General Meta-Model Framework for Surrogate-Based Numerical Optimization

We present a novel, general framework for surrogate-based numerical optimization. We introduce the concept of a modular meta model that can be easily coupled with any optimization method. It incorporates a dynamically constructed surrogate that efficiently approximates the objective function. We consider two surrogate management strategies for deciding when to evaluate the surrogate and when to evaluate the true objective. We address the task of estimating parameters of non-linear models of dynamical biological systems from observations. We show that the meta model significantly improves the efficiency of optimization, achieving up to 50% reduction of the time needed for optimization and substituting up to 63% of the total number of evaluations of the objective function. The improvement is a result of the use of an adaptive management strategy learned from the history of objective evaluations.