Gnubg and other bots: odd-ply vs. even-ply

[Scott Steiner]

Hello again,

I have a question concerning gnubg's analysis' ply level. I suppose it
is a general question that affects all bots, not just gnubg.

Reading posts on this NG, I've heard several times that gnubg's analysis
at 1-ply can be unfortunate at times and produce not so good results and
that one should rather use 2ply beacuse it gives much more accurate
results, which I do anyway. Further, I got the impression that odd-ply
analysis is not so good as even-ply, so I ask if there is any truth to
this and briefly how to explain this.

For instance, I usually analyse using gnubg's preset worldclass++
setting i.e on 2-ply. On some moves where I have a dispute with gnubg's
2-ply analysis, I evaluate the top moves or the cube decision at 3-ply
and see how my play ranks there. Sometimes I would go up to 4-ply,
although very rarely because that takes too long. I notice quite often
that the even-ply analysis are usually close and agree among each other
but they disagree with the odd-ply analysis i.e 2 and 4 ply would agree
but 3ply gives different results. Actually I noticed that quite often
when going through some examples of Robertie's Advanced Backgammon.

So, I guess the question is what is behind this odd-ply and even-ply
analysis level and which should I trust more and what pitfalls are there
when using either one.

thx!

[Gary Wong]

Scott Steiner <nos...@nospam.nospam> writes:
> So, I guess the question is what is behind this odd-ply and even-ply
> analysis level and which should I trust more and what pitfalls are there
> when using either one.

You are quite right that several types of positions exist where odd- and
even-ply evaluations disagree. Unfortunately, there is no general
procedure to decide which one to trust more; in fact, if the two disagree
by a significant amount, then that is a good indication that you should
be suspicious of both.

A long time ago, I found some results which made me suspect that 1-ply
evaluations were _worse_ than 0-ply for an old version of gnubg.
However, a subsequent longer run of tens of thousand of games
indicated that 1-ply was indeed stronger by a small but significant
amount (+0.012ppg cubeless, +/-0.004). On average, you can expect n+1
ply evaluations to be more accurate than n-ply evaluations; there is
nothing to suggest that even or odd n is better overall.

Cheers,
Gary.