The Gaussian Process Autoregressive Regression Model .

Multi-output regression models must exploit dependencies between outputs tomaximise predictive performance. The application of Gaussian processes (GPs) tothis setting typically yields models that are computationally demanding andhave limited representational power. We present the Gaussian ProcessAutoregressive Regression (GPAR) model, a scalable multi-output GP model thatis able to capture nonlinear, possibly input-varying, dependencies betweenoutputs in a simple and tractable way: the product rule is used to decomposethe joint distribution over the outputs into a set of conditionals, each ofwhich is modelled by a standard GP. GPAR's efficacy is demonstrated on avariety of synthetic and real-world problems, outperforming existing GP modelsand achieving state-of-the-art performance on the tasks with existingbenchmarks.