RE: [computer-go] Sylvain's results

[Tom Cooper]

There are multiple possible definitions of what it means for a player
to be the same strength on two different sized boards.  It is impossible
to pit a 9x9 player against a 19x19 player.  If two people
use different definitions of 'same strength', they are bound to disagree
about which size people are better at.

Off the top of my head, 'same strength'could be defined as
i) Same amount of experience of both games,
ii) Drops the same number of points over the course of the game
   when playing a perfect player,
iii) As ii), but scaled for the board area,
iv) Typically reads the same number of moves,
v) Same win rate on average when playing typical human players,
vi) Same win rate on average when playing Gnu Go.

At 05:42 12/04/2007, you wrote:


On Wed, 2007-04-11 at 17:29 -0700, David Fotland wrote:
> No, humans are much weaker on 9x9 than on 19x19.

David,

I saw this on Sensei's Library that indicates larger boards
are harder:

[ snip ]

  In [ext]The Theory & Practice of Go, Korschelt describes an
experimental 21x21 goban that he constructed and turned over to his
Master, Murase Shuho, for testing. He describes a sample game that was
played out to about 130 moves before ending. Only the first 57 were
shown.

Korschelt remarked that the game "took on a freer and more deeply
involved character, but ... at the same time the difficulty of keeping
command of the game grew at an extraordinary rate." He goes on to note
that on a 19x19 board, "too many unexpected situations turn up for
beginners," and speculates that if the board were to increase to 23x23,
"not even the best players could any longer maintain a comprehensive
view of the countless possible combinations."

[ snip ]

This seems to also match my intuition, even though I'm not as good a
player as you are.   If you extrapolate backwards, to 7x7, 5x5, 3x3
it seems clear that smaller boards are less complex and therefore
easier to master.   I cannot believe 9x9 is harder than 19x19 and
I don't care how strong the player is who says that - I don't believe
it.

- Don


___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

--
This email has been verified as Virus free
Virus Protection and more available at http://www.plus.net


___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] people are weaker at 9x9 go

[Darren Cook]

> No one plays 9x9 go.  There is no literature on 9x9 theory, and almost no
> examples of professional play. 

9x9 *is* played by professionals, at least in Japan, but naturally with
neither big prize money, nor students wanting to pay to learn it, it is
not played that seriously, and usually with quick time limits.

The best source of 9x9 professional games is the Mini-go TV series,
which finished a few years back, but should have close to 700 games if
you can track down the complete set. There is even a book published of
the game records. The closest I've managed to get to that is I know
someone who has received a photocopy of it...

As I've said before, when, for instance, a 9-dan and a 3-dan would meet
on the TV program, it was almost always the 9-dan who would win. Both
players would generally show knowledge of 9x9 opening theory (*),
meaning it is either studied behind closed doors, or (more likely) that
they were regular viewers!

Darren

*: Playing a move that the commentator would call "bad" in the first
half a dozen moves was rare. The games were almost always decided by
life and death mistakes or minor endgame mistakes (which is why the
weaker professional dan player generally lost).

___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] people are weaker at 9x9 go

[Magnus Persson]

Quoting Darren Cook <[EMAIL PROTECTED]>:


9x9 *is* played by professionals, at least in Japan, but naturally with
neither big prize money, nor students wanting to pay to learn it, it is
not played that seriously, and usually with quick time limits.


You can find 9x9 pro games here (but it might be necessary to register to
download them, I do not remember):

http://gobase.org/9x9/

My impression for going through these games is that the level of play is very
high. Except for the first opening moves I think even MoGo would have a really
hard time in many positions found in these games. The players seems to be able
to increase the complexity at every move.

-Magnus
___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Speed go next thing to explore

[compgo123]

High d players( > = 5d) play 1 5 games sometimes on internet. But not for 
serious games. I think 10 8 (5 8 at most) is about the fastest a serious 19x19 
game can be played comfortably between kyu players. It does not mean the game 
will last for a long time. Most lasts about 30 to 40 minutes. There are players 
playing 1 3 games, but their fun is not mainly in the game itself. There is 
also the netlag issue.
 
Daniel Liu 
 
 
-Original Message-
From: [EMAIL PROTECTED]
To: [EMAIL PROTECTED]
Sent: Thu, 12 Apr 2007 12:02 AM
Subject: [computer-go] Speed go next thing to explore


CrazyStone made appearance yesterday on KGS making rather impressive 
record. I think it rank peaked at 1d and ended at 2k. It was playing 
at speed limits of 10 minutes absolute, which seems hard for most 
humans. 
 
Also it seemed that people did not escape from the games. Rather few 
unfinished games. 
 
I think speed chess computers surpassed humans quite a while go? With 
specific approach to go programming thats a frontier where progress 
can be made? 
 
Petri Pitkänen 
-- Petri Pitkänen 
e-mail: [EMAIL PROTECTED] 
___ 
computer-go mailing list 
[EMAIL PROTECTED] 
http://www.computer-go.org/mailman/listinfo/computer-go/ 

AOL now offers free email to everyone.  Find out more about what's free from 
AOL at AOL.com.
___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Sylvain's results

[steve uurtamo]

>> No, humans are much weaker on 9x9 than on 19x19. 

> With all due respect, that's absurd.  If that were true, then all
> we would have to do is move to smaller boards if 19x19 were not
> challenging enough.

You've almost gotten it right.  In fact, 9x9 go is used to teach people
the rules of the game, and to help get them acquainted with fighting,
since it's basically a big fight from start to finish with too small of a
board to think about much else.  13x13 is used to get those same
students accustomed to thinking about the rest of the board, and
19x19 is the real game.  9x9 and 13x13 are cute, and it's quite
impressive that UCT-based programs are getting very good at playing
at the smaller size, but it's unrelated, in an important way, to the larger
size board.  they're different games entirely.

