Bayesian optimization for policy search via online-offline experimentation

Benjamin Letham, Eytan Bakshy

Link To Paper: https://jmlr.org/papers/v20/18-225.html

Journal Of Paper: Journal of Machine Learning Research

Confirmed Open Access: Yes

Topics From Call For Papers:

Bayesian Optimization for AutoML
Meta-Learning and Learning to learn
Applications of AutoML

Broader Impact Statement On Ethical And Societal Implications: The paper is on leveraging offline simulators and multi-task modeling to accelerate tuning of online systems. The methodological improvements and new understanding gained in the paper can lead to improvements in the many fields in which Bayesian optimization has impact, as in many of them it is also possible to construct low-fidelity simulators. It particularly enables systems improvements in settings where even Bayesian optimization may be too costly. We do not foresee any negative ethical or societal impact.