This paper evaluates five methods from the field of
computational intelligence and machine learning with regard to
their suitability for total vehicle concept design. These methods
are used to mimic the input-output relation of a complex
model, based on a limited number of expensive simulations
(metamodeling). Important steps of this model building process
will be investigated and possibilities for model evaluation will
be discussed. As the theoretical background is provided, Artificial
Neural Network (ANNs), Multivariate Adaptive Regression
Splines (MARS), Gaussian Process Models (GPMs), Regression
Trees (RTREES) and Support Vector Machines (SVMs) are
implemented and compared regarding their accuracy and speed
of modeling and computation. A setting from an early stage of
total vehicle concept design serves as use case.
It can be summarized that the implemented metamodels, except
RTREES and SVMs, achieve good approximation accuracy. The
ANN proved to be superior method compared to MARS and
GPMs. However, the training of the ANN took the most time,
which may be of relevance if the modeling process has to be
carried out very frequently. Then GPM or MARS offer a good
compromise between training duration and accuracy.
«
This paper evaluates five methods from the field of
computational intelligence and machine learning with regard to
their suitability for total vehicle concept design. These methods
are used to mimic the input-output relation of a complex
model, based on a limited number of expensive simulations
(metamodeling). Important steps of this model building process
will be investigated and possibilities for model evaluation will
be discussed. As the theoretical background is provided, Artificial...
»
Login
BibTeX