Abstract
A critical bottleneck for scientific progress is the costly nature ofcomputer simulations for complex systems. Surrogate models provide an appealingsolution: such models are trained on simulator evaluations, then used toemulate and quantify uncertainty on the expensive simulator at unexploredinputs. In many applications, one often has available data on related systems.For example, in designing a new jet turbine, there may be existing studies onturbines with similar configurations. A key question is how information fromsuch "source" systems can be transferred for effective surrogate training onthe "target" system of interest. We thus propose a new LOcal transfer LearningGaussian Process (LOL-GP) model, which leverages a carefully-designed Gaussianprocess to transfer such information for surrogate modeling. The key novelty ofthe LOL-GP is a latent regularization model, which identifies regions wheretransfer should be performed and regions where it should be avoided. This"local transfer" property is desirable in scientific systems: at certainparameters, such systems may behave similarly and thus transfer is beneficial;at other parameters, they may behave differently and thus transfer isdetrimental. By accounting for local transfer, the LOL-GP can rectify acritical limitation of "negative transfer" in existing transfer learningmodels, where the transfer of information worsens predictive performance. Wederive a Gibbs sampling algorithm for efficient posterior predictive samplingon the LOL-GP, for both the multi-source and multi-fidelity transfer settings.We then show, via a suite of numerical experiments and an application for jetturbine design, the improved surrogate performance of the LOL-GP over existingmethods.