Hi all,
Actually, the article on floating-point effects is an interesting one too:
The Ghost in the Model (and Other Effects of Floating Point Arithmetic)
by Gary Polhill, Luis Izquierdo and Nick Gotts
<
http://jasss.soc.surrey.ac.uk/8/1/5.html>
It reminds me of a worry that I've had about the validity of agent-based
models that use numbers of unthinking precision (by which I mean, just
calling a variable a float and leaving issues of precision to the compiler).
Does it make sense to measure many of the variables we use in ABMs to
15-decimal places?
For example, suppose I'm building a model and need to describe (and give a
value) to my agent's happiness, and I'm then going to be basing some agent
interaction on the relative levels of those agents' happinesses. Would it
really make sense to be saying an agent with happiness 1.000002 is happier
than one with happiness 1.000001? I'm not sure how many levels of my own
personal happiness I can discriminate (more than three, less than a million)
but my compiler can discriminate between orders of magnitude more, and will
do so unless I tell it otherwise.
So what's the solution? I can think of two possibilities:
(1) If I'm looking at lots of agents and lots of interactions, stick
some noise terms in. That way, sometimes 1.00002 is bigger than 1.00001 and
sometimes it isn't;
(2) Always round every evaluation of a variable to the nearest
appropriate precision. So if I want a happiness scale from 0 (misery) to 1
(joy) and I reckon I can distinguish 10 levels of happiness, my agents'
happiness can only ever take values [0, 0.1, 0.2.. 1.0]
Any thoughts?
Dr. Robert Holmes
PO Box 2862, Santa Fe, NM 87504
mobile: (505) 310-1735
web: www.holmesacosta.com
-------------- next part --------------
An HTML attachment was scrubbed...
URL: /pipermail/friam_redfish.com/attachments/20050203/42ad428c/attachment.htm
Locked
