My question is related to multivariable joint distribution. I have one source variable "x" and multiple receivers "y1" "y2" "y3". I have each joint distribution p(x,y1),p(x,y2), p(x,y3). My question is how do I get p(x) from combination of the 3. The issues in my mind are
If I calculate p(x) from p(x,y1). I already think that this should be exactly the p(x) obtained from other joint distributions. But in a real scenario, we have to estimate these distributions which would yield different marginal for p(x).
I do not have code yet but if someone can point out the direction then it would really be helpful
I worked on modeling sensors with belief networks in my dissertation. See: http://riso.sourceforge.net My dissertation is a little ways down on the page. A model for sensors which measure the same thing is described in Section 6.5.
In brief, when you have multiple measurements y1, y2, y3
of the same thing x
, you can model the joint probability of all of them as p(x, y1, y2, y3) = p(y1 | x) p(y2 | x) p(y3 | x) p(x)
, where each p(y | x)
is a measurement model, i.e., it represents the way that the measurement is a function of the thing being measured. Then the goal is to compute p(x | y1, y2, y3)
. It turns out that's proportional to p(y1 | x) p(y2 | x) p(y3 | x) p(x)
, with the constant of proportionality being whatever is needed to make the expression integrate to 1 over x
. I.e., to combine information from multiple sensors, given this model, you multiply them together.
If you open a question on stats.stackexchange.com, I can say more about it. Hope this helps.