Neural networks smart grid based optimisation for expensive functions
Résumé
Bayesian optimisation is an emerging machine
learning technique known to be efficient especially for optimising
functions which are expensive to evaluate. In Bayesian
optimisation, a Gaussian process model of the unknown function
is identified based on available data. Its estimate of the
unknown function and the associated uncertainties are used
to build a so-called acquisition function which does a tradeoff
between exploitation and exploration. The latter is then
iteratively maximised to find candidates which are promising
to be close to the optimum. In this paper, an alternative version
of Bayesian optimisation, where the Gaussian process model is
replaced by a neural network model, is proposed. As shown in
the numerical illustration of this paper, this alternative version
will require less computation time when facing optimisation
problems with initially large data sets. Since neural networks
do not naturally provide an information about the quality of
the estimates, a different strategy for the exploration objective
of our approach is proposed. The efficiency of the proposed
approach is illustrated and compared to Bayesian optimisation
on different case studies.
Domaines
Sciences de l'ingénieur [physics]
Origine : Fichiers produits par l'(les) auteur(s)