pari with:
> a=vector(5000);n=2;
> for(t=2,99999,p=nextprime(n+1);a[p-n]=a[p-n]+1;
> for(p=2,50,printp1(a[p]," ");p=p+1)
Transcribing error: this might work more correctly ... :-)
a=vector(5000);n=2;
for(t=2,99999,p=nextprime(n+1);a[p-n]=a[p-n]+1;n=p);
for(p=2,50,printp1(a[p]," ");p=p+1)
- regards
- jb
--------------------------------------------------------------
Catalan-Mersenne Number -- from Wolfram MathWorld
http://mathworld.wolfram.com/Catalan-MersenneNumber.html
--------------------------------------------------------------
-
>> for(p=2,50,printp1(a[p]," ");p=p+1)
For sufficiently large p, 1 < pr(gap=2) / pr(gap=4) < 1+e , where `e'
is arbitrarily small. Follows from the Hardy-Littlewood conjecture.
- regards
- jb
-------------------------------------------------------------
http://mathworld.wolfram.com/CousinPrimes.html
http://mathworld.wolfram.com/TwinPrimes.html
http://en.wikipedia.org/wiki/Cousin_prime
http://en.wikipedia.org/wiki/Twin_prime
-------------------------------------------------------------
-
>> jazze...@hotmail.com (-) wrote:
>>> Using Pari with:
>>> a=vector(5000);n=2;
>>> for(t=2,9999999, p=nextprime(n+1);a[p-n]=a[p-n]+1;n=p);
>>> for(p=2,50,printp1(a[p]," ");p=p+1)
jazze...@hotmail.com (-) wrote:
> For sufficiently large p, 1 < pr(gap=2) / pr(gap=4) < 1+e , where `e'
> is arbitrarily small. Follows from the Hardy-Littlewood conjecture.
We do not have any strict inequality of pr(gap=2) > pr(gap=4).
Further calculation beyond 10^10 reveals a crossover; perhaps
better to say: | pr(gap=2) - pr(gap=4) | < e, for large p, small e.
- regards
- jb
-------------------------------------------------------------
Ramanujan–Petersson conjecture
http://en.wikipedia.org/wiki/Ramanujan_conjecture
-------------------------------------------------------------