We currently use Metaheuristics.jl's population initialisation, which generates random (I believe uniform) points within the bounds of each variable. This does not necessarily generate feasible points, while having feasible points could speed up the optimisation and increase the probability of convergence.
Possible alternatives to look at are:
Multivariate normal distribution around the optimal solution (I briefly looked into this, but it didn't seem to make a big differences. I did not play too much with the variance though)
Repair method after randomly generating points (i.e. disregard infeasible points or project them onto feasible space)
We currently use Metaheuristics.jl's population initialisation, which generates random (I believe uniform) points within the bounds of each variable. This does not necessarily generate feasible points, while having feasible points could speed up the optimisation and increase the probability of convergence.
Possible alternatives to look at are: