Re: [Bug-gnubg] evaluations and "distribution of rolls" feature disagree

[Philippe Michel]

On Tue, Dec 12, 2017 at 02:43:42PM +0200, Rambiz Khalili wrote:
> Hi guys,
> I am not sure if I am reporting a bug. However in the position
> 710JgAC27QwGAA:cIkyAAAAAAAA
> with the roll of 54, if you look at "distribution of rolls" , 54 should be
> played 13-9 8-3 on any depth. But if you click on hint 13-9 8-3 is a big
> blunder!
> 
> How comes? Don't the "distribution of rolls" feature and "hint" use the
> same evaluation engine?

This seems to be a bug in the distribution of rolls feature. It may well 
have been there since its very beginning.

With your example, at depth 1, the distribution of rolls says that for 
54 the best move is 13/9 8/3 for an equity of -0.331. This is a 0-ply 
evaluation and a hint at 0-ply gives the same result.

When depth is raised to 2, the second layer of rolls are evaluated, 
their equities averaged and this, -0.518, is the 1-ply evaluation of 
13/9 8/3. You can check this with a 1-ply hint (this would not be as 
clear if the position at hand was not cubeless, although it wouldn't 
change the issue).

The problem is that the distribution of rolls merely expanded the 0-ply 
favorite and didn't check that 13/9 8/3 was still the best play for 54 
at 1-ply.