GNU Backgammon

Forum Archive : GNU Backgammon

Rollout settings

From:   Ian Dunstan
Date:   6 August 2003
Subject:   Gnubg settings revisted
Forum:   GammOnLine

> Should we open up an new dialogue where GNU users voice their
> favorite rollout settings (preferably with justifications) or is the
> space so vast that in the end we'll have a favourite setting for
> each poster, and be no further along?

I think Gnubg users will all develop their own favourite settings from
personal experience, I know I have. That said, I've certainly been
influenced by the recent findings of Albert and Zorba, in particular.
Overall, I get the feeling that Gnubg users are getting closer to some
pretty robust, but time efficient rollout settings. For what it's
worth, here is what I consider to be pretty good settings at present.
I'll ignore the recent debate we've had about cubeless vs cubeful
rollouts and state everything in cubeful terms.

1) Comparing chequer plays that are close in intent and execution.

By this I mean plays that are not trying to do very different things,
such as a blitz play versus a priming play vs a running play etc. In a
lot of positions where we are comparing two (or more) similar plays I
think we overdo the need for a full rollout, truncated is quite often
sufficient IMO. In these cases I like:

Play: 0-ply Cubeful / Cube: 33% / Truncated at Depth = 12 (at
truncation point do 2-ply 50% Play evaluation)

I certainly haven't investigated this fully, but the truncation depth
can be quite important in terms of an odd/even effect. Eg, I find even
depths are generally a much better choice than odd depths. You might
think, 11 vs 12, what difference can that make? Quite a lot from some
positions I've looked at; even seems to give you equities closer to a
full rollout result than do odd plys in my experience. In part, I'm
pretty sure this has something to do with Gnubg's often "bad" 1-ply
and sometimes "doubtful" (Albert may disagree) 3-ply evaluations
(Proviso: I can remember a couple of fairly recent positions [last six
months] where 1-ply and 3-ply evaluations were vastly better than
0-ply and 2-ply).

The cube 33% value is a sort of "hedge your bets" option with me. In
some of my previous full rollout trials I found anything less than
100% cube was left wanting; 50% was usually ok. On the other hand,
Albert did a lot of trials and found 25% cube to be good in many (if
not most) situations. The original benchmark positions that I trialled
a few months ago were deliberately pretty tough, but in light of
Albert's findings, I've now compromised to doing all my "everyday"
cubeful rollouts at 33%. Hardly a very scientific approach I know.

2) Full rollouts (Quickies)

Play: 0-ply Cubeful / Cube: 33%

0-ply plays pretty well so I don't have too many problems using it.
Maybe 0-ply cube is ok in a lot of positions too (was there a problem
with 0-ply cube and money positions a while back, I think there was?),
but it didn't do very well in my benchmark trials (mostly money
positions) so I've tended to avoid it (I've noticed Zorba likes 0-ply
Cube for a lot of his quick rollouts).

If I want to beef the rollout up a bit and use 2-ply Play, I've
settled with the following:

Play: 50% speed Skip 1-ply pruning and add upto 3 moves for 2-ply
evaluation if they are within 0.03 of 0-ply's top choice.

This may seem pretty skimpy, but it worked pretty well in my trials.
Also, I've been playing against Gnubg's Supremo setting in matches
over the last few weeks and analysing my play and Gnubg's moves with
the above 2-ply settings. In what must be close to 1000 moves made by
Supremo settings, the reduced strength settings I've advocated above
have only said Supremo was "wrong" on about 5 occasions (I wish they
had been so kind to my plays :). Each time the equity involved was <
0.015 and this gives me confidence that I have some pretty strong
settings for doing rollouts with. Also, Zorba did some excellent
trials quite recently where he showed that 2-ply 50% Play was very
good (and 33% was reasonable), but 25% was a bit dubious. This is my
interpretation of his results, not his.

If you want to start a rollout at 2-ply and finish it at 0-ply, I
would choose a separate evaluation depth of between 6 and 12. I don't
think the even/odd effect is so important here, but I still usually
stick to an even number. I'm sure 12 will get you a better result than
6, but depth 12 changeover rollouts can take a long time, even 6 is

These are just my recommendations to date, I make no claim to them
being ideal. If I've misinterpreted or misrepresented other findings
mentioned here (Albert or Zorba) please let me know.
Did you find the information in this article useful?          

Do you have any comments you'd like to add?     


