Smoothing a discrete set of data

[GerritM]

"Paul Moore" <gustav at morpheus.demon.co.uk> schreef in bericht
news:3csmdso5.fsf at morpheus.demon.co.uk...
<..snip...>
> I have a set of data, basically a histogram. The data is pretty
> variable, and I'd like to "smooth" it to find trends. Actually, this
> comes up a *lot* for me - some examples: sampled IO rates on a machine
> - I want to look for trends (is IO higher overnight or during the day,
> etc) or fuel consumption for my car (do I use less fuel since I had
> the service).
<..snip...>
> Can anyone help me? I can't believe that this problem has never come
> up before, but I can't find any literature on it.
>
Yes, this problem is very common. For instance in Physics it occurs very
often. In physics the approach would be to fit "lines", where the line form
depends on the particular physics.

In your example it sounds as if less theory is available (how about
operations research?), which means that you have less a priori knowledge of
the signal. Your brain picks out the evident peaks.

One approach is to filter out the high frequent noise by simply applying a
low pass filter, ie a weighted average of neighboring points.

A more complex approach would be to assume some distribution in peaks (for
instance Gaussian) and fit a set of these peaks over the data. Depending on
your expectation the peaks can be parameterized for width and amplitude for
instance.

I expect that in the end your interpretation is still required, but the job
might be eased by this kind of computer assistance.

> Thanks in advance,
> Paul Moore.

regards Gerrit

--
www.extra.research.philips.com/natlab/sysarch/