Skip to Main Content
This paper discusses mixed integer evolution strategies (MIES) assisted by metamodels based on radial basis function networks (RBFN). The goal is to make MIES more suitable for optimization with time consuming evaluation functions. A novelty of the presented research is that RBFN are studied for metamodeling in heterogeneous (mixed-integer) parameter spaces. A heterogeneous metric (HEOM) is adopted that is in conformity with the design of the MIES. In addition, cross- validation based optimization techniques are suggested for adjusting hyper-parameters of the model and avoid singularities. Empirical studies on prediction of random sets indicate good prediction capabilities of the proposed RBFN for functional landscapes of moderate dimension/smoothness. The influence of the training set size as well as of the dimension on computational complexity and accuracy of the RBFN is investigated. In the metamodel-assisted MIES, a RBFN metamodel is build and updated after each generation. The metamodel is used for selecting a small subset of offspring individuals from a bigger set of variations and thereby increase the number of promising solutions in the offspring population. The algorithm is designed in a way that in case of failure of the metamodel (e.g. "random" predictions) the metamodel-assisted MIES behaves like a standard MIES. Experimental results, both on artificial test problems and a real world application, namely the optimization of feature detectors in ultrasound images, indicate a clear acceleration that can be achieved by using heterogeneous RBFN.