Re: [computer-go] Re: computer-go Digest, Vol 43, Issue 8

[Magnus Persson]

Quoting Don Dailey <[EMAIL PROTECTED]>:



My feelings on this seem to match at least one source:

Look here:http://senseis.xmp.net/?Komi

Here is an excerpt:

It is widely believed that the correct komi is independent of board size
for all but the smallest boards. For area scoring, this would give 7 for
9x9+, 8 for 8x8, 7 for 7x7, 4 for 6x6, 25 for 5x5 (w cannot live), 0 for
4x4, 9 for 3x3, 4 for 2x2 with a superko rule, and 0 for 1x1. (these
need to be verified)


I corrected this sensei page to give komi 9 for 7x7 and added a link  
to the sgf file John Tromp provides with the analysis.


I played a lot with Valkyria on 7x7 and although it proofs nothing it  
is really convincing that 9 is the correct komi.


Similarily one find that in very simple games on 9x9, but where the  
moves are good solid shape white almost always win with 0.5 points  
with 7.5 komi. Thus if one designs and opening book for 9x9 (as I  
tried to) one should try to complicate things as black and play simple  
as white early on. The exception may be when black opens at 5,5 in the  
center. Then often white ends up trying to live with two groups which  
can be very difficult against a competent opponent.


-Magnus

___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: computer-go Digest, Vol 43, Issue 8

[Christoph Birk]

On Feb 11, 2008, at 9:39 PM, Don Dailey wrote:

My feelings on this seem to match at least one source:

Look here:http://senseis.xmp.net/?Komi

Here is an excerpt:

It is widely believed that the correct komi is independent of board  
size
for all but the smallest boards. For area scoring, this would give  
7 for
9x9+, 8 for 8x8, 7 for 7x7, 4 for 6x6, 25 for 5x5 (w cannot live),  
0 for

4x4, 9 for 3x3, 4 for 2x2 with a superko rule, and 0 for 1x1. (these
need to be verified)


For 7x7 the komi is 9:
 http://senseis.xmp.net/?7x7BestPlay

Christoph

___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] gpugo

[dhillismail]

Hi Florian,

?

 That sounds like an interesting project. I've looked at this a little. If 
it were me, I would start by implementing the simplest possible MC playouts in 
gpgpu,?as efficiently as possible, before looking at adding things like 
patterns.?

 I also heartily encourage you to learn about genetic algorithms. (But?not 
really because of?RAVE.)

- Dave Hillis

-Original Message-
From: Florian Erhardt <[EMAIL PROTECTED]>
To: computer-go 
Sent: Mon, 11 Feb 2008 4:06 pm
Subject: [computer-go] gpugo



Hello !?
?
My name is Florian Erhardt, I am a bachelor student of computer sciences and am 
in the process of optimizing libEGO for gpgpu. For now I implemented the SFMT 
(even I can do copy and paste) on the gpu and am now atomizing the MC to be 
done by the gpu. If everything works as I planned it, I'll have a releasable 
program in a few months (I still have to learn lots about parallelizing, gpu 
programming and programming in general [maybe leaning about genetic algorithms 
can't hurt either - The RAVE algorithm is more like a genetic algorithm - Right 
?] - and I have to go to university :-) ).?
?
For now I'm using a MC-engine with UCT, trying out how much patterns and other 
things might make the results better.?
Now as I understand it, using patterns, groupstatus, ... during the MC-playout 
makes the results from the playout more meaningful (stronger - 30k instead of 
random), so if I would make the playout with a small engine (the easiest way to 
use the power of the gpu) the result should be more meaningful too. Has anyone 
done any test like that (like use gnugo level 0 instead of an MC-playout) ? 
Does anyone have a minimalistic non-MC go engine I could look at ? One more 
thing - has anyone tried using quasi-MC for go ??
?
Well - that's all folks.?
?
mfg?
?
Florian Erhardt?
___?
computer-go mailing list?
[EMAIL PROTECTED]
http://www.computer-go.org/mailman/listinfo/computer-go/?



