Design and analysis of 'noisy' computer experiments

摘要: of functions calculated by long running computer codes. The literature in this area commonly assumes that the objective function is a smooth, deterministic inputs. Yet it well known many simulations,especiallythoseofcomputational fluidandstructuraldynamicscodes,oftendisplaywhatonemightcall numerical noise: rather than lying on smooth curve, results appear to contain random scatter about trend. This paper extends previous optimization methods based interpolating method ofkriging case such noisy experiments. Firstly, we review how kriging interpolation can be modified filter out noise. We then show adjust estimate error prediction so approaches optimization, as maximizing expected improvement, continue work effectively. introduce problems associated with noise and demonstrate our approach using computational fluid dynamics problems.