[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[Bug-gnubg] Re: Battle of the Bots

From: Joseph Heled
Subject: [Bug-gnubg] Re: Battle of the Bots
Date: Sun, 05 Jan 2003 19:42:24 +1300
User-agent: Mozilla/5.0 (X11; U; Linux i686; en-US; rv:1.2) Gecko/20021202

Attached are 48 games played with gnu ply2. Please comment on the mistakes which make it win 10% less games than SN 4. There must be several here ...


Neil Kaz wrote:
My 3ply/precise Snowie rollout has this position as having a cubeless equity
of .554 but my new GNU 2-ply rollout has the equity as about .785 and was
winning the game 10% less from the backgame player's side.

I'll run a full 1296 games 2ply overnight since this takes some time ( I did
385 games in a couple of hours)

This is a very straightforward position with all the skill on the backgame
player's side and a good test of how well a bot can play a backgame.

To do this rollout cubeless with GNU I need to do it by clicking on analyse
and then rollout (off course after setting the rollout and turning off
Jacoby rule)


----- Original Message -----
From: "Joseph Heled" <address@hidden>
To: "Neil Kaz" <address@hidden>
Cc: "Albert Silver" <address@hidden>; "'Joern Thyssen'"
<address@hidden>; "'Øystein Johansen'" <address@hidden>
Sent: Saturday, January 04, 2003 2:28 PM
Subject: Re: Battle of the Bots

Neil Kaz wrote:

By all means..I would use the Jacobs MET until someone, like myself can


out a better one. I trust that you guys will teach me how to make my own


as I obtain data.

I want to use the new GNU version which is supposed to be superior in
backgames, but from rollouts I am doing of a very basic backgame where all
the skill is needed on the backgame player's side , it appears to be FAR
behind Snowie 4 or, perhaps I don't have the proper weights installed.

Can you post the position and Snowie results please? I would not get too
  concerned about one specific position, or even several. Also, play
will be at 2 or 3ply, while rollouts usually are 0ply, and the error
rate of 2ply (in crashed positions) is about half as far as I can tell.


I am very confused with this "weight" stuff and wish there was a simple


just to DL everything I need to run GNU at its highest strength.

----- Original Message -----
From: "Albert Silver" <address@hidden>
To: "'Joern Thyssen'" <address@hidden>; "'Joseph Heled'"


"'Neil Kaz'" <address@hidden>
Cc: "'Øystein Johansen'" <address@hidden>
Sent: Saturday, January 04, 2003 12:26 PM
Subject: Battle of the Bots


As you may have seen (I'm sure Neil has), GammonVillage has
begun a series called the Battle of the Bots, in which Snowie, Jellyfish
and GNUBG will play matches. They realize no statistical info can be
obtained by this on which bot is stronger, and say so clearly, however,
hope to generate interesting discussions on key positions. The impartial
(cough!) judge of the errors will be Snowie 4's 3-ply analysis. I asked
about the settings, MET, version, and DB that would be used. Obviously
this hadn't been considered very deeply so was what I suggested. I
should add that they don't know it is me. My suggestion is the Jacobs
MET, and a 13 point DB to be built. What about the playing settings? I'm
running on an abysmally slo machine, but if they have decent hardware,
then I'd like to make a bold suggestion: some setup of move filters,
allowing at least 2-3 moves to be analyzed at GNU's 3-ply.
Here is what I've noticed in my experience with analyzing with
GNU: it analyzes at least 99% of the time the top 2-3 moves in its list
at 2-ply. It does change it's choice of best move with 3-ply a few
times. I can't estimate a percentage. The 3-ply evaluation is almost
always the same as its rollouts. Even the errors pointed out by Snowie 4
(and GNU points out errors by Snowie 4 too BTW) are corrected invariably
by its 3-ply. So: can anyone suggest a top-level 3-ply play (maybe only
analyzing 2-3 moves at 3-ply) that would be about the same speed as
Snowie 4's 3-ply which is very slow. They plan to use Snowie 4 in the
contest later, so that is the limit of the acceptable.
I can't even test this properly on my laptop, to which I'm
confined until I get a new motherboard, otherwise I would. Also could
you give me the exact script to develop the 13-point DB as well as the
expected size?


-----Original Message-----
From: address@hidden
[mailto:address@hidden On Behalf Of
Joern Thyssen
Sent: Saturday, January 04, 2003 2:21 PM
To: Neil Kaz
Cc: Øystein Johansen; GnuBg Bug (E-mail)
Subject: Re: [Bug-gnubg] Re: Long GNU rollout results

On Sat, Jan 04, 2003 at 10:40:36AM -0600, Neil Kaz wrote

Thx everyone..I want to perfectly understand what GNU is doing in


and how/why it is presenting the data like this. I hope to be 100%


that GNU is doing everything like it is supposed. ..ie perfectly


the plays and cubes according to match score.

gnubg always evaluate plays and cube according the current score and
cube in rollouts.


Bug-gnubg mailing list

Attachment: neil31.sgf
Description: x-sgf

reply via email to

[Prev in Thread] Current Thread [Next in Thread]