A knowledge-based approach to response surface modelling in multifidelity optimization

This paper is concerned with approximations for expensive function evaluation - the expensive functions arising in an engineering design context. The problem of reducing the computational cost of generating sufficient learning samples is addressed. Several approaches of using a priori knowledge to achieve computational economy are presented. In all these, the results of a cheap model are treated as knowledge to be incorporated in the training process. Several approaches are described here: in particular, we focus on neural based systems. This approach is then developed as a new knowledge-based kriging model which is shown to be as accurate as neural based alternatives while being much easier to train. Examples from the domain of structural optimization are given to demonstrate the approach.