The multipoint approximation method as a parallel optimisation framework for problems with computationally expensive responses

The multipoint approximation method was developed as an optimisation framework for engineering problems with computationally expensive and noisy responses. It combines metamodeling techniques with the trust region approach. In this method, the original optimisation problem is replaced by a sequence of approximate problems, which are solved in a sequence of trust regions. In this paper, a new implementation of the multipoint approximation method is presented, which adheres to the paradigm of parallel computations. The algorithm presented here adjusts to the available computational resources by managing the calls to the external computationally expensive software (typically, computational fluid dynamics) according to the number of available processors. The latest redevelopment of the multipoint approximation method not only assures that the available computational resources are used uniformly but also improves non-local properties of the optimisation algorithm.