0
$\begingroup$

I am having a problem of maximizing a sum of sigmoid functions over different time instants with some constraints.

Considering the standard sigmoid function $f(x)=\frac{1}{1+e^{-\alpha x}}$ and it's derivative $f'(x)=f(x)[1-f(x)]$

In my case, it is slightly different, the sigmoid function at time instant $n$ is defined as $f_n(x)=\frac{1}{1+e^{-\alpha(\frac{x}{y_n}-z)}}$, where z>0 and $\alpha>0$ are constants and $y_n>0$ is defined for every time instant $n$.

I need to find $x$ that maximize the following:

$\sum_{n=1}^N \frac{1}{1+e^{-\alpha(\frac{x}{y_n}-z)}}\:\: -c x \:\:$ such that $\:\:x \leq X$, $\:\:\:(1)$

where $c>0$ is some constant represents a cost value, $X$ is another constant represents the maximum value of $x$ and $N$ is total number of time instants.

One of the solution I was thinking about is to use Lagrangian:

$L(x,\lambda)=\sum_{n=1}^N \frac{1}{1+e^{-\alpha(\frac{x}{y_n}-z)}}\:\: -c x \:\:- \lambda (x-X)$

where $\lambda$ is the Lagrange multiplier, since we can find the derivative of equation (1) over $x$.

I tried this method but after getting $\frac{\partial L(x,\lambda)}{\partial x}=0 \:\:$ I could't solve it.

I am not sure this type of optimization problem (1) is solvable or not. And if not is there some type of approximation/relaxation can be used to solve it.

I don't have much experience in optimization problem. So, I was hopping to get some help.

-Thanks

$\endgroup$
0
$\begingroup$

You might get a cue from the following intuition. Suppose the parameters $\alpha$, $y_n$'s and $z$ are such that each of the sigmoids have a very sharp rising edge. In that case, the sum of sigmoids will essentially look like a step function, rising at $zy_n$. Needless to say, $-cx$ is a line with a negative slope.

Now geometrically, a sum of a step function (with $zy_n$'s as transition points) and a negative slope line must have local maximas at $yz_n$'s only. Moreover, the local maximas grow in value from the left to the right on the real line. Given an upper bound $X$, one chooses as solution the highest transition point lesser than $X$.

For the case at hand, you might want to perform a golden section search between $\frac{zy_{n-1}+zy_{n}}{2}$ and $\frac{zy_{n}+zy_{n+1}}{2}$ $\forall n$, and store these values. Also find the value of the function at $X$ and finally, choose the maximum among all these. This is only intuitive and the correctness (or otherwise) needs to be established.

$\endgroup$

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.