Abstract
Sparse Gaussian Processes are a key component of high-throughput BayesianOptimisation (BO) loops; however, we show that existing methods for allocatingtheir inducing points severely hamper optimisation performance. By exploitingthe quality-diversity decomposition of Determinantal Point Processes, wepropose the first inducing point allocation strategy designed specifically foruse in BO. Unlike existing methods which seek only to reduce global uncertaintyin the objective function, our approach provides the local high-fidelitymodelling of promising regions required for precise optimisation. Moregenerally, we demonstrate that our proposed framework provides a flexible wayto allocate modelling capacity in sparse models and so is suitable broad rangeof downstream sequential decision making tasks.