On Sun, Oct 26, 2008 at 21:22, Don Dailey [EMAIL PROTECTED] wrote:
I'm doing a small study of the scalability of the reference bot at
various numbers of playouts.
I still need a lot more games, but in general you eventually start to
see a point of diminishing returns for each doubling. I
OK, after dicking about for a few hours with git and Mercurial I
decided against using any of them. I keep getting errors or
completely fail to understand how it works. It's just not intuitive
enough to get going quickly.
Moreover, if my goal is to get newcomers up and running quickly, I
A while ago I implemented what I thought was a fairly straightforward
way to deal with transpositions. But to my surprise it made the
program weaker instead of stronger. Since I couldn't figure out
immediately what was wrong with it, I decided to leave it alone for
the time being.
Just
- Original Message
From: Mark Boon [EMAIL PROTECTED]
snippage
Let me first describe what I did (ar attempted to do): all nodes are
stored in a hash-table using a checksum. Whenever I create a new node
in the tree I add it in the hash-table as well. If two nodes have the
same
On 27-okt-08, at 11:51, terry mcintyre wrote:
- Original Message
From: Mark Boon [EMAIL PROTECTED]
snippage
Let me first describe what I did (ar attempted to do): all nodes are
stored in a hash-table using a checksum. Whenever I create a new node
in the tree I add it in the
When a child has been sampled often through some other path a naive
implementation may initially explore other less frequently visited
children first. The new path leading to the transposition may
therefore suffer from some initial bias. Using state-action values
appears to solve the problem.
[EMAIL PROTECTED] wrote on 27-10-2008 14:57:54:
On 27-okt-08, at 11:51, terry mcintyre wrote:
- Original Message
From: Mark Boon [EMAIL PROTECTED]
snippage
Let me first describe what I did (ar attempted to do): all nodes are
stored in a hash-table using a checksum.
On 27-okt-08, at 12:45, Erik van der Werf wrote:
Using state-action values
appears to solve the problem.
What are 'state-action values'?
Mark
___
computer-go mailing list
computer-go@computer-go.org
Reinforcement Learning terminology :-)
In Go the state is the board situation (stones, player to move, ko
info, etc.), the action is simply the move. Together they form
state-action pairs.
A standard transposition table typically only has state values; action
values can then be inferred from a
The following modification to AMAF seems to perform better and scale better. The idea is to weight the moves at the beginning of the playout heavier than the
moves at the end of the playout. It's probably not a new idea.
This code from the reference implementation:
wins[mv] += sc;
Interesting. I had tried something much more simple. I added two wins
for the first move in the sequence, figuring that a move being first
in the sequence should have more weight than the rest. But to my
surprise that played much worse, winning only 37%. Maybe I made a
mistake and I should
Great information. I'll include a version of this to my scalability
study. Is this the C version?
wins[] and hits[] are integer arrays and weight is a fraction less than
1.0, so I'm not sure how this works. Did you change hits and wins to be
doubles?
There are many enhancements possible,
Yes, they become doubles.
Don Dailey wrote:
Great information. I'll include a version of this to my scalability
study. Is this the C version?
wins[] and hits[] are integer arrays and weight is a fraction less than
1.0, so I'm not sure how this works. Did you change hits and wins to be
On Sun, Oct 26, 2008 at 4:22 PM, Don Dailey [EMAIL PROTECTED] wrote:
I imagine that it approaches some hypothetical
level in an asymptotic way.
For most board positions, it is reasonable to expect that there exists
a single move for which the asymptotic Monte Carlo value is higher
than the
It's actually more bare-bones (in the sense that there is less code) if you
consider the update loop that I showed.
Don Dailey wrote:
Great information. I'll include a version of this to my scalability
study. Is this the C version?
wins[] and hits[] are integer arrays and weight is a
Now that Leela and Many Faces v12 are available for any Windows user
to purchase and run (and Fuego is free to tinker with), has anyone
tried them against the old guard of commercial programs? KCC Igo,
Haruka, Go++, and HandTalk haven't competed in a while so it is hard
to tell how much
A post from Michael Williams led me to review this mail below once
more. I hadn't looked at the code of Don's reference bot very closely
until now and instead relied on the description he gave below:
On 23-okt-08, at 14:29, Don Dailey wrote:
Let me give you a simple example where we set
Ian Osgood wrote:
(For that matter,
it isn't a foregone conclusion that they are better; GNU Go won the 2008
US computer go tournament against a field MC programs.)
Believe me, in match long enough to exclude pure luck, with the MC
programs running on something faster than a Pentium 4, it IS
Maybe Don built it that way so that the playouts could handle integer komi and
the possibility of a draw. In that case, it would neither add one nor subtract
one.
