buggnubg - Bugs: bug #6423, cube rollout gives wrong results

 
 

bug #6423: cube rollout gives wrong results

Submitted by:  M.Alma <misja>
Submitted on:  Fri 07 Nov 2003 08:08:07 PM UTC  
 
Originator Name: misjaItem Group: Other bug
Category: RolloutsRelease: 0.14-devel
Privacy: PublicOperating System: Windows XP
OS Type: NoneProcessor Type: None
Severity: 3 - NormalOpen/Closed: Closed
Assigned to: Joern Thyssen <thyssen>Priority: 5 - Normal
Status: Invalid

Add a New Comment(Rich Markup)
   

You are not logged in

Please log in, so followups can be emailed to you.

 

Sun 09 Nov 2003 05:30:33 PM UTC, comment #3:

Thank you for your quick answer. Let me start by saying that although my bug might sound negative, I do highly appreciate the gnubg project and that I am amazed how much progress you have booked since the last time I saw gnubg (I think that was one year ago).

You made 2 points in your answer, both of which I would like to comment on.
In point 1) you say the I can't compare the cubeless and cubefull GWCs because gnubg plays the position according to score.
I agree that checker play at certain scores can be very different depending on the position of the cube. But not at this score: It is 0-0 in a seven point match and the cube is still in the middle. Still checker plays might be different but not so much that they give this result; that one starts to win more games when the cube is with the opponent.

In point 2) you mention the possibility that gnubg might have taken some poor doubles, or that the strange rollout result comes from gnubg's potentially poort understanding of prime vs prime.
To check this I have done the following: I have installed the latest build I could find (08-11) and rolled out the position again. This time it still thinks the position is a pass but a double as well, albeit by a narrow margin (0.12%). But I haven't been able to find the option to save the rollout trials to a SGF file.
If you could tell me where to find this then I will send them to you.

M.Alma <misja>
Sun 09 Nov 2003 12:11:33 PM UTC, comment #2:

(1) You can't compare GWCs obtained in a cubeless rollout with GWCs obtained from a cubeful one. Note gnubg plays according to score, so the play may be very different between the two.

(2) The "no double" and "double, take" rollouts and two independent rollouts, so they may very well lead to very different GWCs. For example, the "double, take" branch is almost always played to the truncation point, whereas the "no double" branch will often result in "double, pass" after a few rollouts (if X hit), effectively truncating the rollout after very few plies. OTOH, gnubg may take some bad doubles for O in the "no double" branch leading to the strange "too good" result.

In conclusion, there are no bugs in the rollout code -- this particular rollout is very tricky due to gnubg's potentially poor understanding of prime-vs-prime.

The newest versions of gnubg are able to save the rollout trials to an SGF file which can be analysed afterwards. If you do that, I'll bet you either see (a) the "no double" branch truncated with "double, pass" after very few plies, or (b) very bad takes by gnubg in the "no double" branch.

Joern Thyssen <thyssen>
Project AdministratorIn charge of this item.
Fri 07 Nov 2003 08:12:06 PM UTC, comment #1:

second try to send the attachment ..

M.Alma <misja>
Fri 07 Nov 2003 08:08:07 PM UTC, original submission:

When I rolled out a cube decision I got wrong results:
The side that gave away the cube won more games in the cubeful rollout than in the cubeless rollout. This should be the other way around.

This all resulted in a wrong cube decision: A position that is an easy take according to a Snowie rollout (and also gnubg's own evaluation) now becomes too good to double and a pass according to the rollout.

My rollout settings:
108 trials, truncated after 11 plies,
rollouts were cubeful but truncated cubelass at bearoff database,
variance reduction was on, quasi random dice,
Seed: 1182907264, cube decisions same as checker play,
use same settings for both players, use player0 setting for truncation point.
Checker play for 'first player both' was world class.
Settings which I haven't mentioned were off or default.

Here are the results I got:
Cubeless rollout MWC: 53.45%
Cubefull:
1. No double: 56.24%
0.758 0.101 0.003 0.242 0.063 0.002
2. Double, Take: 56.73%
0.770 0.110 0.003 0.230 0.060 0.002
3. Double, Pass: 56.17%

If I understand this right then these results are telling that by doubling one could gain 1.2% game winning chances...

The position is included in the attachment.

M.Alma <misja>

 

(Note: upload size limit is set to 16384 kB, after insertion of the required escape characters.)

Attach File(s):
   
   
Comment:
   

Attached Files
file #807:  sik-07112003-1.txt added by misja (2KiB - text/plain)

 

Depends on the following items: None found

Items that depend on this one: None found

 

CC list is empty

 

Do you think this task is very important?
If so, you can click here to add your encouragement to it.
This task has 0 encouragements so far.

Only logged-in users can vote.

 

Please enter the title of George Orwell's famous dystopian book (it's a date):

 

 

Follow 4 latest changes.

Date Changed By Updated Field Previous Value => Replaced By
Sun 30 Nov 2003 11:05:35 AM UTCthyssenOpen/ClosedDeclined=>Closed
Sun 09 Nov 2003 12:11:33 PM UTCthyssenOpen/ClosedOpen=>(Error - Not Found)
  Assigned toNone=>NA
  StatusNone=>(Error - Not Found)

Back to the top


Powered by Savane 3.1-cleanup1