Evolutionary optimization of computationally expensive problems via surrogate modeling

We present a parallel evolutionary optimization algorithm that leverages surrogate models for solving computationally expensive design problems with general constraints, on a limited computational budget. The essential backbone of our framework is an evolutionary algorithm coupled with a feasible sequential quadratic programming solver in the spirit of Lamarckian learning. We employ a trust-region approach for interleaving use of exact models for the objective and constraint functions with computationally cheap surrogate models during local search. In contrast to earlier work, we construct local surrogate models using radial basis functions motivated by the principle of transductive inference. Further, the present approach retains the intrinsic parallelism of evolutionary algorithms and can hence be readily implemented on grid computing infrastructures. Experimental results are presented for some benchmark test functions and an aerodynamic wing design problem to demonstrate that our algorithm converges to good designs on a limited computational budget.