# sum out a random variable

I have a bayesian network like that:

B -> A <- F

And this are the values for A:

P(A=true | B=true, F=false) = 0.01

P(A=true | B=true, F=true) = 0.92

P(A=true | B=false, F=false) = 1.00

P(A=true | B=false, F=true) = 1.00

Now I want to sum out the variable B because I need the values P(A=true | F=true) and P(A=true | F=false). What I thought I have to do is sum this lines where the F values are the same and B is true and false. So line 1 and line 3 on the one side and line 2 and 4 on the other. My result would be:

P(A=true | F=true) = 1.92

P(A=true | F=false) = 1.01

I was quite sure this is right but now I wonder why the sum is greater then 1!? I thought probabilities have to go between 0 and 1? Am I doing something wrong or where is my fault?

-
You should change the variable $C$ to $F$ in your graphical model in order to be consistent with the rest of your question. – emrea May 28 '12 at 22:41
sorry, you're right :) – Bazinka May 29 '12 at 10:34

Don't you need the prior probabilities of $B$ and $F$?
$P(A=t|F=t) = \frac{P(A=t,F=t)}{P(F=t)} = \frac{P(A=t,B=t,F=t)+P(A=t,B=f,F=t)}{P(F=t)}$
$= \frac{P(A=t|B=t,F=t) P(B=t) P(F=t) +P(A=t|B=f,F=t)P(B=f)P(F=t)} {P(F=t)}$
$= 0.92 P(B=t) +1P(B=f)$
Isn't that a different approach? I wanted to do it with marginalization like $\sum_{B} P(A | B, F)$ – Bazinka May 29 '12 at 10:41