My question refers to the section "4.1 Optimizing EI in the TPE algorithm" on page 4 of the paper Algorithms for Hyperparameter Optimization (https://papers.nips.cc/paper/2011/file/86e8f7ab32cfd12577bc2619bc635690-Paper.pdf).

The authors provide the following step, where $\gamma := p(y < y^*)$:

$$\ell(x) \int_{-\infty}^{y^*} (y^* - y) p(y)dy = \gamma y^* \ell(x) - \ell(x) \int_{-\infty}^{y^*}p(y)dy$$

However, this is where I arrive:

\begin{align} \ell(x) \int_{-\infty}^{y^*} (y^* - y) p(y)dy &= \left[ \ell(x) \int_{-\infty}^{y^*} y^* p(y)dy \right] - \left[ \ell(x) \int_{-\infty}^{y^*} y p(y)dy \right]\\ &= \left[y^* \ell(x) \int_{-\infty}^{y^*} p(y)dy \right] - \left[ \ell(x) \int_{-\infty}^{y^*} y p(y)dy \right]\\ &= y^* \ell(x) p(y < y ^*) - \left[ \ell(x) \int_{-\infty}^{y^*} y p(y)dy \right]\\ &= y^* \ell(x) \gamma - \left[ \ell(x) \int_{-\infty}^{y^*} y p(y)dy \right]\\ &= \gamma y^* \ell(x) - \ell(x) \int_{-\infty}^{y^*} y p(y)dy \\ \end{align}

The difference between my result and the authors' result is that I still have a $y$ in the remaining integral. I got $\int_{-\infty}^{y^*} y p(y)dy$ but the authors got $\int_{-\infty}^{y^*}p(y)dy$. Do you know how to explain this? Your tips would be much appreciated!


1 Answer 1


That's a typo, it should indeed be $\gamma y^{*} \ell(x) - \ell(x)\int_{-\infty}^{y^*}y p(y) dy$ as you pointed out.

The reason this works is because the derivation can actually continue from the previous expression $\ell(x)\int_{-\infty}^{y^*}(y^*-y)p(y)dy$ as follows:

Let $$I := \int_{-\infty}^{y^*}(y^*-y)p(y)dy$$

Noting that $I$ doesn't depend on $x$.

Now substitute to get \begin{align*} EI_{y^*}(x) &= \frac{\ell(x)I}{\gamma\ell(x) + (1-\gamma)g(x)} \\ &= I \times \frac{1}{\gamma + (1-\gamma)\frac{g(x)}{\ell(x)}} \\ &= I \times \left(\gamma + (1-\gamma)\frac{g(x)}{\ell(x)}\right)^{-1} \end{align*} Since $I$ does not depend on $x$ it suffices to simply evaluate $\left(\gamma + (1-\gamma)\frac{g(x)}{\ell(x)}\right)^{-1}$ in order to determine the next candidate point to evaluate the fitness function.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

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