optimization of rule-based model on discrete variables

[Greg Ewing]

On 15/06/21 12:51 am, Elena wrote:
> I see what you mean, so I try to explain it better: Y is a vector say [y1,
> y2, ... yn], with large (n>>10), where yi = f(Xi) with Xi = [x1i, x2i, ...
> x10i] 1<=i<=n. All yi and xji assume discrete values.
> 
> I already have a dataset of X={Xi} and would like to find the rules f able
> to minimize a complicated-undifferenciable Real function g(f(X)).
> Hope this makes more sense.

Hmmm, so the problem breaks down into two parts:
(1) find a vector Y that minimises g
(2) find a set of rules that will allow you to predict each component
of Y from its corresponding X values

Is that right?

> x1...x10 are 10 chemical components that can be absent (0), present (1), 
> modified (2). yi represent a quality index of the mixtures and g is a 
> global quality of the whole process.

I ztill don't really understand. What are you going to do with this
function f once you have it?

I would have thought the idea was that if someone gives you a new
mixture X[n+1] you can use f to predict how well it will work.
But that will just give you a y[n+1], and it's not clear what to
do with that. Do you append it to Y and feed an n+1 component
vector into g?

I think I still need more information about the underlying problem
before I can help you much.

-- 
Greg