Consider the Weibull distribution with parameter $\theta$, fixed $\lambda$ and p.m.f : $$ f_Y(y)=\frac{\lambda y^{\lambda -1}}{\theta^{\lambda}}\exp(-(\frac{y}{\theta})^{\lambda}) $$ It can be shown that this distribution is from exponential family because: $$ f_Y(y)={\lambda y^{\lambda -1}}\exp(-y^{\lambda}\theta^{-\lambda}-\lambda\log\theta) $$ where $$ b(y)={\lambda y^{\lambda -1}} $$ $$ \eta=\theta^{-\lambda} $$ $$ T(y)=-y^{\lambda} $$ $$ a(\eta)=\lambda\log\theta $$ As I saw in Andrew Ng's notes here: http://cs229.stanford.edu/notes/cs229-notes1.pdf , It can be seen naturally that : $$ \theta=\exp{({\frac{\log\eta}{-\lambda}})} $$ or $$ \theta=\exp{({\frac{\log(\beta X)}{-\lambda}})} $$

where $\beta$ is our line parameter we want to estimate.My problem is actually, how to calculate $ E(T(y)|X,\beta)$ as response value the algorithm should return for a given $x$ because It seems algebraically hard for me to calculate it.

I should mention that in the examples that had been provided in Ng's notes, $T(y)$ was equal to $y$ and the calculation of $ E(T(y)|X,\beta)$ was fairly easy and he gave tidings us that most of the time $T(y)=y$ is established but unfortunately in this distribution It didn't happen.

  • $\begingroup$ Is your difficulty in computing $E(y^\lambda)$? $\endgroup$ – Glen_b Jul 10 '15 at 3:26
  • $\begingroup$ yes.I don't know if my method is actually right. and If this approach is ok, then how should I calculate it? $\endgroup$ – CoderInNetwork Jul 10 '15 at 8:36

If the hurdle is $E(y^\lambda)$, a simple substitution serves:

$f_Y(y)=\frac{\lambda y^{\lambda -1}}{\theta^{\lambda}}\exp(-(\frac{y}{\theta})^{\lambda})$

$E(y^\lambda)=\int_0^\infty y^\lambda \frac{\lambda y^{\lambda -1}}{\theta^{\lambda}}\exp(-(\frac{y}{\theta})^{\lambda}) dy$

let $u=(\frac{y}{\theta})^{\lambda}$, $du=\frac{\lambda}{\theta}(\frac{y}{\theta})^{\lambda-1}dy$

$E(y^\lambda)=\theta^\lambda \int_0^\infty u\, \exp(-u)\,du =\theta^\lambda $

Alternatively, you could look at the Wikipedia page for the Weibull distribution, which gives (in your parameterization) $E(y^r)=\theta^r\,\Gamma(1+\frac{r}{\lambda})$, and substitute $r=\lambda$.

A little simple manipulation of information from your question will let you write $\theta^\lambda$ in terms of $X\beta$.

| cite | improve this answer | |

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.