More new features than ever.  Check out the new AIM(R) Mail ! - 
http://webmail.aim.com
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: computer-go Digest, Vol 43, Issue 8

[Don Dailey]

David Schneider-Joseph wrote:
> On Feb 11, 2008, at 8:42 PM, Don Dailey wrote:
>
>> David Schneider-Joseph wrote:
>>> On that topic - might it be possible that the notion of a "proper
>>> komi", derived as it is from "the hand of God" (perfect play), will
>>> invariably be too high for any actual go players which would be an
>>> interesting match for each other?
>> I guess it's possible.   I don't think it's likely but I guess nobody
>> can say with 100% certainty what the correct komi really is at any
>> non-trivial board size.
>
> Why not likely?  It seems a virtual guarantee to me.  By definition,
> komi is proportional to the value of moving first.  Likewise, by
> definition, your skill is the amount of value you get out of a move. 
> Therefore, better players should play with higher komi.
Hi David,

It's possible (even easy) to construct  positions where one side has a
win,  but the win requires careful accurate play or it loses.  Such
positions may actually be a practical advantage to the losing side if
two equal players do not understand how to play it. 

The opening position in GO is such a position.   I believe that if you
pick the "correct" komi, whatever that may be,  it's probably easier for
white to win.  

This would imply an adjustment downward from "god's number."   This is
essentially your argument and I agree with it.

But how much adjustment?This is where we disagree.   You seem to
believe that the adjustment should be quite large.   I disagree because
even though I believe the white pieces are easier to play,   I still
believe that a won position is still an advantage for reasonably
competent players.   A strange consequence of your position is that you
have to believe that a human player should prefer to start the game from
a dead lost position.   

For instance if 12.0 is God's komi and 9.5 is man's komi, then 9.5 gives
even chances in a position that is actually lost, and anything higher
gives white a practical advantage in a dead lost position!

Even though I believe as you do that it takes more skill to equalize
with white (given the correct komi),  I believe that 1/2 point more or
less gives one side a winning game,  and that is enough for players of
modest skill to have the better winning chances. 

It's pretty clear however that white is easier to play ...

If you play random vs random,  3.5 seems to be right komi.Since we
both agree that komi should be AT LEAST 7.5,  this implies that it's
easier to play the white pieces for a player of limited skill (of course
assuming komi is set correctly, whatever that may be.)  And sure enough,
if you use weak but not random program, the komi required jumps up very
quickly.  Even very weak programs seem to require about 7.5 komi,  if
they are beyond just weak beginner.  

But then even programs enormously stronger still require 7.5 komi.  

My feelings on this seem to match at least one source:

Look here:http://senseis.xmp.net/?Komi

Here is an excerpt:

It is widely believed that the correct komi is independent of board size
for all but the smallest boards. For area scoring, this would give 7 for
9x9+, 8 for 8x8, 7 for 7x7, 4 for 6x6, 25 for 5x5 (w cannot live), 0 for
4x4, 9 for 3x3, 4 for 2x2 with a superko rule, and 0 for 1x1. (these
need to be verified)

Despite all of this,  I allowed the possibility that it's possible that
even God cannot win at 7.5 komi. 


- Don



> ___
> computer-go mailing list
> computer-go@computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/
>
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: computer-go Digest, Vol 43, Issue 8

[Michael Alford]

Don Dailey wrote:



I tried Alford's value of 9.5 komi and white is even more happy, showing
about 0.547  in the score.


I don't believe what Alford says about 9.5 being the correct komi for
9x9.Where does that information come from?



Japanese tv games.
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: computer-go Digest, Vol 43, Issue 8

[Stuart A. Yeates]

On Feb 12, 2008 2:10 PM, Don Dailey <[EMAIL PROTECTED]> wrote:
>
>
> Andy wrote:
> > But the program isn't stronger than pros, so how can it give better
> > information about proper komi?
> Pro's cannot give you statistical information on komi unless you simply
> collate several thousand pro games.
>
> I don't think you need a particularly strong program,  just good
> programs.If you notice that over thousands of games 6.5 is gives
> black a statistically significant edge, and 8.5 gives white a
> statistically significant edge,  you know (at least for programs) that
> 8.5 is too high.

I think you might need a strong program with either (a) no built-in
knowledge about the game of go (i.e. pure UCT with no open book, no
heuristics, etc) or (b) with built-in knowledge which can be shown to
be of equal benefit to both black and white.

I'm guessing that (a) will happen before (b).

cheers
stuart
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: computer-go Digest, Vol 43, Issue 8

[David Schneider-Joseph]

On Feb 11, 2008, at 8:42 PM, Don Dailey wrote:


David Schneider-Joseph wrote:

On that topic - might it be possible that the notion of a "proper
komi", derived as it is from "the hand of God" (perfect play), will
invariably be too high for any actual go players which would be an
interesting match for each other?

