Multi-objective optimization using surrogate functions

In multi-objective optimization problems, expensive high-fidelity simulations are commonly replaced by response surfaces approximating the parameter-criteria dependencies. Response surface or metamodel approximations allow for fast evaluations of a large set of designs. As it is known from the literature, the problem of finding the optima of a set of vectors can be implemented efficiently as a search strategy. Such an approach does not yield good results, though, if there is a strong bias within the function that disadvantages Pareto-optimal designs in objective space. As an answer we suggest a novel algorithm that produces a trajectory towards the Pareto front starting from an initial design. The applicability of our algorithm is limited to the case where no local Pareto fronts exist along the trajectory.