RE: [Bug-gnubg] 2-ply vs. 3-ply

[Ian Shaw]

 rambiz khalili wrote:
        
        
        hi all,
        i would like to know why gnu 2-ply overestimates the equity of a
certain move or position in comparison with gnu 3-ply? to me it seems
like a SYSTEMATIC bug in the evaluation function! 
        many people i have spoken to say that gnu 2-ply is more reliable
than gnu 3-ply! is it true?
        thx for considering my question
        regards
        rambiz 


Hi rambiz,

Short answer: No! It is not true.

Long answer:

This is a question that recurs on the mailing list from time-to-time, in
one form or another.

The evaluations will be different because:

A) 3-ply is looking one roll deeper than 2-ply.
B) That last play is effectively looking at the position from the
point-of-view of the other player. The neural network's weights will be
different.

I've not seen any evidence that 2-ply is more reliable than 3-ply. In
tests of 3-ply vs 2-ply, I believe that 3-ply has a very small edge. If
your many people would care to back 2-ply against 3-ply, they'd probably
find many takers!

A counter example. Several people have noted that, in advanced-anchor
holding games, 2-ply evaluations are too optimistic for the holding
player, and that 3-ply evaluations are closer to rollout results. There
have been posts at GammonU in the past about this. In this sense, 3-ply
is MORE reliable for these holding game cubes.

There is likely to be an odd-even ply effect, that is common to most
two-player games, where each side is trying to maximise it's own equity
by giving it's opponent the worst possible best move. There has been
discussion on this in the past. You can search the archives for stuff.
The thread containing the following post is useful: 

http://lists.gnu.org/archive/html/bug-gnubg/2003-06/msg00207.html

I think Thomas Hauk's thesis, which used gnubg for its tests, covers the
topic, too. http://www.cs.ualberta.ca/~jonathan/Grad/hauk.pdf

I hope this helps.
Regards,
Ian