I think that I've played 15 games on a 9x9 board, and several
hundred on a 19x19 board.  I have no idea what it would mean to
be "the world's best 9x9 player", because nobody keeps track of
these things and I can't easily go get a book about 9x9 theory.
It's an unimportant game from the point of view of the typical 19x19
player, but it is a separate and distinct game -- if you study and play
19x19 go for most of your life, you will no longer be spending much time
playing the 9x9 version, and probably won't be all that interested in
'mastering' it.  similarly, if you play 9x9 go for too long, it will take you 
far
too long to understand why everything is so much harder on a 19x19 board.

So yes, humans become weaker at a game that they don't study,
write books about, spend thousands of years trying to master, etc.

It doesn't mean that 9x9 is harder *as a game*, it just means that
it holds no real interest *as a serious game*.

To belabor this even further, if you were to take everyone playing 19x19
go right now and convince them that no such thing ever existed and that
only 9x9 go exists, in a short amount of time they would all be very much
stronger 9x9 players (relative to the game itself), because 9x9 is a less deep
game.  something like "the best" opening move might be found, for instance,
whereas nobody knows this same fact on a 19x19 board.

s.





   

Never miss an email again!
Yahoo! Toolbar alerts you the instant new Mail arrives.
http://tools.search.yahoo.com/toolbar/features/mail/
___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] people are weaker at 9x9 go

[Don Dailey]

On Wed, 2007-04-11 at 23:02 -0600, Arend Bayer wrote:
> 
> I agree with all David Fotland has been saying. I think every strong
> go player would agree.

I also agree, now that I know he wasn't claiming 9x9 was more
profound.  

- Don


> In fact, I think I am stronger than most European 4ds at 9x9, simply
> because I realize it is a serious game, and I realize how early you
> have to start reading out variations deep enough until you can count
> out the endgame. Every other 4d could get there pretty quickly, and a
> professional who studies 9x9 for a week could probably write a book
> that would give me a lot more insights, and improve my 9x9 go much
> more than reading 5 books on 19x19 go. 
> 
> Arend
> ___
> computer-go mailing list
> [EMAIL PROTECTED]
> http://www.computer-go.org/mailman/listinfo/computer-go/

___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

RE: [computer-go] Sylvain's results

[Don Dailey]

On Wed, 2007-04-11 at 22:24 -0700, David Fotland wrote:
> > I cannot believe 9x9 is harder than 19x19 and
> > I don't care how strong the player is who says that - I don't 
> > believe it.
> > 
> > - Don
> 
> I don't believe it either :)  Sorry for the misunderstanding.  I was making
> a statement about how strong people are at a pair of games, not a statement
> about which game is harder.
> 
> David

I see that now, but I was starting to wonder how you could be so good at
GO and  programming and not understand this.But I see now that you
had just chosen a different frame of reference.You were comparing
go skill in humans to computers,  and to a lesser degree pointing out
that your training and skill has been directed at 19x19, not 9x9.   

- Don

> 

___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] people are weaker at 9x9 go

[terry mcintyre]

From: Darren Cook <[EMAIL PROTECTED]>

>The best source of 9x9 professional games is the Mini-go TV series,
>which finished a few years back, but should have close to 700 games if
>you can track down the complete set. 


Where would one begin to track down this show?

>There is even a book published of
>the game records. The closest I've managed to get to that is I know
>someone who has received a photocopy of it...







   

The fish are biting. 
Get more visitors on your site using Yahoo! Search Marketing.
http://searchmarketing.yahoo.com/arp/sponsoredsearch_v2.php___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Sylvain's results

[Don Dailey]

On Wed, 2007-04-11 at 23:15 -0700, terry mcintyre wrote:
> Humans don't have much experience with 9x9 go. In such tight confines,
> there is a premium for precise reading; there is little margin for
> error. It is much harder to escape, and harder to trade territory for
> influence. There is also, as Dave Fotland observed, little established
> literature on the 9x9 game. We take for granted a great deal of
> knowledge about fuseki and joseki and opening theory on the 19x19
> board; there is no equivalent store of knowledge about 9x9 openings.
> 
> That said, if humans were to invest the effort in the 9x9 game, the
> standard of play would rapidly increase. 

Yes,  the game would be easier to master, but the effort has been 
spent in mastering 19x19.

- Don



> 
> 
> 
> __
> Sucker-punch spam with award-winning protection.
> Try the free Yahoo! Mail Beta.
> ___
> computer-go mailing list
> [EMAIL PROTECTED]
> http://www.computer-go.org/mailman/listinfo/computer-go/

___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

[computer-go] Re: libego feedback

[Łukasz Lew]

On 4/11/07, Darren Cook <[EMAIL PROTECTED]> wrote:

Hi Lucasz,


Hi,


I spent some time studying your libego code today. It was educational
(I'd thought you were getting the speed by using assembler in key
functions, so I was surprised and pleased it is all pure C++).


I'm happy I refrained for going into asm (I was considering HLA). :)


I've some
questions and comments. If you want to reply to any of these on the
computer go mailing list that is fine by me.


OK :)




OPTIMIZATION

* In a few places you use a 1 element array. E.g. in class uct_t: tree_t
tree[1]; Is this faster than simply using tree_t tree; ? Is there a
standard name for this type of optimization?


Some time ago I was programming in C with g++ compiler style.
So I used no references. It was convenient convention for me to
declare every complex variable as one element array to have pointed
bind to the name. And I used those pointers
everywhere. I took this idea from Fruit (chess program) source.



* In a couple of places in class uct_t you use the trick of moving a
parameter to be a template parameter. This is one of the nice things
about C++ and I use it a lot (when I care about speed). I wondered why
you didn't use it more?


