I correctly, and somewhat confidently took below.
Why? Because I have a clear memory of listening to a youtube
lecture by Mochy where he gave a very similar position and explained
that although we know the slogan "three roll positions are passes", the
fact that the cuber has one number (21) that loses immediately is enough
to swing the position to a take.
XGID=-DB-------------------a-e-:2:-1:-1:00:0:4:3:0:10
X:eXtremeGammon O:Daniel
Score is X:4 O:0. Unlimited Game, Jacoby Beaver
+13-14-15-16-17-18------19-20-21-22-23-24-+
| | | O O |
| | | O O |
| | | O |
| | | O |
| | | |
| |BAR| |
| | | X |
| | | X |
| | | X | +---+
| | | X | | 4 |
| | | X X | +---+
+12-11-10--9--8--7-------6--5--4--3--2--1-+
Pip count X: 8 O: 8 X-O: 4-0
Cube: 4, X own cube
X on roll, cube action
Analyzed in 4-ply
Player Winning Chances: 76.13% (G:0.00% B:0.00%)
Opponent Winning Chances: 23.87% (G:0.00% B:0.00%)
Cubeless Equities: No Double=+0.523, Double=+1.045
Cubeful Equities:
No redouble: +0.702 (-0.247)
Redouble/Take: +0.949
Redouble/Pass: +1.000 (+0.051)
Best Cube action: Redouble / Take
eXtreme Gammon Version: 2.10
Strange anaysis result. In such a short race the equities can be
calculated exactly (and are probably in a bearoff database). No redouble shoud be 0.700, Redouble/Take 0.980. How can XG get these wrong when all
its other numbers are accurate ?
On 12/15/2021 5:59 PM, Philippe Michel wrote:It's interesting that XG made an error here. I was actually surprised to see a take
Strange anaysis result. In such a short race the equities can beIt gets it right with 5-ply or higher (which corresponds to GNU 4-ply
calculated exactly (and are probably in a bearoff database). No redouble shoud be 0.700, Redouble/Take 0.980. How can XG get these wrong when all its other numbers are accurate ?
or higher). So my guess is that a 4-ply evaluation is effectively
a "truncated rollout" with the checker play being governed by the
neural net directly, and the truncated evaluation also being calculated
by consulting the net rather than the database?
as
On 12/15/2021 5:59 PM, Philippe Michel wrote:
Strange anaysis result. In such a short race the equities can be
calculated exactly (and are probably in a bearoff database). No redouble
shoud be 0.700, Redouble/Take 0.980. How can XG get these wrong when all
its other numbers are accurate ?
It gets it right with 5-ply or higher (which corresponds to GNU 4-ply
or higher). So my guess is that a 4-ply evaluation is effectively
a "truncated rollout" with the checker play being governed by the
neural net directly, and the truncated evaluation also being calculated
by consulting the net rather than the database?
It's certainly a possibility, but then in what circumstances is the
bearoff database used ?
Sysop: | Keyop |
---|---|
Location: | Huddersfield, West Yorkshire, UK |
Users: | 300 |
Nodes: | 16 (2 / 14) |
Uptime: | 54:21:36 |
Calls: | 6,712 |
Files: | 12,243 |
Messages: | 5,355,326 |