Hi,
I'm sending this via direct e-mail as opposed to the GNU mailing list
because the post is in HTML and I'm not sure whether the mailing list
would preserve this in a readable format. Michael Depreli posted the
following in GoL regarding some very strange GNU behaviour in 0-ply cube
decisions which I am sending to you.
Albert
Examples Of GNU 0-ply cube oddness.
Posted By: Michael Depreli <address@hidden>
Date: Saturday, 7 June 2003, at 3:07 a.m.
Here's some evidence that GNU 0 ply cube is unreliable for Double/No
Double decisions.
These were the first three double/no double decisions from one of the
money sessions between SW and GNU and nearly all the others I looked at
suffer from the same peculiarities. I haven't shown the positions
because they are not relevant.
Notice the No/Double & Double/Take equities are always identical and in
fact GNU never doubles
regardless of whether it's an optional double or not. S4's 1ply (GNU 0
ply Equiv) is shown in red.
Cube decision
0-ply cubeless equity
+0.509
0.635 0.297 0.014 - 0.365 0.070 0.001
Cubeful equities:
1.
No double
+0.770
2.
Double, pass
+1.000
+0.230
3.
Double, take
+0.770
+0.000
Proper cube action:
Optional double, take
SW4: 0.687/0.774
Cube decision
2-ply cubeless equity
+0.514
0.634 0.304 0.019 - 0.366 0.075 0.002
Cubeful equities:
1.
Double, take
+0.780
2.
Double, pass
+1.000
+0.220
3.
No double
+0.634
-0.146
Proper cube action:
Double, take
SW4: 0.683/0.784
Cube decision
0-ply cubeless equity
+0.535
0.669 0.266 0.009 - 0.331 0.076 0.001
Cubeful equities:
1.
No double
+0.846
2.
Double, pass
+1.000
+0.154
3.
Double, take
+0.846
+0.000
Proper cube action:
Optional double, take
SW4: 0.738/0.853
Cube decision
2-ply cubeless equity
+0.548
0.674 0.263 0.010 - 0.326 0.071 0.001
Cubeful equities:
1.
Double, take
+0.874
2.
Double, pass
+1.000
+0.126
3.
No double
+0.794
-0.080
Proper cube action:
Double, take
SW4: 0.792/0.882
Cube decision
0-ply cubeless equity
+0.599
0.654 0.340 0.026 - 0.346 0.074 0.002
Cubeful equities:
1.
No double
+0.962
2.
Double, pass
+1.000
+0.038
3.
Double, take
+0.962
-0.000
Proper cube action:
No double, take (0.0%)
SW4: 0.667/0.741
Cube decision
2-ply cubeless equity
+0.550
0.636 0.338 0.027 - 0.364 0.083 0.003
Cubeful equities:
1.
Double, take
+0.853
2.
Double, pass
+1.000
+0.147
3.
No double
+0.761
-0.093
Proper cube action:
Double, take
SW4: 0.692/0.721
Re: Examples Of GNU 0-ply cube oddness.
Posted By: Ilia Guzei <address@hidden>
Date: Saturday, 7 June 2003, at 3:32 a.m.
In Response To: Examples
<http://www.gammonline.com/members/board/config.cgi?read=54891> Of GNU
0-ply cube oddness. (Michael Depreli)
Michael,
I am not sure how much trust one can put into the cube evaluations at
low plies. Often S4 doubles at its 2-ply level but reverses the decision
at the 3-ply level by a significant margin. The 1-ply level is probably
even less reliable.
Also note how different the equities reported by Snowie and GNUBG are. I
still wonder why the two bots produces equity differences of such
magnitude.
Ilia
Re: Examples Of GNU 0-ply cube oddness.
Posted By: Michael Depreli <address@hidden>
Date: Saturday, 7 June 2003, at 6:15 a.m.
In Response To: Re: Examples
<http://www.gammonline.com/members/board/config.cgi?read=54892> Of GNU
0-ply cube oddness. (Ilia Guzei)
I'm already on record as stating I would never use GNU 0-ply or SW 1-ply
for evaluating cube decisions.
I posted these example because I still see rollouts posted here using
0-ply cube which I really don't trust.
To be honest I don't actually know the overall effect it has on rollouts
but it can't be good when these positions show GNU 0-ply refuses to
double!
Sure S4 changes it's decision from 2 to 3 ply because it's another ply
of accuracy with sound mathematical doubling formulae behind it.
My guess is S4 1-ply uses a simplistic cubeless to cubeful conversion
and bases it's decision around some borderline level, which although not
good enough for serious stundents IMHO at least looks more sensible than
GNU's attempt.
I guess if you analaysed a huge number of cube decisions you could come
up with a cubeless equity which would average out being borderline for
double/no double decsion. Maybe even some of Janowski's formula could
improve on this method.
Michael