I guess it's possible.   I don't think it's likely but I guess nobody
can say with 100% certainty what the correct komi really is at any
non-trivial board size.


Why not likely?  It seems a virtual guarantee to me.  By definition,  
komi is proportional to the value of moving first.  Likewise, by  
definition, your skill is the amount of value you get out of a move.   
Therefore, better players should play with higher komi.

___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: computer-go Digest, Vol 43, Issue 8

[Don Dailey]

David Schneider-Joseph wrote:
> On that topic - might it be possible that the notion of a "proper
> komi", derived as it is from "the hand of God" (perfect play), will
> invariably be too high for any actual go players which would be an
> interesting match for each other?
I guess it's possible.   I don't think it's likely but I guess nobody
can say with 100% certainty what the correct komi really is at any
non-trivial board size.  


- Don




>
> On Feb 11, 2008, at 7:35 PM, Andy wrote:
>
>> But the program isn't stronger than pros, so how can it give better
>> information about proper komi?
>>
>> On Feb 11, 2008 6:09 PM, Christoph Birk <[EMAIL PROTECTED]
>> > wrote:
>>
>> On Mon, 11 Feb 2008, Don Dailey wrote:
>> > I don't bet,  but if I did,  I would bet that it's 7 or 8, and I'm
>> > fairly certain that with best play the game would end with 7 extra
>> > points for black.
>> >
>> > I think this was discussed at great length 2 or 3 years ago.
>>
>> I know ... I brought it up again because of Mogo's success.
>> A very (!) strong program should be able to tell us the proper
>> komi.
>>
>> Christoph
>>
>> ___
>> computer-go mailing list
>> computer-go@computer-go.org 
>> http://www.computer-go.org/mailman/listinfo/computer-go/
>>
>>
>> ___
>> computer-go mailing list
>> computer-go@computer-go.org 
>> http://www.computer-go.org/mailman/listinfo/computer-go/
>
> 
>
> ___
> computer-go mailing list
> computer-go@computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: computer-go Digest, Vol 43, Issue 8

[Don Dailey]

Andy wrote:
> But the program isn't stronger than pros, so how can it give better
> information about proper komi?
Pro's cannot give you statistical information on komi unless you simply
collate several thousand pro games.

I don't think you need a particularly strong program,  just good
programs.If you notice that over thousands of games 6.5 is gives
black a statistically significant edge, and 8.5 gives white a
statistically significant edge,  you know (at least for programs) that 
8.5 is too high.  

Although it's possible that black has a won game at 8.5 komi,  the
evidence from computer play is just the opposite. You would have to
assume that a computer is a better fighter when down, or conversely gets
lazy when winning.Somehow that is difficult to believe.

Also,  you can try giving mogo a 6.5, and 8.5 komi and searching the
second position (it seems to always play e5 on the first move.)At
6.5 komi,  after  black e5  white thinks it is slightly losing.  At 8.5
komi white thinks it is slightly winning!  At 7.5 komi it also
thinks white is winning slightly.

I tried Alford's value of 9.5 komi and white is even more happy, showing
about 0.547  in the score.


I don't believe what Alford says about 9.5 being the correct komi for
9x9.Where does that information come from?

- Don


>
> On Feb 11, 2008 6:09 PM, Christoph Birk <[EMAIL PROTECTED]
> > wrote:
>
> On Mon, 11 Feb 2008, Don Dailey wrote:
> > I don't bet,  but if I did,  I would bet that it's 7 or 8, and I'm
> > fairly certain that with best play the game would end with 7 extra
> > points for black.
> >
> > I think this was discussed at great length 2 or 3 years ago.
>
> I know ... I brought it up again because of Mogo's success.
> A very (!) strong program should be able to tell us the proper
> komi.
>
> Christoph
>
> ___
> computer-go mailing list
> computer-go@computer-go.org 
> http://www.computer-go.org/mailman/listinfo/computer-go/
>
>
> 
>
> ___
> computer-go mailing list
> computer-go@computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: computer-go Digest, Vol 43, Issue 8

[David Schneider-Joseph]

On that topic - might it be possible that the notion of a "proper  
komi", derived as it is from "the hand of God" (perfect play), will  
invariably be too high for any actual go players which would be an  
interesting match for each other?


On Feb 11, 2008, at 7:35 PM, Andy wrote:

But the program isn't stronger than pros, so how can it give better  
information about proper komi?


