Surrogate Assisted Evolutionary Algorithms

Evolutionary algorithms are a class of population based random optimization algorithms. In solving practical optimization problems, they don’t rely on the derivative information of objective function, even don’t need to know the explicit mathematical expression, and only require the function values of corresponding variables, which we call fitness evaluation. However, in engineering optimization area, there exists a class of problems which cost high in time or other expenses during fitness evaluation, such as, aerodynamic design, structure design, circuit design, and drug design. Here we can combine evolutionary algorithms with surrogate models, whose main aim is to reduce the number of high cost fitness evaluation. This thesis makes deep investigation on surrogate assisted evolutionary algorithms, and aims to propose a high efficient algorithm framework, making evolutionary algorithms appropriate for this class of high cost problems. This thesis first introduced evolutionary algorithms in concise and approximate models adopted at present in detail, and then deeply surveyed existed surrogate assisted evolutionary algorithm frameworks, later determined to have a deep research on gaussian process model assisted evolutionary algorithms and conducted corresponding experiments.In all kinds of machine learning models, Gaussian process model is a more powerful model, because it can not only predict fitness values, but also the confidence of predicted fitness values. Because coarse approximate models can mislead evolutionary algorithms trapping into local optimum, the confidence information provided by gaussian process model can help improve the efficiency of real fitness evaluation. So in the forth chapter we investigated gaussian process model assisted differential evolution algorithm on5dimensional benchmark functions. However, in practical use of gaussian process models, we find an important shortcoming of them is the high training cost rapidly increased with sample number. In order to reduce the training cost of gaussian process models, we proposed to adopt local ensemble gaussian process models. Different from independent gaussian process models, local ensemble gaussian process models share the same model parameters. At last, we did experiments on local ensemble gaussian process models assisted covariance matrix adaptation evolution strategy, and compared5different sampling strategies. On8benchmark functions, we demonstrated the ensemble models can provide trustable fitness value and its confidence. During compared sampling strategies, we found combining lower confidence bound and clustering technique improves in more efficient global search.