[Bug-gnubg] rollout cube decision -- money play: BUG

[Robert-Jan Veldhuizen]

Hi team,

Look at this:

PosID: 794ggABb2wRDAA
MatchID: cAkAAAAAAAAA

A cube decision in a money session. 4-ply evaluation (24 candidates, 1.00 
tolerance) gives this answer:

        Win     W(g)    W(bg)   L(g)    L(bg)   Equity  (cubeful)
static:  69.8%   14.8%    0.2%    2.7%    0.0%  (+0.396  (+0.831))
 1 ply:  70.2%   16.6%    0.2%    2.6%    0.0%  (+0.404  (+0.887))
 2 ply:  68.8%   14.5%    0.1%    3.0%    0.0%  (+0.376  (+0.798))
 3 ply:  71.0%   17.8%    0.2%    3.5%    0.1%  (+0.420  (+0.969))
 4 ply:  69.4%   15.6%    0.2%    3.0%    0.0%  (+0.388  (+0.855))

Double, take        : +0.855
Double, pass        : +1.000   (+0.145)
No double           : +0.417   (-0.438)
Correct cube action: Double, take

Now (cubeful, variance reduced) rollouts come up with this:

                              Win    W(g)   W(bg)  L(g)   L(bg) 
Equity  Cubeful
             Centered 1-cube 0.6886 0.1440 0.0015 0.0368 0.0006 +0.3724 
+0.6093
              Standard error 0.0008 0.0008 0.0001 0.0007 0.0001  0.0017 
+0.0143
     Player shos owns 2-cube 0.7032 0.1855 0.0051 0.0391 0.0013 +0.5565 
+0.5777
              Standard error 0.0012 0.0018 0.0006 0.0009 0.0002  0.0036 
+0.0092

Full 0-ply, 1944 trials

                              Win    W(g)   W(bg)  L(g)   L(bg) 
Equity  Cubeful
             Centered 1-cube 0.7045 0.1764 0.0024 0.0303 0.0003 +0.4089 
+0.7264
              Standard error 0.0009 0.0008 0.0001 0.0008 0.0001  0.0022 
+0.0215
     Player shos owns 2-cube 0.7014 0.1789 0.0058 0.0369 0.0008 +0.5498 
+0.5745
              Standard error 0.0013 0.0025 0.0010 0.0011 0.0002  0.0044 
+0.0105

Full 1-ply(small), 1296 trials

                              Win    W(g)   W(bg)  L(g)   L(bg) 
Equity  Cubeful
             Centered 1-cube 0.6983 0.1578 0.0017 0.0324 0.0008 +0.3974 
+0.6775
              Standard error 0.0007 0.0008 0.0000 0.0007 0.0001  0.0017 
+0.0155
     Player shos owns 2-cube 0.7061 0.1800 0.0042 0.0326 0.0008 +0.5629 
+0.6439
              Standard error 0.0011 0.0021 0.0006 0.0008 0.0002  0.0033 
+0.0131

Full 2-ply(small) 33%, 1296 trials

All three results indicate a pretty big drop, and the 2-ply one suggests 
taking is a huge blunder: 2*(0.6439-0.5)=0.2878.

But there must be a bug somewhere in these reports; the cubeful equity 
after taking just can't be right if you look at the win/loss breakdown. 
Even if opponent never uses the cube, his equity should be better after 
taking than what's reported; and the cubeful equity should be lower or at 
best equal to the (cubeless) "equity" in the bottom row; opponent should 
get some use out of his recube.

I set the same position up as in a 25pt match, score 0-0, Snowie MET. A 
cubeful, variance-reduced rollout of the position after double/take (opp 
owns 2-cube) then gives this answer, which seems correct and is much 
different from the ones above:

PosID: 794ggABb2wRDAA
MatchID: QQkgAwAAAAAA

                              Win    W(g)   W(bg)  L(g)   L(bg) 
Equity  Cubeful
            Current position 0.7034 0.1867 0.0058 0.0396 0.0009 +0.5611 
+0.5005
              Standard error 0.0013 0.0023 0.0008 0.0012 
0.0002  0.0043  0.0092

Full 0-ply, 1296

                              Win    W(g)   W(bg)  L(g)   L(bg) 
Equity  Cubeful
            Current position 0.7013 0.1794 0.0058 0.0374 0.0009 +0.5519 
+0.4823
              Standard error 0.0013 0.0025 0.0010 0.0011 
0.0002  0.0045  0.0094

Full 1-ply(small), 1296

Both indicating a borderline drop/take decision, which is in agreement with 
Jellyfish and SW rollouts of the money game position, BTW.

I think I've been noticing weird things with money game rollouts more 
often; probably just "rollout cube decision" and then just the bottom row 
for double/take. GNUBG sometimes makes very big passes out of positions 
that are just small passes at worst I think.

I hope it's just some bug in a formula somewhere. Maybe gammons are counted 
twice, or gammons are counted as backgammons? I think there might be a 
relation to the gammon wins, but I'm not sure. Anyway, something is not 
right here.

Something else: perhaps it's easier for most people if GNUBG would double 
the equities of the bottom row in money game rollouts of cube decisions? 
That would also be more in line with the way match cubes are done (drop 
when take equity for doubler > +1.000)

Greetings,

--
Robert-Jan (Zorba on FIBS)