On Feb 11, 2008 6:09 PM, Christoph Birk <[EMAIL PROTECTED]> wrote:
On Mon, 11 Feb 2008, Don Dailey wrote:
> I don't bet,  but if I did,  I would bet that it's 7 or 8, and I'm
> fairly certain that with best play the game would end with 7 extra
> points for black.
>
> I think this was discussed at great length 2 or 3 years ago.

I know ... I brought it up again because of Mogo's success.
A very (!) strong program should be able to tell us the proper
komi.

Christoph

___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/


___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: computer-go Digest, Vol 43, Issue 8

[Andy]

But the program isn't stronger than pros, so how can it give better
information about proper komi?

On Feb 11, 2008 6:09 PM, Christoph Birk <[EMAIL PROTECTED]> wrote:

> On Mon, 11 Feb 2008, Don Dailey wrote:
> > I don't bet,  but if I did,  I would bet that it's 7 or 8, and I'm
> > fairly certain that with best play the game would end with 7 extra
> > points for black.
> >
> > I think this was discussed at great length 2 or 3 years ago.
>
> I know ... I brought it up again because of Mogo's success.
> A very (!) strong program should be able to tell us the proper
> komi.
>
> Christoph
>
> ___
> computer-go mailing list
> computer-go@computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/
>
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: computer-go Digest, Vol 43, Issue 8

[Christoph Birk]

On Mon, 11 Feb 2008, Michael Alford wrote:

i believe correct komi for 9x9 with pros is 9.5


That's way too large.

Christoph

___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: computer-go Digest, Vol 43, Issue 8

[Christoph Birk]

On Mon, 11 Feb 2008, Don Dailey wrote:

I don't bet,  but if I did,  I would bet that it's 7 or 8, and I'm
fairly certain that with best play the game would end with 7 extra
points for black.

I think this was discussed at great length 2 or 3 years ago.


I know ... I brought it up again because of Mogo's success.
A very (!) strong program should be able to tell us the proper
komi.

Christoph

___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: computer-go Digest, Vol 43, Issue 8

[Don Dailey]

Christoph Birk wrote:
> On Mon, 11 Feb 2008, Don Dailey wrote:
>> Is your question whether 7.0 or 8.0 is the best komi?   Or do you
>> suspect a different 1/2 komi value is best?
>
> I wonder what the true komi is ... I don't know (nobody knows?) if
> it's fractional or not; eg. for 7x7 it is 9.0.
I think the "true" komi must be an even number.   For CGOS, the question
is what is the "best" komi since drawn games are simply not allowed.  

I don't bet,  but if I did,  I would bet that it's 7 or 8, and I'm
fairly certain that with best play the game would end with 7 extra
points for black. 

I think this was discussed at great length 2 or 3 years ago.

- Don


>
> Christoph
>
> ___
> computer-go mailing list
> computer-go@computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/
>
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: computer-go Digest, Vol 43, Issue 8

[Michael Alford]

Christoph Birk wrote:

On Mon, 11 Feb 2008, Don Dailey wrote:

Is your question whether 7.0 or 8.0 is the best komi?   Or do you
suspect a different 1/2 komi value is best?


I wonder what the true komi is ... I don't know (nobody knows?) if
it's fractional or not; eg. for 7x7 it is 9.0.

Christoph

___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/



unlurk

i believe correct komi for 9x9 with pros is 9.5

lurk
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: computer-go Digest, Vol 43, Issue 8

[Christoph Birk]

On Mon, 11 Feb 2008, Don Dailey wrote:

Is your question whether 7.0 or 8.0 is the best komi?   Or do you
suspect a different 1/2 komi value is best?


I wonder what the true komi is ... I don't know (nobody knows?) if
it's fractional or not; eg. for 7x7 it is 9.0.

Christoph

___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: computer-go Digest, Vol 43, Issue 8

[Hideki Kato]

Olivier Teytaud: <[EMAIL PROTECTED]>:
>> That translates to mean that MoGo no longer uses upper confidence
>> bounds, and only uses means.  It also means that MoGo will _never_
>> explore improbable children (after a few sims) unless the RAVE value
>> yields an unusually high estimate for it.  Is all of that correct?
>>
>
>Precisely: I don't see why you would be wrong, but empirically for 9x9,
>we have played games against high-level humans and for the (few :-) )
>games that mogo lost, we tried to see which moves were erroneously chosen
>by mogo; if we restart mogo at the same position with a huge 
>computation time (30 minutes of a fast octocore) mogo always changed his
>mind and moves to a better move.