I tried to have two board_t::play functions - separate for black and white.
This way I could avoid some branches, but unfortunately it turned out
the it is slower.
I guess it's because jump prediction worked not good when it had to
deal with twice as big code. (code cache-swapping  could be a reason
too)



* uct.cpp, Many node_t functions can be marked const.


I have not tried to optimize uct yet.




ACCURACY

* In uct_t::do_playout(), when two passes in a row then you break and
score the game at that point. However I don't see anything to stop the
two passes happening anywhere in the tree, which would upset accuracy.


They can happen anywhere in the tree, pass is just another move.
I do not see why it should upset the accuracy.




Why not do simple_playout::run() after the two passes to make sure the
board is settled?  BTW, the only other exit point from that loop already
does this, so simple_playout::run() could be moved out of the loop to
just before the call to play_board->winner().


Because of seki :)



Note: I believe two UCT pass nodes in a row in a UCT tree is quite rare,
so while it will make little difference to playing strength, it should
also make little difference to speed, and code complexity stays
basically the same.



Because my own Go research goes in other direction and I get no
feedback about UCT
in libego therefore uct.cpp is just proof-of-concept by now.




UCT GENERALLY

* Is the use of is_mature() (to require 100 playouts before expanding
out a UCT node) simply to save memory (and some speed)? Or does it also
increase playing strength as well? I know this was discussed on the
computer go list, but only remember it for saving memory.



UCT adds one node each time it makes a new playout. Equally well it
could be 2 nodes or one node for each 2 playouts. It's arbitrary.
I chosen one node for each 100 playouts because it gets about 100*n
playouts to get each child visited n times.




* In a real game, the tree (i.e. the uct_t object) is thrown away by
each call to genmove. Wouldn't it be better to make this a global, and
then when a move is chosen just delete the sub-trees for the moves that
weren't chosen.

Pros: It will start with a more informative tree, instead of having to
build it from scratch each time. This should be the equivalent of an
extra 20-50% playouts, for free.

Cons: More code complexity. Also, the program can start playing more
weakly if the user uses undo_move() a lot (not weaker than without the
idea, but it may be more confusing).

Note: More memory usage, but no different to if number of playouts were
increased, so I don't think this is a con.


It's a good idea, It was a next thing on my UCT to-do list. But as I
said before.
uct.cpp is very far in quality from the rest of libego library.




Darren

P.S. Thanks very much for this library. It allows me to immediately
experiment with some UCT ideas without having to spend a month
developing, debugging and optimizing my own code. (If those ideas turn
out to be good then of course afterwards I will spend that month writing
my own version. But for the moment I can work on the fun bit!) If my
experiments work out well I'll let you know.


Thanks for Your thanks and remarks. That's my motivator ;)

Regards,
Łukasz Lew

PS
Can You connect MLSN to all those dictionaries that Wakan uses?
http://wakan.manga.cz/
(EDICT primarily)




--
Darren Cook
http://dcook.org/mlsn/ (English-Japanese-German-Chinese free dictionary)
http://dcook.org/work/ (About me and my work)
http://dcook.org/work/charts/  (My flash charting demos)

___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: libego feedback

[Chris Fant]

> * In uct_t::do_playout(), when two passes in a row then you break and
> score the game at that point. However I don't see anything to stop the
> two passes happening anywhere in the tree, which would upset accuracy.

They can happen anywhere in the tree, pass is just another move.
I do not see why it should upset the accuracy.


Most people do not allow the pass move until no other non-eye-filling
moves remain.  By accuracy,  think he means that the winner of a game
that is not completely finished cannot easily be determined.
___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] people are weaker at 9x9 go

[Chris Fant]

>The best source of 9x9 professional games is the Mini-go TV series,
>which finished a few years back, but should have close to 700 games if
>you can track down the complete set.


Where would one begin to track down this show?


http://home.wwgo.jp/minigo/
___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Sylvain's results

[forrest curo]

  In [ext]The Theory & Practice of Go, Korschelt describes an
experimental 21x21 goban that he constructed and turned over to his
Master, Murase Shuho, for testing. 
  
Anyone who wants to try this can play email-style games on 
dragongoserver.net -- up to 25X25.


Forrest Curo
___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

[computer-go] Computer match time

[compgo123]

I noticed that GnuGo is alsousing 10 min.time limit. People's complaining about 
Crazystone's time can be viewed as the program is getting better.
 
GnuGo uses 10 min time limit. It's short for serious human games. I think part 
of the reason is that computer program didn't play that well before and not 
warrent a serious game. But now computer program is getting stronger and they 
can play a serious game. Probably we should set up a standard for matches 
between computer programs and humans. A program should follow this standard, 
otherwise the ranking and results is not recognized. In principle longer time 
is to the advantage of computer program.
 
I suggest we setthe the time standard as following
 
1. For test games the match time is 10 min.each. The rank and results obtained 
are not recognized as representing the computer state of the art.
2. For non tournament matches the time is set as 23 min. each. The rank and 
results obtained will be recognized as representing the state of art for the 
program.
3. For tournament matches the time is set as 30 min. each. The rank and results 
obtained will be recognized.
4. For mtahces between the computer programs and the professional Go players 
the time is set as 1 hour each. The rank and results will be recognized.
 
 
Above is for 19x19games. Since the byo yomi is hard to implement in computer 
programs, we may just use one time limit.
 
This is just my suggestions.
 
