Re: [Bug-gnubg] crashed position

[Ian Shaw]

I'll just emphasize Joseph's explanation. 


Mark wrote:
" I'd imagine a crashed position is one where you're bearing in against an 
opponent anchor and have to start dismantling your beautiful barricade as the 
checkers come in."

This is not the essence of a crashed position. It's not about bearing in 
against an anchor (although some of these will be in the crashed category). A 
position is crashed if most of  a players chequers have ended up on his own 1 
and 2 points, giving him little flexibility. He has no control of his inner 
board or the outfield, which allows his opponent to spread his chequers around 
without fear of being hit and contained. This calls for different tactics from 
the standard contact positions. 

Perhaps a couple of diagrams will help (view with fixed-width font). Player X 
has the same structure each time; the difference is in the opponent's position. 
In the crashed position X can safely slot the front of the prime to roll it 
forward, but this would be too dangerous when O still has good structure.


    GNU Backgammon  Position ID: /z4AADBsuxsEAA
                    Match ID   : cAngAAAAAAAE
    +24-23-22-21-20-19------18-17-16-15-14-13-+  O: White
    | O  O             |   |                X |  0 points
    | O  O             |   |                  |  
    | O  O             |   |                  |  
    | O  O             |   |                  |  
    | 8  O             |   |                  |  
    |                  |BAR|                  |v 7 point match (Cube: 1)
    |                  |   |                  |  
    |                  |   |                  |  
    |                X |   | X                |  
    | O     X  X  X  X |   | X  X             |  On roll
    | O     X  X  X  X |   | X  X             |  0 points
    +-1--2--3--4--5--6-------7--8--9-10-11-12-+  X: Blue

    GNU Backgammon  Position ID: sN0tADBsuxsEAA
                    Match ID   : cAngAAAAAAAE
    +24-23-22-21-20-19------18-17-16-15-14-13-+  O: White
    |             O  O |   | O  O  O  O     X |  0 points
    |             O  O |   | O  O  O          |  
    |                  |   | O  O             |  
    |                  |   |                  |  
    |                  |   |                  |  
    |                  |BAR|                  |v 7 point match (Cube: 1)
    |                  |   |                  |  
    |                  |   |                  |  
    |                X |   | X                |  
    | O     X  X  X  X |   | X  X             |  On roll
    | O     X  X  X  X |   | X  X             |  0 points
    +-1--2--3--4--5--6-------7--8--9-10-11-12-+  X: Blue


I hope this helps.

-- Ian




-----Original Message-----
From: address@hidden [mailto:address@hidden On Behalf Of Joseph Heled
Sent: 08 February 2012 23:09
To: Mark Higgins
Cc: address@hidden; Øystein Schønning-Johansen
Subject: Re: [Bug-gnubg] crashed position

CRASHED attempts to capture the positions where one side has only a
small number of "active pieces". The number of active pieces has been
arbitrarily set at 6, and the definition requires that you have at
most 6 checkers not on points 1 or 2, accounting for the possibility
of  one checker from 2 sent back after the rest piled on point 1.

The most important part in this celebration of arbitrary decisions was
to use a definition which is non cyclic - positions resulting from a
crashed positions should be crashed. When this is violated,
performance deteriorates since each net is trained only on it's own
kind of positions.

That was my experience anyway. I will be happy to see someone coming
up with a better definition and performance. GNUbg pathetic play in
many backgame situations leaves it open to abuse from humans.

-Joseph

On 9 February 2012 00:23, Mark Higgins <address@hidden> wrote:
> Revisiting this one - I read the eval.c ClassifyPosition code, so have a
> decent idea of how gnubg defines "crashed" (it's not what I described
> below).
>
> What I don't get is why it uses this particular definition.
>
> ie I'd imagine a crashed position is one where you're bearing in against an
> opponent anchor and have to start dismantling your beautiful barricade as
> the checkers come in.
>
> So why isn't crashed something simple like "contact, and at least one player
> has all their checkers at their nine point or closer"? Seems like that's
> roughly when you'd start caring about how to bear off against an anchor.
>
> Or maybe you'd replace "nine point" with "six point" if you want to get
> closer to the end of the game. But I don't really see why how many checkers
> are on the 1 or 2 point specifically matter than much (vs the 3 point, or
> why >1 checker is the threshold vs >0 checkers).
>
> Anyone remember the motivation for the current definition?
>
>
>
> On Dec 17, 2011, at 12:22 PM, Øystein Schønning-Johansen wrote:
>
>
>
> On Sat, Dec 17, 2011 at 4:05 PM, Mark Higgins <address@hidden> wrote:
>>
>> I'm trying to find the exact definition gnubg uses for a "crashed"
>> position.
>>
>> The one reference I've found (Thomas Haug's thesis) says it's contact,
>> plus the restrictions that the player has fewer than 7 pieces remaining with
>> none in the opponent's 1 or 2 position. Is that correct?
>>
>> If so, can someone give a little color on why those particular
>> restrictions? eg why is it contact if the player has a piece on the
>> opponent's 2 position, but crashed if it's on their 3 position?
>>
>
> The source is the documentation!
> http://cvs.savannah.gnu.org/viewvc/gnubg/gnubg/eval.c?revision=HEAD&view=markup
>
> Search for the function called ClassifyPositon()
>
> -Øystein
>
>
>
> _______________________________________________
> Bug-gnubg mailing list
> address@hidden
> https://lists.gnu.org/mailman/listinfo/bug-gnubg
>

_______________________________________________
Bug-gnubg mailing list
address@hidden
https://lists.gnu.org/mailman/listinfo/bug-gnubg