[computer-go] Depth dependent evaluation effects on monte carlo searches

[Jason House]

Magnus Persson wrote:
> Quoting Jason House <jason.james.house at gmail.com>:
>
>> Looking at a single color, the winning percentage seems to shift by 
>> 0.2 to 0.4%... About what I'd expect to see.  What confuses me though 
>> is how to interpret the jump back and forth as the color changes 
>> (about 8%).  Are the percentages always the winning percentage for 
>> black?  Or is it the winning percentage for the color to move?
>
> It is the latter, the winning percentage for the move just played.
>
> Root is the position after the first move of black with white to move, 
> I changed
> the colors in your table.
> W: 54.3%
> B:         46%
> W: 54.6%
> B:         46.3%
> W: 54.7%
> B:         47.6%
> W: 54.9%
> B:         47.0%

Taking that into account, the winning percentages for white would be:

W: 54.3%
B:         54%
W: 54.6%
B:         53.7%
W: 54.7%
B:         52.4%
W: 54.9%
B:         53.0%

>> If it's the winning percentage for the color to move, it seems really 
>> strange that it'd go up for both colors as the principle variation 
>> went on.
>
>
> If it goes up for both players it might mean that the games get more 
> hot, that
> is, a pass become more and more catastrophic.
>
> A more boring explanation is that as you go deep in the tree, for 
> statistical
> reasons the program most likely finds moves that have been 
> overestimated, and
> with deeper search these values will come closer to the average.

I'm guessing the 2nd explanation is probably the case.

Thanks a lot for generating that data for me.  It was interesting to 
see.  I think it definitely shows the fluctuation by color to move.  
I'll probably try and generate some pure MC numbers (without growing 
trees) to see how they look.  When I generate the data, I'll post it to 
the list.