Hello,

I have a discrete multivariate Kalman filter. I have 30 variables (or
fields or inputs) I can track them every day.

However, two things can happen:
1- One of these variables can become irrelevant in the state equation
2- I have a large pool of "potential" variables. These variables are
not relevant to my filter at the moment, but they could become in the
future. To give you an idea of the magnitude, I have a poll of 500
"potential" variables and a maximum of 1 or 2 could become relevant in
my filter each day.

How would you test for (1) irrelevant variables and (2) test for
potential inclusions in the model.

My problem is not related to meteorology, but I figure that you guys
could give me indications as you are pros in data assimilation.

If I had to invent a problem that is similar to mine in terms you are
familiar with, I would say that I want to explain wind velicity as the
sum of all nearby air currents. Sometimes air currents disappear. And
sometimes major air currents come from further away and should included
in the equation.

Thanks

There's nothing special about a Kalman filter. Just treat it as
a potential multivariate normal models, including whatever you want,
for deviations around the nomimal current state.

The usual drill is how much variance reduction you get vs. how
many degrees of freedom you're putting in.

Chi squared and the rest, on which I'm not particularly expert.

--
Ron Hardin


On the internet, nobody knows you're a jerk.

On 7 Aug 2006 14:23:27 -0700,
, in
.com wrote:

+ How would you test for (1) irrelevant variables and (2) test for
+ potential inclusions in the model.

Run a sensitivity study to determine which variables are important,
and which are not. Just shove these terms into google:

model sensitivity study

--
Consulting Minister for Consultants, DNRC
I can please only one person per day. Today is not your day. Tomorrow
isn't looking good, either.
I am BOFH. Resistance is futile. Your network will be assimilated.