Hi all,
It has been a long time since my last post here!
I have been away from computer go since about one year, but the publishing
process is really slow. This paper describes research I was doing in 2012
and, since the reviewers were not quite convinced, I had to do more
experiments in
Hi Aja,
The testing program codes different problems in the same sgf file
like in:
loadsgf sgf/seki/case1.sgf 4
14 genmove w
#? [B2|J3]
loadsgf sgf/seki/case1.sgf 6
16 genmove w
#? [B2]
If you ignore the move numbers, j3 is not even a legal move. Unfortunately,
move
Rémi Coulom wrote:
Accelerated UCT does this:
https://www.conftool.net/acg13/index.php/Hashimoto-Accelerated_UCT_and_Its_A
pplication_to_Two-Player_Games-111.pdf?page=downloadPaper
https://www.conftool.net/acg13/index.php/Hashimoto-Accelerated_UCT_and_Its_
Hi,
To set the number of playouts, you need:
uct_param_player ignore_clock 1
uct_param_player max_games *number*
for a typical testing environment, you probably also need:
uct_param_player ponder 0
uct_param_player reuse_subtree 0
also
uct_param_search max_nodes *number*
Maybe a tournament is not the best way to see quality computer/human games.
There are better ways to measure the computer/computer performance and the
human/human performance is not interesting here. We could simply schedule
computer/human games on KGS (e.g., 3 times a year, one in each time zone
Darren Cook wrote:
Hence the proposal to use alpha-beta as the top-level search,
using MCTS with about 50K playouts at the nodes.
This is not anyhow close to what I am researching so take my
word on this as just a though. But alpha-beta is a method to
find outliers. When a program is happy
Ingo Althöfer wrote:
... don't let that fool you, Zen with over a minute per
move will be a hell of an opponent.
From where do you have this knowledge?
Or is it just your opinion?
It is my opinion based on my own experience. I can't
tell about Zen. But I know how hard it is to code
a program
Nick, thanks again for organizing and the new table.
A conclusion that can be draw from the table is that,
except for the 2nd and 3rd place of Pachi and MF which
was very disputed and anyone could have won, the
different ranks give very consistent results. Almost all
games against lower ranked
Don Dailey wrote:
I don't understand your concept of wrong direction.When
you expand the tree for ANY move you get more information about
how good or how bad that move is, so it's always helpful (on
average) to have more time.
I think Hideki's argument is about: more simulations won't
Don Dailey wrote:
Are you trying to say that heavy playouts are better?
Who is going to argue with that?I agree completely.
Are you trying to make the point that there are very simple
to understand positions that computers cannot easily solve?
I agree with that. Are you trying to say
Hi.
I know this subject comes to the list from time to time,
but since hardware and prices change, it is interesting
to read what people say.
I purchased an i7 920 two years ago. I was very happy
with its overclocked performance (with a Gigabyte
GA-EX58-UD5 mobo) but it is starting to have
Rémi Coulom wrote:
Also, I think we did not yet advertise that paper Lukasz wrote
with me last year:
http://www.mimuw.edu.pl/~lew/files/Simulation-based%20Search%20of%20Combinatorial%20Games.pdf
This one doesn't work for me. I get :
Access forbidden!
Error 403
Jacques.
Don Dailey wrote:
Sometimes nobody is using it, and if there is only 1 player
it cannot play games of course. You can test that by putting
up a second client to see what happens.
I'll check it out.
At the moment it is not working. When I test, I run a 19x19
test pack downloaded from:
Don Dailey wrote:
Ok, I gave it a kick start. Should be coming up shortly.
Its working fine now. Thanks.
Jacques.
___
Computer-go mailing list
Computer-go@dvandva.org
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
David Doshay wrote:
I can invoke cgosview-darwin from the command line in a terminal
window only if I do not include the:
-server cgos.boardspace.net -port 6819
Hi,
This happens in all OSs, not just Mac-OS. The correct syntax is:
cgosview cgos.boardspace.net 6819
_Without_ -server or
Pachi won a 7h game against Zhou Junxun 9p, who is an active 9p
player and has won at least one world title. I would say this put
Pachi at top amature or beginning pro level.
Simplifying: All pros are 9d. The handicap system is not fine grained
enough to measure differences at pro level. Any
Hiroshi Yamashita wrote:
wow!
cross-table is great!
Thanks.
I agree. Another advantage of that table is that it gives
insight on how well the pairing algorithm is doing. The
ideal table should have empty spaces in the northeast
and southwest corners with the games along the main
diagonal. Of
Hi,
I have compiled different collections from
over 100K sgf files filtering all repeated,
short handicapped, wrong size, wrong player level,
blitz time settings, etc.
The last one I did which is use in many of my learned
patterns and in my paper:
http://www.dybot.com/papers/meval.htm
Has
Ingo Althöfer wrote:
How difficult would it be to design a program
that generates such bot-difficult semeais
(or at least building blocks for such)
automatically?