Daniel Liu 

AOL now offers free email to everyone.  Find out more about what's free from 
AOL at AOL.com.
___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] people are weaker at 9x9 go

[Sanghyeon Seo]

> >The best source of 9x9 professional games is the Mini-go TV series,
> >which finished a few years back, but should have close to 700 games if
> >you can track down the complete set.
> Where would one begin to track down this show?
http://home.wwgo.jp/minigo/


MINIGO ICHIBAN SHOUBU would be a reading of Japanese ミニ碁一番勝負.
(Translation would be "Mini go sudden death".)

Indeed, Japanese Wikipedia has an article on this TV series:
http://ja.wikipedia.org/wiki/%E3%83%9F%E3%83%8B%E7%A2%81%E4%B8%80%E7%95%AA%E5%8B%9D%E8%B2%A0

According to the article, this weekend early morning 30-minute program
continued from 1987 to 2002. Long running series!

--
Seo Sanghyeon
___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: libego feedback

[Don Dailey]

On Thu, 2007-04-12 at 15:25 +0200, Łukasz Lew wrote:
> > * In a real game, the tree (i.e. the uct_t object) is thrown away by
> > each call to genmove. Wouldn't it be better to make this a global,
> and
> > then when a move is chosen just delete the sub-trees for the moves
> that
> > weren't chosen.
> >
> > Pros: It will start with a more informative tree, instead of having
> to
> > build it from scratch each time. This should be the equivalent of an
> > extra 20-50% playouts, for free.

Lazarus does this, but after the opponent moves, there is very little
tree remaining.   I do this out of general principle (might as well 
take the free tree no matter how tiny)  but in practical terms I doubt 
it adds more than 5 ELO points to the program.   

But there is another way to make it pay off more reasonably.  After you
make a move,  kill all the siblings and start thinking as if you were
the opponent.   Then when it's your turn again,  prune the tree again
to the relevant branches.   Then you will get a modest improvement.

However, Lazarus thinks on the opponent time by assuming the program
will play what Lazarus thinks is best.  I did some informal calculations
and I believe this is more effective.  Lazarus get the full benefit of
the opponents time if it predicts correctly, otherwise almost no benefit
but it is able to predict a lot of moves, especially in sharp positions
where this is important.


> > Cons: More code complexity. Also, the program can start playing more
> > weakly if the user uses undo_move() a lot (not weaker than without
> the
> > idea, but it may be more confusing).
> >
> > Note: More memory usage, but no different to if number of playouts
> were
> > increased, so I don't think this is a con.
> 
> It's a good idea, It was a next thing on my UCT to-do list. But as I
> said before.
> uct.cpp is very far in quality from the rest of libego library.
> 

___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Speed go next thing to explore

[Don Dailey]

On Thu, 2007-04-12 at 08:02 +0300, Petri Pitkanen wrote:
> CrazyStone made appearance yesterday on KGS making rather impressive
> record. I think it rank peaked at 1d and ended at 2k. It was playing
> at speed limits of 10 minutes absolute, which seems hard for most
> humans.

> Also it seemed that people did not escape from the games. Rather few
> unfinished games.
> 
> I think speed chess computers surpassed humans quite a while go? With
> specific approach to go programming thats a frontier where progress
> can be made?


Speed chess isn't hard for humans at all, they play quite well.  But
it's been clearly shown that humans improve with time in a very 
similar way to computers in chess, but even more so.   At really
fast time-controls this improvement is very obvious and clear,
but we just don't notice it as much at longer time controls even
though it is there just as strong.

One thing is clear, humans improve MORE at longer time-controls 
than computers, despite most people intuition that this doesn't
happen,
so at longer time controls humans have a definite advantage.  It's
also been shown that strong computers improve more quickly than
weak computers.   It seems that the quality of the evaluation
function (both in humans and computers) is a factor.   That's
why in recent years much more emphasis is on the evaluation function
of computer chess.  In previous times you could just build a super-fast
searcher and if it had a reasonable evaluation function it could
compete.  Not true any more.  

CrazyStone (any UCT program) is going to be relatively weak at the
10 minute time control.   If you double the time-control to make
it easier for humans, this will  help but don't expect a major
turnaround.  CrazyStone will play a much stronger game at 20 minutes
also.  It's not clear how much it will help the humans, but I would
not expect more than 1 kyu.   If you make it more like tournament
time controls you might make a 2 or 3 kyu difference (if that) but
it won't suddenly drop to the high or mid kyu range.  If it's
realistically 2 kyu strength at 10 minutes, it's probably still
stronger than 5 kyu at any reasonable time control.  I would put
it at 3 or 4 if 2kyu is correct at 10 minutes.

However, there is also another factor.  As more computers play and
compete, humans will learn to beat them.   I believe the UCT style
will be much more difficult to take advantage of because they do 
not have idiosyncratic weaknesses to the extent of older generation 
programs, but still there will be this factor pushing their ranks 
down a bit. 

There was once a study of chess players and their abilites at 5 minute
chess.   The study started with a questionaire.   It's extremely common
for chess players to say, "I'm really quite terrible at speed chess, I
blunder all the time and I'm just much better at longer time controls."

The questionaire verified that many chess players feel that other
players
are better than them at speed chess.   But the study showed
with empirical data that
your tournament ratings accurately predicted speed chess performance, 
even of the players who thought they were unusually weak.  
Also, the issue of "blunders" was addressed.   Blunders are often cited
as the problem with speed chess, that the extra pressure of the clock
makes too many blunders and the games are somehow "not real" as someone
here recently complained.   Of course in chess everytime there is a
blunder the ego says the game didn't really count, it was a "fluke", it
was not characteristic of my usually strong game.

I cannot remember the exact experimental setup of the study,  but I
remember that the "fluke" factor was shown to be false too.   Of course
you do blunder more, but so do computers.  Humans always had the feeling
they would stop blundering if there were give just a little bit more
time but the fact of the matter is that humans will blunder at any 
time-control.  They blunder less at longer time-controls but the blunder
rate and quality of moves does not suddenly improve at some magic 
point.   It's a very mudane and predictable fall-off.

