Re: [Bug-gnubg] Strange cube decision

[Ned Cross]

In general with these types of positions, the reason for the bot's 
ambiguous opinion is the fact that there is zero volatility.  Whatever 
the correct decision is this turn, there is no sequence that will change 
it for next turn. 

In the example given, repeated rolls of 22 are the only thing bad that 
can happen for X, and O's situation does not change until X starts 
bearing off.

For the plies where GNU (incorrectly) thinks the winning % is in the 
lower end of the doubling window, GNU (correctly) sees no market losing 
sequences, hence the 'optional double'.  My guess is that 4-ply says 
"no-double" instead of "optional double" because of slight round-off error.

Note that with deeper plies, (0->2->4) the win% (and hence the no-double 
or double/take equity) is creeping up as GNU gets closer to 
'recognizing' the position.

For the plies where GNU (correctly) thinks the winning % is in the 80% 
range, the 1-ply look-ahead only sees that doubling is better then not 
doubling, but 3-ply see that doubling this roll really doesn't matter 
since it's still a double/pass next roll no matter what.

The analogous checker-play situation is when we see a bot make moves 
that seem counter-intuitive to us near the end of a game in a gin 
position. Since the bot (correctly) sees little or no equity difference, 
it makes apparently  random moves.

Ned

Jim Segrave wrote:

> Someone at dailygammon queried this gnubg analysis and I have to say
> that I find the results a bit odd. 
>
>