2
$\begingroup$

I have trouble understanding, why my solution is wrong.

Given:

$P(A) = 0.5\\ \forall i: P(X_i |A ) = 0.2\\ \quad P(X_i|\neg A) = 0.6$

(And all $X_i$ are dependant on and only on $A$.) This is bad mathematical formulation. There was a graph, with $A$ at top and arrows from $A$ to all $X_i$

Asked: $P(A|X_1 X_2 \neg X_3)$

The solution should be $\approx 0.1818...$ As far as i know my numerator is correct and i don't understand why my denominator isn't.

My Approach:

I use the equation $P(A|B) = P(A B)/P(B)$ over and over.

$P(A|X_1 X_2 \neg X_3) = P(A X_1 X_2 \neg X_3)/P(X_1 X_2 \neg X_3) = \\ \frac{P(\neg X_3 | A X_1 X_2) P(AX_1X_2)}{P(X_1 X_2 \neg X_3)} = \frac{P(\neg X_3 | A X_1 X_2) P(X_2|AX_1) P(AX_1)}{P(X_1 X_2 \neg X_3)} =\\ \frac{P(\neg X_3 | A X_1 X_2) P(X_2|AX_1) P(X_1|A) P(A)}{P(X_1 X_2 \neg X_3)}$

Since $X_i$ is never dependant on any other $X_j$ i can simplify into:

$\frac{P(\neg X_3 | A) P(X_2|A) P(X_1|A) P(A)}{P(X_1 X_2 \neg X_3)}$

All the numbers for the numerator are given: $\frac{0.8\cdot 0.2 \cdot 0.2 \cdot 0.5}{P(X_1 X_2 \neg X_3)} = \frac{0.016}{P(X_1 X_2 \neg X_3)}$

So now i need to find out what $P(X_i)$ is. $P(X_i) = P(X_i|A) P(A) + P(X_i| \neg A) P(\neg A)\\ = 0.2 \cdot 0.5 + 0.6 \cdot 0.5 = 0.4$

Since $X_i$ are indpenendent, $P(X_1 X_2 \neg X_3) = P(X_1) P(X_2) P(\neg X_3)\\ = 0.4 \cdot 0.4 \cdot 0.6 = 0.096$

So $P(A|X_1 X_2 \neg X_3) = \frac{0.016}{P(X_1 X_2 \neg X_3)} = \frac{0.016}{0.096} = 0.166...$ which is not equal to $0.1818...$

What is my mistake with the denominator? Have i expanded incorrectly? Is my understanding of dependant/independant variables wrong?

(I know, no homeworks. This is from a udacity course. Complete solution is available. I just want to know, where i have done a mistake with my approach, because clearly i have difficulties with understanding dependent variables)


Official solution:

Get the numerator: $P'(A|X_1 X_2 \neg X_3) = 0.016$ (as with my solution)

Get the numerator: $P'(\neg A|X_1 X_2 \neg X_3) = 0.072$ (double checking with $ = 0.4 \cdot 0.6 \cdot 0.6 \cdot 0.5$)

Both P's are not weighted and together they should add up to 1. So $0.016 + 0.072 = 0.088$ is the weighting factor.

$\Rightarrow \frac{P'(A|X_1 X_2 \neg X_3)}{P'(A|X_1 X_2 \neg X_3) + P'(\neg A|X_1 X_2 \neg X_3)} = \frac{0.016}{0.016+0.072} = \frac{0.016}{0.088} = 0.1818...$

$\endgroup$
  • $\begingroup$ Your denominator is correct but your numerator is not. $P(\neg X_3|A)\neq 0.6$. $\endgroup$ – Mike Earnest Jun 5 '18 at 17:19
  • 2
    $\begingroup$ that was just a typo, on my paper it was 0.8. The numerator is still 0.016 $\endgroup$ – hr0m Jun 5 '18 at 19:00
1
$\begingroup$

Your statements about independence are somewhat lax, and this is also what seems to have caused the error – I think this is a good opportunity to become more clear about what (in)dependence means and doesn't mean.

You write "All $X_i$ are dependent on and only on $A$", and you later conclude that the $X_i$ are independent. It's not entirely clear to me what exactly "being dependent on and only on $A$" means, but from the official solution, it seems that the intended meaning is: The $X_i$ are random variables whose joint distribution is as if you first throw a coin (event $A$) and then independently choose the $X_i$ depending on $A$, with the given conditional probabilities. This does not mean that the $X_i$ are independent; in fact, the fact that they all depend on the same event $A$ makes them dependent.

$\endgroup$
  • $\begingroup$ That's exactly what i thought my mistake was. However i am very confused. So if they are dependent then $P(X_1 X_2 X_3) \neq P(X_1)P(X_2)P(X_3)$ However, why can i simplify $P(\neg X_3 | A X_1 X_2) = P(\neg X_3 |A)$? $\endgroup$ – hr0m Jun 5 '18 at 19:19
  • $\begingroup$ All this juggling around the $|$ is really confusing :) $\endgroup$ – hr0m Jun 5 '18 at 19:21
  • 1
    $\begingroup$ @hr0m: $P(\neg X_3 \mid A X_1 X_2) = P(\neg X_3 \mid A)$ (note that the spacing looks better if you use \mid instead of the vertical bar) because the $X_i$ "depend on each other only through $A$" -- if you know $A$, you know all there is to know about the distribution of $X_3$, there's no additional information in $X_1$ and $X_2$. But if you don't know $A$, then there is information in $X_1$ and $X_2$ about $X_3$ (since there's information about $A$), so they're not independent, just independent conditional on $A$ (see en.wikipedia.org/wiki/Conditional_independence). $\endgroup$ – joriki Jun 5 '18 at 19:45
  • 1
    $\begingroup$ @hr0m: I suspect you mean $P(X_1X_2\neg X_3)$, since that's what you need in your calculation. You can compute either of these just like you computed $P(X_i)$: \begin{eqnarray*} P(X_1X_2\neg X_3)&=&P(X_1X_2\neg X_3\mid A)P(A)+P(X_1X_2\neg X_3\mid\neg A)\\ &=&0.2\cdot0.2\cdot0.8\cdot0.5+0.6\cdot0.6\cdot0.4\cdot0.5 \\&=&0.088\;. \end{eqnarray*} (That the probabilities conditional on $A$ or $\neg A$ are products follows from conditional independence with respect to $A$.) Using this in your calculation yields the right result: $$ \frac{0.016}{0.088}=\frac2{11}=0.\overline{18}\;. $$ $\endgroup$ – joriki Jun 5 '18 at 21:02
  • 1
    $\begingroup$ Aaaah of course. Thank you! It is so obvious now! $\endgroup$ – hr0m Jun 6 '18 at 5:50

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.