Computer generated problems are not new. Thomas
Wolf, the author of Go Tools generated a set
of tsumego problems automatically
There was another nice example in the last Tournament.
http://files.gokgs.com/games/2011/1/9/PueGo-pachi2.sgf
Fuego was leading easily (70% according to its own WR)
until move 189.
Imo, move 190 White .. T7 is a big mistake allowing
B to play 191 Black M13. After this move the semeai
is
The discussion changed the subject to joseki, but the
paper is not about joseki at all. (There is a poster
about joseki in the same website.)
The Power of Forgetting is an improvement to the
previous Last-Good Reply idea.
The results are spectacular and implementation is
super-simple. Looks
I typed:
Towards perfect play of Scrabble Sheppard
(with quotes as above) in Google scholar
Followed the links from the first URL until
http://arno.unimaas.nl/show.cgi?did=10724
then file 2, which is:
http://arno.unimaas.nl/show.cgi?fid=7134
and was able to download a .pdf
To my surprise
Hi Francois,
You mean all the 3x3 patterns? I'm only using 3x3
patterns that occur a number of times in my training
collection.
If 3x3 is the only pattern size you use, that probably does
not apply. But if your pattern system supports other sizes
then you must have implemented some kind of
Hi Francois, Welcome
For reference I need about 100k playouts with
RAVE to get 50% winrate against GnuGo 3.8 L10.
Yes that's more or less expected. At least before the
big improvements (yet to come ;-)
In my case I do a lot of testing at 4x1 because
the games are around 15 seconds long
I don't agree the first two games were that easy.
In the second game the bot was ahead most of the game
and failed in life and death in the top right corner.
It needed more CPU power probably.
It is a pity that you couldn't rent a better computer. The
amazon cloud was 26 ECU (1 ECU = 1 core
Olivier Teytaud wrote:
whereas I am looking for
(2) input = a database D
output = frequently matched patterns
This is what I do:
I have thousands of separate SGF files and a program to merge
them all in a single binary file. This program filters: handicap
games, insufficient rank in
In recent years at this time of the year it was known where
and when the Olympiad and the ACG congress will be. The
call for papers usually end at the end of Feb or beginning
of March.
Does anyone (Rémi? Hideki?) know where it will be and when?
Jacques.
Petr Baudis wrote:
I don't quite understand. I think most programs do not
disallow eye-filling in the tree stage, only in the MC
simulation stage. So your simulations will get misevaluated,
but given enough time to reach final position, tree will
always consider evaluating the eye-filling
Hideki Kato wrote:
| . . . . . . .
| . # # . # . .
| . O # . # . .
| O O O . # # .
| # # O O O # .
| . # # . . . .
---
Here, the problem is filling own eye at 1-1 for B, right?
Yes, that is the figure I meant, but it is not related with
nakade, only with filling your own
Don Dailey wrote:
To the best I can estimate it is something like 300 ELO
now which means in a short match there is almost no chance
for the human.
It is more than that, around 500. Only a match with a handheld
device could be seen as similar strength:
Top humans are near 2800
About the convergence proof to go:
As far as I can understand Álvaro's proof sounds correct,
but that possibly applies to most games, but not go as
humans play it. Because what we play is MCTS go an almost
identical game where filling your own eyes is not allowed.
All MCTS programs finish their
Intriguing! Do you think it actually improves its strength,
or is that just an experiment?
How does it know where the losing variation starts? Is
it updating some pattern weights, or updating some persistent
game tree?
I guess it is just an opening book that keeps win/loss info.
When a
If you censor at the root, you only have 361 losses that
you can endure before you are out of ideas. I guess you
implicitly meant censoring at move 3 or something.
Of course not. You start with a standard book. I my case,
learned from 50K games with about 25K nodes. You are just
adding
You said: When a branch has losses and 0 wins, you just censor that
move. Apparently you meant leaf and not branch.
Well I did not precise how you grow the tree. Obviously, the initial
offline learned tree has all nodes with more than one visit as the
number of visits is the criterion to build
Thank you Nick!
I like the idea very much and look forward to participate in 2011.
My suggestions are:
1. Use separate board sizes and also combined. So you could award
three different winners: a. 19x19, b. smaller boards and c. combined.
Combined would be the sum of all points awarded on all
On anti-MCTS bot strategy:
I don’t know of a strategy, but there sure are principles.
I can state one as a proverb:
We you clarify, you are helping the bot.
E.g., If a connection works but is not obvious, if a semeai
can be won but is not obvious, etc. the bot has to discover
it for each
Oops. Should be:
When you clarify, ..
Jacques.
___
Computer-go mailing list
Computer-go@dvandva.org
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
About the previous question, in which phase should more time
be given:
Don Dailey wrote:
In 9x9 go the first few moves are extremely critical and if you
start badly you cannot recover.
Petr Baudis wrote:
This depends a lot, e.g. on 9x9 it is probably worth spending a lot
of time in the
38 matches
Mail list logo