Since ACO often exploits the weight of individual components (e.g. edges in a TSP), it is particularly poorly suited to continuous domains with non-separable fitness functions. Any reason you're not willing to use standard continuous-space techniques such as ES, PSO, or DE?

Thanks stephen. But what do you mean by non-separable fitness functions?

In a separable fitness function (e.g. f = f1 + f2), each dimension can be optimized one at a time. The optimal value for each dimension (f1* and f2*) will then lead to the global optimum for the overall function. In many benchmark functions, a separable function is made non-separable through the use of a rotation matrix.

Bahram this is maybe what you are looking for:

REAL-PARAMETER OPTIMIZATION USING STIGMERGY

Peter Korošec, Jurij Šilc

Computer Systems Department

Jožef Stefan Institute, Ljubljana, Slovenia

{peter.korosec,jurij.silc}@ijs.si

Google, pdf, ...

I'd say that ACO algorithm might be usefull when it is very hard to compute gradient, when we have to optimize function with many variables (1000000). But it's only my hypothesis.