A lot of players (despite the fact that they feel they are weak at
speed chess) somehow believe that their best chance against a really
strong player is speed chess, because they might capitalize on a
blunder.   This is not the case.   If anything, the experience (and
better evaluation function) makes the better player even more superior.

There are some players who specialize in speed chess so I'm not saying
it isn't possible to better at one than the other.   You can train
specifically for speed chess and gain some skills that improve your
results specifically for this aspect of the game. 

I'm taking well known concepts from chess and computer chess and
extrapolating to GO.   There is no reason to believe this is 
somehow completely irelevant to the game of GO. Even though GO
is a unique game, it's still played in the same universe and these
seem to be universal "truths"

Re: [computer-go] Computer match time

[Jason House]

Not having byo yomi because it's tough to code isn't really a good
argument.  If we want (non-computer-go) people to take the results
seriously, the game timing should be the same as what people naturally do.
I personally am hesitant to play games with fixed time limits because I
don't trust myself the use the time appropriately throughout the game.

On 4/12/07, [EMAIL PROTECTED] <[EMAIL PROTECTED]> wrote:


 I noticed that GnuGo is alsousing 10 min.time limit. People's complaining
about Crazystone's time can be viewed as the program is getting better.

GnuGo uses 10 min time limit. It's short for serious human games. I think
part of the reason is that computer program didn't play that well before and
not warrent a serious game. But now computer program is getting stronger and
they can play a serious game. Probably we should set up a standard for
matches between computer programs and humans. A program should follow this
standard, otherwise the ranking and results is not recognized. In principle
longer time is to the advantage of computer program.

I suggest we setthe the time standard as following

1. For test games the match time is 10 min.each. The rank and results
obtained are not recognized as representing the computer state of the art.
2. For non tournament matches the time is set as 23 min. each. The rank
and results obtained will be recognized as representing the state of art for
the program.
3. For tournament matches the time is set as 30 min. each. The rank and
results obtained will be recognized.
4. For mtahces between the computer programs and the professional Go
players the time is set as 1 hour each. The rank and results will be
recognized.


Above is for 19x19games. Since the byo yomi is hard to implement in
computer programs, we may just use one time limit.

This is just my suggestions.

Daniel Liu
 --
AOL now offers free email to everyone. Find out more about what's free
from AOL at *AOL.com * .

___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Speed go next thing to explore

[terry mcintyre]

An advantage of speed go for computers is Moore's Law: as processors improve, 
computers do more work in a given time. No such law applies to human brains. It 
is reasonable to expect a doubling in computer go performance every eighteen 
months or so - perhaps more so with the trend to multicore processors.  More 
speed, more computation, fewer blunders, higher rankings.

Can anyone share their estimates of RAM used per 100k playouts, or other 
appropriate measure?









   

8:00? 8:25? 8:40? Find a flick in no time 
with the Yahoo! Search movie showtime shortcut.
http://tools.search.yahoo.com/shortcuts/#news___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Computer match time

[Don Dailey]

On Thu, 2007-04-12 at 11:18 -0400, Jason House wrote:
> Not having byo yomi because it's tough to code isn't really a good
> argument.  If we want (non-computer-go) people to take the results
> seriously, the game timing should be the same as what people naturally
> do.  I personally am hesitant to play games with fixed time limits
> because I don't trust myself the use the time appropriately throughout
> the game.  

Byo yomi of course is a better time control for computers, even if it's
slightly more difficult to implement.   

It also has the advantage that humans may consider it more "fair" to
them even though it isn't.

- Don




___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: libego feedback

[Chris Fant]

But there is another way to make it pay off more reasonably.  After you
make a move,  kill all the siblings and start thinking as if you were
the opponent.   Then when it's your turn again,  prune the tree again
to the relevant branches.   Then you will get a modest improvement.


This is what my engine does.  But it is still very weak.  But I'm
still working at it.
___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Speed go next thing to explore

[terry mcintyre]

byo yomi is probably an advantage for a computer program. As long as one's move 
is made within the byo yomi period, one may continue playing indefinitely. If a 
UCT program plays a "good enough" move just before the flag drops, this will 
keep human opponents under a great deal of pressure.

I am speaking from experience -- in a recent game against a slow opponent, I 
hoped to catch him in time trouble -- I had five minutes on the clock when he 
went into byo yomi, but he altered his playing style to pick any acceptable 
move just before his period expired, and I used a good bit of my time to find a 
reasonable counter, lol. No doubt, with practice, I'll manage my time better, 
but this is something which can surely be exploited by any algorithm, such as 
UCT, which progressively improves play as more time is available.




   

Finding fabulous fares is fun.  
Let Yahoo! FareChase search your favorite travel sites to find flight and hotel 
bargains.
http://farechase.yahoo.com/promo-generic-14795097___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Computer match time

[Don Dailey]

On Thu, 2007-04-12 at 10:38 -0400, [EMAIL PROTECTED] wrote:
> I noticed that GnuGo is alsousing 10 min.time limit. People's
> complaining about Crazystone's time can be viewed as the program is
> getting better.
>  
> GnuGo uses 10 min time limit. It's short for serious human games. I
> think part of the reason is that computer program didn't play that
> well before and not warrent a serious game. But now computer program
> is getting stronger and they can play a serious game. Probably we
> should set up a standard for matches between computer programs and
> humans. A program should follow this standard, otherwise the ranking
> and results is not recognized. In principle longer time is to the
> advantage of computer program.

Actually,  even though it makes the program stronger it benefits the
human
player even more.   But we can of course test that assertion.