GNU Backgammon

Analyzing GamesGrid matches  (Roy Passfield, Dec 2001) 
Batch analysis tool  (Øystein Johansen, June 2004)  [GammOnLine forum]
Cache size  (Ned Cross+, Mar 2004)  [GammOnLine forum]
Compiling for Windows  (Øystein Johansen, Jan 2002) 
Edit mode removing checker from bar  (Scott Steiner+, May 2003) 
Entering an annotated match  (Albert Silver, Dec 2003)  [GammOnLine forum]
Error rates: Gnu vs. Snowie  (Raccoon, Mar 2006)  [GammOnLine forum]
Even-ply/odd-ply effect  (Raccoon, Nov 2004) 
Even-ply/odd-ply effect  (Tom Keith+, Oct 2003) 
Even-ply/odd-ply effect  (Scott Steiner+, Dec 2002) 
Filter settings  (Robert-Jan Veldhuizen, Nov 2004)  [GammOnLine forum]
Gnu 0.13 versus Jellyfish and Snowie  (Torsten Schoop, Aug 2003) 
Gnu 0.13 vs. Snowie 4  (Albert Silver, June 2003) 
Gnu 0.14 vs. Jellyfish  (Michael Howard+, July 2003) 
Gnu versus Snowie and Jellyfish  (Michael Depreli, Oct 2005) 
How luck factor is calculated  (Gregg Cattanach, Aug 2002) 
How rollouts work  (Gary Wong, July 1999) 
How to enter an illegal move  (Øystein Johansen, Aug 2003)  [GammOnLine forum]
Importing .gam files  (PAR+, Mar 2005) 
Importing PartyGammon matches  (rew+, July 2006) 
Improving your game using GnuBG  (D.U.G.+, Nov 2002) 
Installing on Windows  (maareyes, Oct 2001) 
Interpreting JSD's  (Adrian Wright+, Feb 2005)  [GammOnLine forum]
JSD's and confidence intervals  (Daniel Murphy+, Jan 2005) 
Logging rollouts  (Øystein Johansen, Oct 2004)  [GammOnLine forum]
Luck rate  (Kees van den Doel+, May 2002) 
MWC versus Equity (EMG)  (Ken+, Apr 2005)  [GammOnLine forum]
Manually entering first roll  (Andreas Graf+, Apr 2005) 
Match equity tables  (Raccoon, July 2005)  [GammOnLine forum]
Personal reflections  (Louis Nardy Pillards, Sept 2002) 
Playing two computers against each other  (Stanley E. Richards+, Mar 2008)  [GammOnLine forum]
Python scripting  (Øystein Johansen+, Nov 2004) 
Quasi-random dice in rollouts  (Ian Shaw, Mar 2004)  [GammOnLine forum]
Question marks in game list  (Jim Segrave, July 2005) 
Questions and answers  (Jim Segrave+, Jan 2003) 
Questions and answers  (Jørn Thyssen, Aug 2002) 
Restarting a rollout with different settings  (Jim Segrave, Apr 2005)  [GammOnLine forum]
Restarting a rollout with different settings  (Robert-Jan Veldhuizen, Apr 2004)  [GammOnLine forum]
Rollout settings  (geoff arnold+, Apr 2007) 
Rollout settings  (Stick+, Nov 2005)  [GammOnLine forum]
Rollout settings  (Robert-Jan Veldhuizen, Mar 2004)  [GammOnLine forum]
Rollout settings  (Ian Dunstan, Aug 2003)  [GammOnLine forum]
Rollout settings for the impatient  (Robert-Jan Veldhuizen, June 2004)  [GammOnLine forum]
Running rollouts in background  (Bruce+, Apr 2004)  [GammOnLine forum]
Saving rollout results from command-line interface  (Jeremy Bagai+, Apr 2006)  [GammOnLine forum]
Saving rollouts  (Mislav Radica+, May 2006)  [GammOnLine forum]
Setting GnuBG's playing strength  (JP White, Sept 2001) 
Setting skill level  (Jim Segrave, Apr 2004) 
Setting up and saving a rollout  (Albert Silver, Dec 2003)  [GammOnLine forum]
What's GNU?  (Gary Wong, Oct 2001) 
Which player is player 0?  (Neil Kazaross+, Oct 2004)  [GammOnLine forum]

[GammOnLine forum]  From GammOnLine       [Long message]  Long message       [Recommended reading]  Recommended reading       [Recent addition]  Recent addition

  Book Suggestions
Computer Dice
Cube Handling
Cube Handling in Races
Extreme Gammon
Fun and frustration
GNU Backgammon
Luck versus Skill
Magazines & E-zines
Match Archives
Match Equities
Match Play
Match Play at 2-away/2-away
Opening Rolls
Pip Counting
Play Sites
Probability and Statistics
Source Code
Strategy--Bearing Off
Strategy--Checker play


Return to:  Backgammon Galore : Forum Archive Main Page