Take the 2-minute tour ×
Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.

I have to develop an optimizer for a simulation. There are 7 reference values $$ r_1, r_2,\ldots,r_7 $$ (certain values which are expected to show up) and 7 corresponding actual values $$ a_1,a_2,\ldots,a_7,$$ the simulations's results. The deviations between reference values and actual values are listed as one single value, the sum of squares of all deviations. $$ f(x)=\sum_{i=1}^{7}(r_i - a_i)^2=\sum_{i=1}^{7}e^2=||e||^2 $$ where $$x_1,x_2,\ldots,x_7 $$ denote the input values.

Given this, I can calculate an arbitrary number of results $f(x)$ for many input vectors $x$, while the results are scalar values. Obviously I want to minimize the deviations. My minimization method makes use of the gradient. How can I calculate the partial derivatives to build up the gradient without knowing the function?

I only know the input values and the result, but the relationships between input values is completely unknown!

share|improve this question
    
Which minimization method are you using? If your simulation runs reasonably fast, you may be able to approximate the gradient numerically using a suitable adaptation of the Newton's formula $\frac{f(x+h) - f(x)}{h}$ for a choice of a small number $h$. Nonetheless, I think in this case a derivative-free method may be more efficient and reliable. –  Libra Nov 26 '12 at 23:05
    
Let's assume that I add $$10^-4$$ to each element of the input vector x. x is not a single value but a vector vector holding the input values, so we can as well write it as $$ f(x_1, x_2, \ldots,x_7)$$. So h is $$h=(10^{-4}, 10^{-4},\ldots, 10^{-4})$$ (7 elements). What does the fraction then look like, what's the h in the denominator? –  bogus Nov 28 '12 at 9:31
    
I have given some details in an answer. –  Libra Nov 28 '12 at 10:08
    
From your question, I'm not clear on whether $x = (a_1,...,a_7)$ or whether it also includes the 7 $r$ values as well? Also, what is the context of the problem: is it to find the "best" simulator, i.e. the simulator that gives the least deviation from reality over a wide range of inputs? If this is the problem, then I would expect you to be asking how to tune the simulator, i.e. optimise over a number of parameter variables. From your question, it seems you are trying to optimise over the inputs $x$, but that will only tell you the points over which the simulator is most accurate. –  Assad Ebrahim Dec 19 '12 at 8:46

1 Answer 1

You can assume:

$$\nabla f(\vec x) = (\frac{\partial f}{\partial x_1},\cdots, \frac{\partial f}{\partial x_7}) \approx (\frac{f(x_1+h)-f(x_1)}{h},\cdots,\frac{f(x_7+h)-f(x_7)}{h})$$

$h$ needs not to be a vector. Then, you can use the information given by the gradient to select the variable(s) to change. For example, adopting a steepest descent you may apply a change proportional to the negative of the (approximate) gradient of the function at the current point (see this Wikipedia entry, for example).

share|improve this answer
    
It may properly occur that the values in x are in different units and thus differ significantly in size. Due to that, I should probably make use of different values of h and consider h a vector of small values –  bogus Nov 28 '12 at 10:37
    
@Bogus: to calculate the partial derivative you have to "move away" of a small quantity $h$ from the current point $x_i$, regardless the size of $x_i$. So, $h$ could be the same for any component of $\vec x$. Nonetheless, perhaps normalizing the dimensions could help. –  Libra Nov 29 '12 at 7:16
    
@bogus: If you use different values of $h$, you lose the benefits of being able to compare the components of the gradient to decide on the direction of steepest descent. –  Assad Ebrahim Dec 19 '12 at 8:41

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

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