 
> I suggest we setthe the time standard as following
>  
> 1. For test games the match time is 10 min.each. The rank and results
> obtained are not recognized as representing the computer state of the
> art.
> 2. For non tournament matches the time is set as 23 min. each. The
> rank and results obtained will be recognized as representing the state
> of art for the program.
> 3. For tournament matches the time is set as 30 min. each. The rank
> and results obtained will be recognized.
> 4. For mtahces between the computer programs and the professional Go
> players the time is set as 1 hour each. The rank and results will be
> recognized.

I think it's a good thing to use the same time-control so that we are
always comparing apples to apples.   But I think 10 minutes is as good
as any.   If the games are being seriously rated, I have no problem with
some other time setting.  

I don't like byo yomi either for computers.  Is there a kind of time
control
that simply adds n seconds to each move?   Here is an example:  You
start
with 5 minutes, but 5 seconds are added to your clock after each
move.   
This means you will never be in such extreme time pressure that you have
to play instantly and you can even build up time by playing moves more
quickly than 5 minute.This is used in chess sometimes.  

If I understand byo yomi correctly,  it does not allocate time in a fair
way.  One player can use less time, but lose the game for instance 
because he spent the time differently.   Time you save is not carried
over.   Do I understand it correctly?   

It would be bad for humans against computers (if my understanding 
is correct) because the best strategy for computers would be to push
the time-control to maximum advantage. It's rather like university
funding where you have to spend a lot of money in a hurry or else you
lose it.  

- Don
 
 

> Above is for 19x19games. Since the byo yomi is hard to implement in
> computer programs, we may just use one time limit.
>  
> This is just my suggestions.
>  
> Daniel Liu 
> 
> __
> AOL now offers free email to everyone. Find out more about what's free
> from AOL at AOL.com.
> 
> ___
> computer-go mailing list
> [EMAIL PROTECTED]
> http://www.computer-go.org/mailman/listinfo/computer-go/

___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Computer match time

[Rémi Coulom]

Don Dailey wrote:

On Thu, 2007-04-12 at 11:18 -0400, Jason House wrote:
  

Not having byo yomi because it's tough to code isn't really a good
argument.  If we want (non-computer-go) people to take the results
seriously, the game timing should be the same as what people naturally
do.  I personally am hesitant to play games with fixed time limits
because I don't trust myself the use the time appropriately throughout
the game.  



Byo yomi of course is a better time control for computers, even if it's
slightly more difficult to implement.   


It also has the advantage that humans may consider it more "fair" to
them even though it isn't.

- Don
  


byoyomi on KGS has this problem:
http://computer-go.org/pipermail/computer-go/2006-February/004317.html
I believe 10 min / game is a usual time control for humans (that's how 
we play blitz at my Go club). It is an advantage for the bot, as 
compared to longer time control, that is for sure. If Crazy Stone 
reaches [1d] on KGS, I will make it play slower :-)


Rémi
___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Computer match time

[Jason House]

I recently found that the output from time_left looks identical for main
time and byo yomi time.  Both give zero stones remaining.  I did discover
that "time_left 0 0" is sent when main time runs out and byo yomi begins.  I
didn't check if that occurs after each byo yomi period is lost as well.

To overcome this issue, I set my bot to use Canadian time.  My engine
doesn't yet support the time_settings command, so I wanted to make the
current state very clear (allowing each time_left command to be interpreted
without prior state).

On 4/12/07, Rémi Coulom <[EMAIL PROTECTED]> wrote:


Don Dailey wrote:
> On Thu, 2007-04-12 at 11:18 -0400, Jason House wrote:
>
>> Not having byo yomi because it's tough to code isn't really a good
>> argument.  If we want (non-computer-go) people to take the results
>> seriously, the game timing should be the same as what people naturally
>> do.  I personally am hesitant to play games with fixed time limits
>> because I don't trust myself the use the time appropriately throughout
>> the game.
>>
>
> Byo yomi of course is a better time control for computers, even if it's
> slightly more difficult to implement.
>
> It also has the advantage that humans may consider it more "fair" to
> them even though it isn't.
>
> - Don
>

byoyomi on KGS has this problem:
http://computer-go.org/pipermail/computer-go/2006-February/004317.html
I believe 10 min / game is a usual time control for humans (that's how
we play blitz at my Go club). It is an advantage for the bot, as
compared to longer time control, that is for sure. If Crazy Stone
reaches [1d] on KGS, I will make it play slower :-)

Rémi
___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Computer match time

[Arend Bayer]

On Thu, Apr 12, 2007 at 10:56:59AM -0400, Don Dailey wrote:
> On Thu, 2007-04-12 at 10:38 -0400, [EMAIL PROTECTED] wrote:
> I don't like byo yomi either for computers.  Is there a kind of time
> control
> that simply adds n seconds to each move?   Here is an example:  You
> start
> with 5 minutes, but 5 seconds are added to your clock after each
> move.   
> This means you will never be in such extreme time pressure that you have
> to play instantly and you can even build up time by playing moves more
> quickly than 5 minute.This is used in chess sometimes.  

This is called Fischer time and hardly used in go yet, even though I
agree with you that it would be a superior time control for go as well.
In fact, if we could just get people used to it, I think almost
everybody would prefer it, but go players are conservative... 

Arend

___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] people are weaker at 9x9 go

[Hideki Kato]

Sanghyeon Seo: <[EMAIL PROTECTED]>:
>> > >The best source of 9x9 professional games is the Mini-go TV series,
>> > >which finished a few years back, but should have close to 700 games if
>> > >you can track down the complete set.
>> > Where would one begin to track down this show?
>> http://home.wwgo.jp/minigo/
>
>MINIGO ICHIBAN SHOUBU would be a reading of Japanese .
>(Translation would be "Mini go sudden death".)