Could we look at some of the records of the games?

-Hideki

>So:
>- theoretically, I don't see any reason for mogo to be asymptotically
>   consistent
>- there are long computation times during which mogo focuses on a bad
>   move
>- however, we have not seen a case of bad move for which mogo keeps
>   this move in case of _very_ long computation times
>
>==> if someone beats the release MoGoR3 with
>   very large computation times (time x nbcores = 4h, 1 to 4 cores)
>   I'm interested in the sgf file and the analysis
>___
>computer-go mailing list
>computer-go@computer-go.org
>http://www.computer-go.org/mailman/listinfo/computer-go/
--
[EMAIL PROTECTED] (Kato)
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: computer-go Digest, Vol 43, Issue 8

[Don Dailey]

Christoph Birk wrote:
> On Mon, 11 Feb 2008, Olivier Teytaud wrote:
>> With 20 minute games, some people succeed in winning games
>> against the release 3 of MoGo. But for
>> X-hours-per-move, I don't know.
>
> What are the self-play results (white vs. black) for "hour-long"
> games of Mogo?
> I am wondering if the proper komi for 9x9 is really 7.5.

Is your question whether 7.0 or 8.0 is the best komi?   Or do you
suspect a different 1/2 komi value is best?

In thousands of tests I have run with strong programs,  7.5 appears to
the best fractional komi.   I believe 7.5 is a win for white with
perfect play,  but I'm not sure - however I am pretty sure it's the best
fractional komi to use because 6.5 is clearly in blacks favor and 8.5 is
clearly in whites favor.


- Don
 



>
> Thanks,
> Christoph
> ___
> computer-go mailing list
> computer-go@computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/
>
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: computer-go Digest, Vol 43, Issue 8

[Christoph Birk]

On Mon, 11 Feb 2008, Olivier Teytaud wrote:

With 20 minute games, some people succeed in winning games
against the release 3 of MoGo. But for
X-hours-per-move, I don't know.


What are the self-play results (white vs. black) for "hour-long"
games of Mogo?
I am wondering if the proper komi for 9x9 is really 7.5.

Thanks,
Christoph
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] gpugo

[Olivier Teytaud]

easiest way to use the power of the gpu) the result should be more meaningful 
too. Has anyone done any test like that (like use gnugo level 0 instead of an 
MC-playout) ? Does anyone have a minimalistic non-MC go engine I could look 
at ? One more thing - has anyone tried using quasi-MC for go ?


I have only disappointing answers :-)

Using a good player instead of MC has been used with negative results -
 the quality of the MC as a player is not the best criterion for
 evaluating a MC as a building block in bandit-based Monte-Carlo planning.
 I have not implemented anything myself, but you'll find informations in
 past posts in the mailing lists.

Quasi-MC is difficult to do - some things have been tried here, without
 positive results (and often negative ones :-) ). More precisely, pairing
 simulations has been tried without success; also, the variance of MC is
 seemingly important as reducing it by averaging is not efficient (at
 least, I believe that :-) ).

Olivier
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

[computer-go] gpugo

[Florian Erhardt]

Hello !

My name is Florian Erhardt, I am a bachelor student of computer sciences 
and am in the process of optimizing libEGO for gpgpu. For now I 
implemented the SFMT (even I can do copy and paste) on the gpu and am 
now atomizing the MC to be done by the gpu. If everything works as I 
planned it, I'll  have a releasable program in a few months (I still 
have to learn lots about parallelizing, gpu programming and programming 
in general [maybe leaning about genetic algorithms can't hurt either - 
The RAVE algorithm is more like a genetic algorithm - Right ?] - and I 
have to go to university :-) ).


For now I'm using a  MC-engine with UCT, trying out how much patterns 
and other things might make the results better.
Now as I understand it, using patterns, groupstatus, ... during the 
MC-playout makes the results from the playout more meaningful (stronger 
- 30k instead of random), so if I would make the playout with a small 
engine (the easiest way to use the power of the gpu) the result should 
be more meaningful too. Has anyone done any test like that (like use 
gnugo level 0 instead of an MC-playout) ? Does anyone have a 
minimalistic non-MC go engine I could look at ? One more thing - has 
anyone tried using quasi-MC for go ?


