Bayesian Optimization Allowing for Common Random Numbers

More Efficient Bayesian Optimization Through the Use of Common Random Numbers optimization is a powerful tool for expensive stochastic black-box problems, such as simulation-based or hyperparameter tuning in machine learning systems. In “Bayesian Allowing Numbers,” Pearce, Poloczek, and Branke show how explicitly modeling random seed Gaussian process surrogate model allows to exploit structure noise benefit from variance reduction provided by common numbers. The proposed knowledge gradient with numbers acquisition function iteratively determines combination input evaluate objective. It automatically trades off reusing old seeds through querying new avoid bias because small number seeds. algorithm analyzed theoretically empirically shows superior performance compared previous approaches on various test problems.