Are you kidding? "ICHIBAN SHOUBU" means one game match. 
#Meijin-sen is a "NANABAN SHOUBU" that means seven games match. 
##"ICHI" is one and "NANA" is seven, in Japanese.

- gg

>Indeed, Japanese Wikipedia has an article on this TV series:
>http://ja.wikipedia.org/wiki/%E3%83%9F%E3%83%8B%E7%A2%81%E4%B8%80%E7%95%AA%E5%8B%9D%E8%B2%A0
>
>According to the article, this weekend early morning 30-minute program
>continued from 1987 to 2002. Long running series!
>
>-- 
>Seo Sanghyeon
> inline file
>___
>computer-go mailing list
>[EMAIL PROTECTED]
>http://www.computer-go.org/mailman/listinfo/computer-go/
--
[EMAIL PROTECTED] (Kato)
___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Computer match time

[Don Dailey]

On Thu, 2007-04-12 at 18:37 +0200, Rémi Coulom wrote:
> byoyomi on KGS has this problem:
> http://computer-go.org/pipermail/computer-go/2006-February/004317.html
> I believe 10 min / game is a usual time control for humans (that's
> how 
> we play blitz at my Go club). It is an advantage for the bot, as 
> compared to longer time control, that is for sure. If Crazy Stone 
> reaches [1d] on KGS, I will make it play slower :-)
> 
> Rémi 

It is clear to me too that faster time control is better for the bot.

- Don


___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Speed go next thing to explore

[Don Dailey]

On Thu, 2007-04-12 at 09:28 -0700, terry mcintyre wrote:
> byo yomi is probably an advantage for a computer program. As long as
> one's move is made within the byo yomi period, one may continue
> playing indefinitely. If a UCT program plays a "good enough" move just
> before the flag drops, this will keep human opponents under a great
> deal of pressure.

Yes, there are tricks but it's very dangerous.   Some great advice
I recieved when I actually played tournament chess was ignore the
opponents clock and focus on your own moves.

The reason that advice is good is that when you are in time
pressure, your skill jumps up, you are espcecially focused,
and if your opponent plays fast to exploit your time pressure,
he is probably not helping his own game - a good situation,
you play better, he plays worse.  Just the opposite of what
you might think.

In fact, I have also heard the advice that if your opponent
is in serious time pressure,  take extra time and
the adrenaline will wash him out - creating a lot of stress.
You can only stay at attention but so long.   



- Don


___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] people are weaker at 9x9 go

[Sanghyeon Seo]

>MINIGO ICHIBAN SHOUBU would be a reading of Japanese .
>(Translation would be "Mini go sudden death".)
Are you kidding? "ICHIBAN SHOUBU" means one game match.


Eh, doesn't "sudden death" mean "one game match" in English?

And I know how to read Japanese, thank you.

--
Seo Sanghyeon
___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] The dominance of search (Suzie v. GnuGo)

[Hideki Kato]

[EMAIL PROTECTED]: <[EMAIL PROTECTED]>:
>Remi,
> 
>Could you change your time as 10 min. plus 8 min. byo yomi? Otherwise it's too 
>short for
>human. It's difficult to get a real results.

Strogly agree! It's too short for me, 4 kyu, to have a meaningful
game with Crazy Stone. I beg you too, Remi.
#GNU bots are 10 min + 20 sec x 5.

- gg

>Daniel Liu
> 
> 
>-Original Message-
>From: [EMAIL PROTECTED]
>To: [EMAIL PROTECTED]
>Sent: Wed, 11 Apr 2007 3:33 PM
>Subject: Re: [computer-go] The dominance of search (Suzie v. GnuGo)
>
>
>[EMAIL PROTECTED] wrote: 
>> I also find this kind of information very interesting and useful. Now > I 
>> have a better
>feel for what kind of scaling is realistic to try for > and how to measure it. 
>> > Putting some recent data points together, it look like giving Mogo 2 > 
>> > orders of
>magnitude more computer power would result in low dan level > 19x19 play? Not 
>the sort of
>thing one can pull out of a back pocket, > but tantalizing. 
>> > I would be very interested to see equivalent scaling numbers from > 
>> > CrazyStone, if Remi
>would be so kind. 
>> > - Dave Hillis 
>I don't have time to do this right now. I have connected Crazy Stone to play 
>10 minute blitz
>on KGS, and will leave it all night long. It reached [1d] for a while, and is 
>currently [1k].
>There is a big surprise effect when humans play Crazy Stone. The first moves 
>look so stupid.
>Usually, a human player will lose the first game, then it will win the next 
>ones. 
> 
>Rémi 
>___ 
>computer-go mailing list 
>[EMAIL PROTECTED] 
>http://www.computer-go.org/mailman/listinfo/computer-go/ 
>
>AOL now offers free email to everyone.  Find out more about what's free from 
>AOL at AOL.com.
> inline file
>___
>computer-go mailing list
>[EMAIL PROTECTED]
>http://www.computer-go.org/mailman/listinfo/computer-go/
--
[EMAIL PROTECTED] (Kato)
___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Computer match time

[compgo123]

Let's add byo yomi time. So it would be
 
1. For test games the match time is 10 min.each plus 5 min for 25 stones byo 
yomi. The rank and results obtained are not recognized as representing the 
computer state of the art.
2. For non tournament matches the time is set as 20 min. each plus 5 min for 25 
stones byo yomi. The rank and results obtained will be recognized as 
representing the state of art for the program.
3. For tournament matches the time is set as 30 min. each plus 5 min for 25 
stones byo yomi. The rank and results obtained will be recognized.
4. For mtahces between the computer programs and the professional Go players 
the time is set as 1 hour each plus the standard byo yomi pro games. The rank 
and results will be recognized.
5. The game uses the Chinese game rule and score counting method.
6. Mistake in calculating score and identifying captured stone is allowed.
 
