ABSTRAK Febrian Akhmad Baehaqi
PUBLIC Alice Diniarti COVER Febrian Akhmad Baehaqi
PUBLIC Alice Diniarti BAB 1 Febrian Akhmad Baehaqi
PUBLIC Alice Diniarti BAB 2 Febrian Akhmad Baehaqi
PUBLIC Alice Diniarti BAB 3 Febrian Akhmad Baehaqi
PUBLIC Alice Diniarti BAB 4 Febrian Akhmad Baehaqi
PUBLIC Alice Diniarti BAB 5 Febrian Akhmad Baehaqi
PUBLIC Alice Diniarti BAB 6 Febrian Akhmad Baehaqi
PUBLIC Alice Diniarti PUSTAKA Febrian Akhmad Baehaqi
PUBLIC Alice Diniarti
This thesis primarily discusses multi-fidelity Kriging methods in predicting
various analytical and engineering cases. Although the Kriging model has cut
the computational cost/expense of numerical simulations, improving the performance
of the Kriging model is still a hot topic nowadays. One of which is the
multi-fidelity Kriging methods which use the data from a cheaper simulations
to assist the building process of the surrogate models. As a result, the multifidelity
Kriging method has more samples and information in the models than
the single-fidelity Kriging method with the same computational cost thus making
the model better. Three multi-fidelity Kriging methods will be studied and
applied in this thesis. They are Cokriging, Hierarchical Kriging, and NARGP
(Nonlinear Autoregressive Gaussian Process). Later, these three frameworks will
be tested into several cases which vary in terms of dimensions, complexity, to
the degree of correlation between the low- and high-fidelity responses. The results
show that Cokriging is the most robust and most consistent in modeling the
cases especially the engineering cases. Hierarchical Kriging can be competitive to
the Cokriging results as long as an extremely good correlation between the lowand
high-fidelity responses is available. NARGP has its own specialization which
works very well when the correlation between low- and high-fidelity responses has
a nonlinearly space-dependent correlation.