Abstract:
|
Many complex processes are now simulated by computer models. Often, such computer models are computationally expensive to evaluate and as a result statistical emulators are commonly used as surrogates. Recent advances in computing and technology have led to an interest in increasingly complex simulations. Such simulations can have high-dimensional inputs, rendering current statistical approaches to emulation less effective. Furthermore when the number of input points is large, current approaches such as Gaussian process models are computationally intractable. Deep learning methods have found wide success in applications with complex structure and high dimensional inputs. In this work we investigate the use of deep learning methods for emulation in computer experiments with high-dimensional inputs. We demonstrate the utility of deep learning for prediction as well as for uncertainty quantification. In addition we investigate the impact of experimental design choices on the performance of deep learning for emulation.
|