The reason to choose 20 min. playing time each is because it's the typical 
match time (total 40 min.) between human players. Someone more familiar with 
the KGS and IGS statistics can back me up on this.
 
 
 
-Original Message-
From: [EMAIL PROTECTED]
To: [EMAIL PROTECTED]
Sent: Thu, 12 Apr 2007 11:10 AM
Subject: Re: [computer-go] Computer match time


On Thu, 2007-04-12 at 11:18 -0400, Jason House wrote:
> Not having byo yomi because it's tough to code isn't really a good
> argument.  If we want (non-computer-go) people to take the results
> seriously, the game timing should be the same as what people naturally
> do.  I personally am hesitant to play games with fixed time limits
> because I don't trust myself the use the time appropriately throughout
> the game.  

Byo yomi of course is a better time control for computers, even if it's
slightly more difficult to implement.   

It also has the advantage that humans may consider it more "fair" to
them even though it isn't.

- Don




___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

AOL now offers free email to everyone.  Find out more about what's free from 
AOL at AOL.com.
___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

[computer-go] Computer tournament at next US Go Congress?

[Peter Drake]

The next US Go Congress will be held July 28 - August 4, 2007, in  
Lancaster, Pennsylvania:


http://congress.usgo.org/

Is anyone else planning to attend?

I would, at the very least, like to gather a casual meeting of Go  
programmers.


It might also be interesting (and a good media opportunity) to hold a  
computer tournament. I see three possibilities:


1) Not require the machines in question to be physically present.  
This would require a certain amount of faith that there were really  
programs at the other end, so it would probably not be appropriate to  
offer any prize for this. This is the easiest to organize, as it  
amounts to convincing Nick to hold that month's KGS tournament on an  
appropriate date.


2) Require people to bring their machines. This is a bit of a pain  
for the programmers.


3) Require all programs to run on identical machines provided on- 
site. Hopefully someone could arrange a local computer vendor to loan  
some fast machines in return for the free press. This approach would,  
unfortunately, exclude SlugGo and anyone that's been building special- 
purpose hardware. I might try to swing this at the 2008 Congress,  
which will be held here in Portland, Oregon, the home of a very large  
Intel facility.


Such a tournament could be followed up by an exhibition match against  
strong humans (including professionals).


What do you think? Is there anyone in the Pennsylvania area that  
would be willing to organize such a tournament this year?


Peter Drake
http://www.lclark.edu/~drake/



___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] people are weaker at 9x9 go

[Darren Cook]

>> The best source of 9x9 professional games is the Mini-go TV series,
>> which finished a few years back, but should have close to 700 games if
>> you can track down the complete set. 
> 
> Where would one begin to track down this show?

Others have already given the main links. While hunting for some others
I found this:
  http://pr.fujitsu.com/jp/news/1998/Mar/minigo.html

I *think* this is saying it has the pro games, with commentary, from the
first 11 years of the show. This amazon page [1] seems to be the same
thing, but says it is out of stock. Does anyone have this CD and can say
what is on it?

Darren


[1]:
http://www.amazon.co.jp/%E6%97%A5%E6%9C%AC%E5%9B%B2%E7%A2%81%E3%82%BD%E3%83%95%E3%83%88-%E5%AF%BE%E5%B1%80%E3%83%9F%E3%83%8B%E7%A2%81-%E3%81%93%E3%82%8C%E3%81%A7%E3%81%82%E3%81%AA%E3%81%9F%E3%82%82%E5%9B%B2%E7%A2%81%E5%88%9D%E6%AE%B5/dp/B8I2TH

___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] OT: sudden death

[Darren Cook]

>> >MINIGO ICHIBAN SHOUBU would be a reading of Japanese .
>> >(Translation would be "Mini go sudden death".)
>> Are you kidding? "ICHIBAN SHOUBU" means one game match.
> 
> Eh, doesn't "sudden death" mean "one game match" in English?

Sudden death means that as soon as one scores a point/goal they become
the winner; it has the implication that the main games has been played
resulting in a draw and this is being used to break the draw [1]. IMHO,
"one game match" or simply "match" is a better translation of ichiban
shoubu.

Darren

[1]: http://www.answers.com/main/ntquery?s=sudden+death&gwp=13
___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Speed go next thing to explore

[Darren Cook]

> Can anyone share their estimates of RAM used per 100k playouts, or
> other appropriate measure?

As part of my study of libego I've had the same question.

100K playouts (from an empty board) creates an UCT tree of around 21,000
nodes, which is using 795K. The node class uses 20 bytes, which is
getting packed to 24 bytes; it looks possible to squeeze it into 16
bytes, but I think that will have a negative impact on speed.

200K playouts creates a tree of around 42,000 nodes, using 1.5MB. So it
seems to be scaling constantly.

The rest of the executable is using just under 4MB of memory.

BTW, the above memory usage is using the object_pool class with default
settings. Actual memory usage for 21,000 nodes is 500K, and 1MB for
42,000 nodes.

Also, this is using the default mature_bias_threshold of 101. If I lower
the threshold to 2, with 100K playouts I get a tree of approx 1.1
million nodes, using 25MB of memory.

Darren

P.S. 100K UCT takes about 3 secs (Celeron 2.8Ghz).
I just tried a 1 million UCT (mature_bias_threshold set back to 101) and
the tree it makes is just over 200,000 nodes. It takes 37 seconds, and
uses 7.2M for the UCT tree (11.1 MB total for the program).

I.e. very dependent on mature_bias_threshold, but it seems it is CPU
bound, not memory bound.

___
computer-go mailing list
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/