Well - that's all folks.

mfg

Florian Erhardt
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: computer-go Digest, Vol 43, Issue 8

[Sylvain Gelly]

> >Thinking a little more about it, I think we have to add an hypothesis
> >which is that, for a given move, the number of AMAF updates if < alpha
> >(nb total UCT updates), with alpha < 1. That seems to hold for most of
> >the updates (with alpha close to 0.5), but there may be cases where it
> >does not hold.
> >
> >
> If I understand well, you say that, in order to ensure consistency,
> we need some assumptions on the AMAF updates,
> i.e. the MC simulations which decide which move will have AMAF updates.

Yes.
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: computer-go Digest, Vol 43, Issue 8

[Olivier Teytaud]

I can't tell if you mean the float version or the double version.  
Using the float version (since it was all I had), I did a fairly 
extensive analysis of the losing move from the MoGo game that Fotland 
added comments to.  My results were posted to this list on 2/1/08 
under the subject, "UCT and solving life and death".  The test was 
run on 4 cores.



Oops, I meant 2 threads (as stated in the original email).


Thanks for your posts,
the position is interesting.

I'll try to see what happens with larger computation times. If it is a 
case of non-consistency it's
interesting :-)I have no go-expertise to guess that this is the bad 
move from mogo, but

I trust you for that :-)

Well, if I find minutes for that after my fight with a furious myrinet 
switch :-)

___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: computer-go Digest, Vol 43, Issue 8

[Olivier Teytaud]

>>> Sylvain wrote:


Thinking a little more about it, I think we have to add an hypothesis
which is that, for a given move, the number of AMAF updates if < alpha
(nb total UCT updates), with alpha < 1. That seems to hold for most of
the updates (with alpha close to 0.5), but there may be cases where it
does not hold.
 


If I understand well, you say that, in order to ensure consistency,
we need some assumptions on the AMAF updates,
i.e. the MC simulations which decide which move will have AMAF updates.

(this would be a good piece of news for the two people trying to get rid
of some bias in the MC :-)  unfortunately, it is difficult to have 
statistics on the
level of mogo depending on that. It it gives some result, it is probably 
only for

huge computation times and very specific positions... we need 100 000 years
before ensuring that with 5% confidence intervals on complete games :-) )

(in empirical cases, I'll try to check the consistency on the example
posted by David & Michael)
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: computer-go Digest, Vol 43, Issue 8

[Sylvain Gelly]

> As far as I see,
> if RAVE gives constant value 0 to one move, it will never be tested if
> other moves
> have non-zero AMAF values.
>
> A move
> with "real" empirical probability 0 of winning and AMAF value of 0.01
> will always be preferred to a non-simulated move with AMAF 0.0, whatever
> may be
> the number of simulations.
I agree, it is why I added a statement about the prior, which implies
that the AMAF value is never 0.0 but at worst decreases like 1/m if m
is the number of AMAF updates for that move.

Thinking a little more about it, I think we have to add an hypothesis
which is that, for a given move, the number of AMAF updates if < alpha
(nb total UCT updates), with alpha < 1. That seems to hold for most of
the updates (with alpha close to 0.5), but there may be cases where it
does not hold.
Maybe I am confused and say unsound things, sorry for that. It is the
kind of things we should discuss in front of a black (or white) board.

Sylvain
___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/

Re: [computer-go] Re: computer-go Digest, Vol 43, Issue 8

[Olivier Teytaud]

A new position is always visited unless the leaf of the tree is the
end of the game. In that case, one player always win, so the other
always win. Then, the losing player will explore all the other moves
to avoid the sure loss. If all moves are still loosing, that will
propagate to the move before, and the exploration will begin and so
on.
 


("min" --> "loss" I guess)

As far as I see,
if RAVE gives constant value 0 to one move, it will never be tested if 
other moves

have non-zero AMAF values.

A move
with "real" empirical probability 0 of winning and AMAF value of 0.01
will always be preferred to a non-simulated move with AMAF 0.0, whatever 
may be

the number of simulations.

So, I don't see why the bandit would be consistent, unless we have
assumptions on the MC or on RAVE values.

I might be completly wrong, as I said
I have only retro-engineered the bandit in mogo until the recent PDF file. I
trust your opinion more than mine :-)

There are people studying some specific positions with surprising behavior,
but I am not working on that with them, they might want to post their
analysis in this mailing-list...




___
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/