Dtyjlui

I have a function with 8 inputs, which yields a single output. I do not know what the function does and so cannot use any derivative-based method to minimise said output.
Currently this is done by picking n random vectors, with some bounds for each input, obtaining an output value for each vector, and picking the lowest output from a vector of n.

I have been told to try Ant Colony Optimisation, however I struggle to see how I could implement that for a function with that many inputs.

Any ideas as to how to approach this problem in a better way than it is currently being done will be much appreciated.

The function itself takes a non-negligible time to run, so I am interested in ways to most efficiently (solving the function as low number of times as possible) find a minimum.

EDIT: The bounds for each of the 8 inputs are (0,1), but could conceivably be tightened a little bit.

EDIT 2: The function is actually a collection of processes, a simulation of sorts, it produces a number of outputs. I have those 8 inputs which I am trying to calibrate such that the outputs closely match the reality I am simulating. So I have 'observed' values for those outputs, and the simulated ones for each set of 8 inputs. That's how the 'loss' is defined: as distance from that fixed set of observed values.
I do not know limits, or whether it is differentiable.

A relatively simple way to optimize a continuous, non-differentiable function is Nelder-Mead simplex optimization (which, as the linked Wikipedia article comments, is "[n]ot to be confused with Dantzig's simplex algorithm for the problem of linear optimization"). Nelder-Mead has been implemented in many computational systems (e.g. optim(...,method="Nelder-Mead") in R, scipy.optimize.minimize(..., method='Nelder-Mead') in Python, etc. etc.; Numerical Recipes will give you versions in FORTRAN, C, and C++). It's also simple enough that you could implement it yourself if necessary. The simplex method works by constructing an $n$-dimensional simplex of evaluation points and then iteratively using simple heuristic rules (e.g. "find the worst point and establish a new trial point by reflecting through the opposing face of the simplex") to update until the simplex converges approximately to a point.

It's hard to say how many function evaluations would be required, but I would say you could expect on the order of dozens to hundreds of evaluations for a reasonably well-behaved 8-dimensional optimization problem. At $approx$ 5 minutes per evaluation that would be tedious but not unfeasible.

It's not clear whether the (0,1) bounds on your inputs are hard constraints (i.e., are you pretty sure that the optimum is inside that space, or do you need to constrain the solution to that space)? If they are, things get a bit harder as most Nelder-Mead implementations don't allow for box constraints (i.e., independent upper/lower bounds on parameters). You can take a look at Powell's BOBYQA method, which might actually be a little more efficient than Nelder-Mead (although also more complex; FORTRAN implementations are available ...)

If you can guess that your function is differentiable (which seems like a good guess if the underlying simulation process is deterministic and doesn't contain sharp if/then switches), then you could also (in principle) compute gradients at each point by finite difference methods, i.e.

$$
frac{partial f(x_1,...,x_i,...,x_n)}{partial x_i} approx (1/delta) cdot left(f(x_1^*,...,x_i^*+delta,...,x_n^*) - f(x_1^*,...,x_i^*,...,x_n^*)right)
$$

for a $delta$ that is neither too small (roundoff/floating-point error) nor too big (approximation error). Then you can use any derivative-based method you like. In practice this isn't always worth it; the combination of cost and numerical error in the finite-difference computations means this method may be dominated by the derivative-free methods described above.