Improving evolutionary optimization with metamodel-based operators

Simulation-driven optimization problems often require large computational resources and as such are often solved with algorithms which rely on surrogates, namely computationally cheaper mathematical approximations of the simulation. A common approach is to use a surrogate in conjunction with an evolutionary algorithm to seek an optimum based on the surrogate's predictions. In this setup the mechanics of the evolutionary operators are unrelated to the surrogate and do not benefit from the information it accumulates during the search. As such this paper proposes new EA operators in which surrogates are intrinsically incorporated. The proposed recombination operator combines local surrogates with an SQP search and the mutation operator uses a global surrogate based on nearest-neighbour distances. Performance analysis based on well-established test functions shows the effectiveness of the proposed implementations.