Fahd:

What you propose could be done quite simply using a single-layer neural
network. Treat each sensor as either a binary response variable, or some
type of continuous value for more complex sensors. Weight the inputs
from each of these sensors and assign some threshold function that
translates weighted combinations of these inputs to your categorical
outputs of interest. Then, have the network 'learn' the relationship
between the inputs and outputs.

On the other hand, if you are sure you want to use Bayesian approach,
then you might try the following way of thinking. 

Recall:

P(model | data) = P(data | model) P(model) / P( data )

represents the posterior probability given prior information (a set of
beliefs), evidence (observations), and a likelihood (the model).

List out your inputs and the outputs, and draw a directed acyclic graph
to represent the 'network' you are interested in, to help you think
about the problem.

The evidence, P(data), is computed by taking the outputs of each of the
sensors and computing the probability of THAT specific combination of
outputs over all possible combinations of outputs.

Assume 'model' here represents the relationship between the 'mood' of
the person as inferred from the biosensors. P(model) would be the prior
probability of that person being in any particular mood. If this is a
time-dependent network, you can compute the prior for this based on
recent historical data; but if this is a static case, then you can
choose to use an uninformative prior, or you may also assume that it is
more likely that the person is in a 'happy mood' some % of the time,
which gives you an informed prior.

P(data | model) is the likelihood of seeing a particular set of outputs
from the sensors given that the person is in a particular mood.

P(model | data) is the posterior probability you are trying to estimate,
that is, the probability that the person is in a specific 'mood' based
on the available data from biosensors.

There are many useful tutorials out there on the web to get you started
with Bayesian networks. You can look into the various literature in the
Machine Learning community to get you started. Some of the work by
Ghahramani is very useful in this respect, as he has worked out the
similarities and relationships between various types of models. Google
it.

I'm not sure why you would need to do a 'regression analysis' if you're
using the bayesian strategy; but then again, i'm not sure you know which
approach you are actually interested in.

hope that helps...
P

.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to