Mark Boon wrote:
A post from Michael Williams led me to review this mail below once more.
I hadn't looked at the code of Don's
From: Ian Osgood [EMAIL PROTECTED]
Now that Leela and Many Faces v12 are available for any Windows user
to purchase
Thanks for the heads-up, I must have missed the announcement.
Do either of these worthy programs work with Wine on Linux?
I recently tried a development version of MGF12
terry mcintyre wrote:
From: Ian Osgood [EMAIL PROTECTED]
Now that Leela and Many Faces v12 are available for any Windows
user to purchase
Thanks for the heads-up, I must have missed the announcement.
Do either of these worthy programs work with Wine on Linux?
You can try the free
On Mon, 2008-10-27 at 17:19 -0200, Mark Boon wrote:
So I understand from the above that when a playout leads to a win
you
add 1 to the wins.
But in the code you subtract one when it leads to a loss.
This is just semantics. In the literal code a win is 1 and a loss is -1
but when I
By your argument, it would seem to make sense to remove this check even if you
don't use my decaying weight.
boolean ok = true;// ok to use this move?
// see if either side has used this move before
for (int j=savctm; ji; j++) {
GNU Go won the tournament at the US Go Congress against several MC
programs including Many Faces and Leela, but the Many Faces that
competed was not quite the newest. David Fotland was working on the
program while in Portland and only got the multi-core (to use both
cores of a duo)
On Mon, 2008-10-27 at 16:08 -0400, Michael Williams wrote:
By your argument, it would seem to make sense to remove this check
even if you don't use my decaying weight.
boolean ok = true;// ok to use this move?
// see if either side has used this move
I think the question is largely meaningless, because few games have
been studied by humans (or human computer programmers) with the depth
and intensity that has been achieved for games like chess and go.
In general, games with many choices and no obvious strategies
are good for people and bad
On Tue, 2008-10-28 at 08:55 +0900, Darren Cook wrote:
Where harder means the gap between top programs and top human players
is bigger, are there any games harder than go? Including games of
imperfect information, multi-player games, single-player puzzle games.
Naturally I'm most interested
Where harder means the gap between top programs and top human players
is bigger, are there any games harder than go? Including games of
imperfect information, multi-player games, single-player puzzle games.
Poetry contests?
I caught the smiley, but if you can define the rules (such that a
At Portland, I ran 3 games against gnugo with the two-CPU version, and won
all three. Version 12 as released is quite a bit stronger than the code I
was using in Portland.
Many Faces version 11 was competitive against the old guard, winning about
30% to 40% against handtalk for example. Version
*
CALL FOR PARTICIPATION
The 2nd Computer Go UEC Cup
The University of Electro-Communication,
Tokyo, Japan, 13-14 December 2008
http://jsb.cs.uec.ac.jp/~igo/2008/eng/index.html
*
Computer Scrabble significantly exceeds humans. A basic monte carlo search
and an endgame solver is very effective. There is probably still much
strength to be gained (very little opponent modeling is done), but it's
already so strong I don't think it's getting much attention.
Looks like
Core Wars , robot soccer (there is a simulation league), pictionary,...
- Dave Hillis
?
- Original Message - From: Darren Cook [EMAIL PROTECTED]?
To: computer-go@computer-go.org?
Sent: Monday, October 27, 2008 8:54 PM?
Subject: Re: [computer-go] OT: Harder than go??
?
Where harder
Will remote computing be allowed, or do we need to have our hardware
on site?
Cheers,
David
On 27, Oct 2008, at 7:21 PM, TAKESHI ITO wrote:
*
CALL FOR PARTICIPATION
The 2nd Computer Go UEC Cup
The University of
Do we have to show up in person, or can our programs be operated for us?
David Fotland
-Original Message-
From: [EMAIL PROTECTED] [mailto:computer-go-
[EMAIL PROTECTED] On Behalf Of TAKESHI ITO
Sent: Monday, October 27, 2008 7:22 PM
To: computer-go@computer-go.org
Subject:
I love pictionary! The computers will be drunk, right?
[EMAIL PROTECTED] wrote:
Core Wars , robot soccer (there is a simulation league), pictionary,...
- Dave Hillis
- Original Message - From: Darren Cook [EMAIL PROTECTED]
mailto:[EMAIL PROTECTED]
To: computer-go@computer-go.org
You can also try the free version of Many Faces of Go 12 at
www.smart-games.com. Many Faces version 11 worked under wine, so this one
should too. If you try it please let me know if it works.
David
-Original Message-
From: [EMAIL PROTECTED] [mailto:computer-go-
[EMAIL PROTECTED] On
36 matches
Mail list logo