Bayesian models for response surfaces and their implications for experimental design